This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
What I Built
Sloth Reads reads your read-it-later pile overnight, so you don't have to.
I bookmark DEV posts all the time. I almost never go back to them. My reading list had posts from 2022 in it, still waiting for a "later" that was never coming.
So I built a sloth to read them for me:
- You save things the way you already do: bookmark a post on DEV (it syncs every night), paste any link, or share it from your phone's share sheet (it installs as an app).
- Overnight, the sloth reads them, slowly and steadily. DigitalOcean Functions wake it every hour from 1 to 5 AM. Each time it reads a few articles with Gemma 4 and writes a short summary, three takeaways, and a verdict: worth reading, skim it, or skip it, with one honest reason written for me, not a generic reader.
- At 6 AM it writes a one-minute morning brief of everything it read.
- The good ones, it reads to me. "Read & listen" opens the article in a clean reader and an ElevenLabs voice reads it out, highlighting each word as it's spoken. I can tap any word to jump there, pick from five voices (switching mid-sentence carries on from the same sentence), and it remembers where I stopped, so tomorrow it picks up from that sentence.
- It counts the minutes I got back. A skipped article saves its whole reading time; a skimmed one saves all but a minute.
Who gets to be lazier? Anyone with a bookmarks folder they feel guilty about.
Demo
Try it: slothreads.souvikbiswas.com (no sign-up). Open the morning brief or any article with "Read & listen".
 |
 |
Code
sbis04
/
sloth-reads
Overnight, an open model reads everything you saved to read later, decides what's worth your time, and briefs you in a minute the next morning. Then it reads the good ones to you, word by word.
Sloth Reads
Save it. Sleep on it.
Overnight, an open model reads everything you saved to read later, decides what's worth your time,
and briefs you in a minute the next morning. Then it reads the good ones to you, word by word
Live: slothreads.souvikbiswas.com · Demo video
What it does
- You save things to read later: bookmark posts on DEV (synced every night), paste any link, or share it from your phone (it installs as an app with a share target).
-
Overnight, the sloth reads them. DigitalOcean Functions wake it every hour from 1 to 5 AM; each time it reads a few articles with Gemma 4 and writes
- a 2 to 3 sentence summary and three takeaways,
- a verdict, worth reading, skim it, or skip it, with one honest reason aimed at you,
- an embedding for search.
- At 6 AM it writes your morning brief…
MIT licensed, including the DigitalOcean deploy files.
How I Built It
Gemma 4 does the reading. In production it's gemma-4-31B-it on DigitalOcean's serverless inference; while building, it was gemma4 on Ollama on my Mac, with exactly the same code (one environment variable switches between them). For each article Gemma writes the summary, takeaways and verdict as JSON, and I give it a one-line profile of me, so "worth reading" means worth my time. It correctly skipped four 2022 hackathon announcements still sitting in my list: "The hackathon dates have passed, making the announcement irrelevant for current development."
bge-m3 does the remembering. Every article gets an embedding (the same open model locally and on DigitalOcean), so I can search my pile by meaning: "transcribing audio with flutter" finds my old Decifer post without those exact words.
DigitalOcean runs the night shift:
-
Functions with scheduled triggers are the sloth's alarm clock:
30 19-23 * * *(1 to 5 AM where I live) reads a few articles each hour;30 0 * * *(6 AM) writes the brief. The function just pings the app, which does the slow work, so the function stays tiny. - App Platform runs the app from the Container Registry (free tier), on my own domain with HTTPS.
- Serverless inference runs Gemma 4 and bge-m3.
Tiger Data is the memory. TimescaleDB on Tiger Cloud holds:
-
itemswith a pgvector embedding and a full-text index; search is hybrid (vector and keyword results merged with reciprocal rank fusion); -
runsandsavedas hypertables (every overnight run, every minute saved); -
saved_daily, a continuous aggregate behind the "minutes saved per day" chart.
ElevenLabs is the voice, with word timings. The stream/with-timestamps endpoint returns audio and the start time of every character, which is what makes the live highlight possible. I stream it in paragraph-sized pieces, so:
- audio starts in about a second;
- generation stops the moment you stop listening (no paying for a whole article you abandoned);
- every piece is cached with its timings, per voice, so replays are free and still highlight;
- resuming or switching voice starts a new piece at the current sentence, which is cached too.
Visitors share a small daily voice budget, checked before every piece, so a busy day on the live site can't drain my credits. Everyone can always read.
Why Does Open Innovation Matter?
Because the sloth has to read everything, including the stuff I'll never read. Most of a reading pile isn't worth your time; that's the point. So the model reads far more than I ever will. With an open model at DigitalOcean's Gemma 4 price ($0.18 per million input tokens, $0.50 per million output), reading and judging an average article costs about a tenth of a cent. A whole night's pile costs less than a cent. That's what lets it read everything, every night, without me thinking about it.
Because what you read is personal. Your reading list says a lot about you. With open weights, the same Gemma and the same bge-m3 run on my own laptop through Ollama; I built and tested the entire thing that way, offline, for free. Anyone who wants their pile to stay on their machine can run it there.
Because the model is a setting, not a contract. Swapping Gemma for a bigger open model, or for one fine-tuned on my own "read / skip" choices, is an environment variable.
Where closed won: the voice. Open text-to-speech is getting good, but ElevenLabs gives me natural voices and character-level timestamps in one streaming call, and that's the feature people actually notice.
My Agent Session
I built Sloth Reads with Claude Code in a day: from reading 481 Week 3 posts to pick a category, to the overnight pipeline, the word highlight, the voice switching and this video.
hi, new week now and new chanllenge
https://dev.to/challenges/hacktoberfest-week3-2026-10-19
do the same research we did for the last challenge, check all the elligible submission, which category would be best to target for better chance of winning, etc
This session is being continued from a previous conversation that ran out of context. The summary below covers the earlier portion of the conversation.
Summary:
Context
The user is Souvik Biswas (GitHub sbis04, DEV sbis04). We built Leaf Peep for the DEV "Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass" (event #79). The deadline was Oct 11, 11:59 PM PDT, and the post is already published. Today is 2026-10-10.
User-stated rules and preferences (keep applying)
-
Commits: "in the commit author just mention me and not claude please anywhere". Never add Co-Authored-By or Claude attribution, and commit as you go. This is saved in memory at
~/.claude/projects/-Users-souvikbiswas-MobileDev/memory/no-claude-commit-attribution.md, indexed in MEMORY.md. - No em dashes in the post.
- Location privacy: "set my current location to somewhere in manhattan don …
Here's a LinkedIn version in plain language. LinkedIn suits a slightly longer story with a personal angle:
I spent this weekend building a little app for leaf peepers 🍁
Every October I end up asking the same question: are the leaves actually turning yet, and where should I go to see them? Foliage reports are weekly and vague, and they never cover the trees on my own street.
So I built Leaf Peep.
📸 Point your phone at a tree and take one photo
🎨 AI gives it a fall colour score, from still green to peak
🗺️ It lands on a shared map, so everyone can see where the colour is right now
📅 Each spot gets a simple forecast, like "peak in about a week"
❤️ You can like other people's trees and climb a friendly leaderboard
No account, no email. Pick any name and start walking. Photos are stripped of location data, and spots are rounded to about a kilometre, so nobody can track your walks.
It started with my own October stroll around Manhattan, but anyone, anywhere can add the trees near them.
The best part for me: it's meant to be ten seconds on your phone, then back outside 🌳
Try it: https://leafpeep.souvikbiswas.com
How I built it: https://dev.to/sbis04/leaf-peep-an-open-model-rea …
I just won the DigitalOcean challenge!
Congratulations, Souvik! 🎉 That's a great result for a weekend project, and it fits: the DigitalOcean side carried the whole live app, with the Droplet, the Gemma 4 scoring through their serverless inference, and the custom domain.
A few things worth doing now:
- Keep the Droplet up for now. Taking it down after judging was the plan, but the winners announcement will likely bring visitors. It costs $12/mo, scoring stays within the daily budget of 500, and your $10 alert on that DigitalOcean account will flag anything unusual. I can check the live site and admin stats whenever you want.
-
Approve new photos. Visitors' photos stay private until you approve them at
/admin, so check the "Photos to review" tile over the next few days. - Celebration posts. I can draft a short X post and a LinkedIn post that thank DEV and DigitalOcean and link the winners announcement, the app and the DEV post. Send me the announcement link and I'll write them.
- README and landing page. I can add a "🏆 Winner, DigitalOcean prize" line to the README and the GitHub Pages site, and commit it under your name only.
Want me to go ahead with any of these?
can you open the leaf peep admin once
I've opened the admin page in Chrome: https://leafpeep.souvikbiswas.com/admin
It's asking for the admin token, because this Chrome tab hasn't saved it. I can't type the token in myself since this is the live site, not a local copy. Paste ADMIN_TOKEN from leaf-peep/.env into the field and click Open dashboard. The browser remembers it after that, so you only need to do this once.
Once you're in, check the Photos to review tile for new photos from visitors. If you'd rather not paste it, I can pull the same numbers (photos, walkers, likes, photos waiting for review) here in the terminal, using the token from .env without displaying it.
lert's stop the digital ocean droplet now, so that it doesn't charge me
The user sent a new message while you were working:
I don't need to keep the webpage active
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
it was mainly for the hackathon
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The photos are backed up (54 files, 13 MB). Now destroying the Droplet:
I destroyed the Droplet, so it won't charge you anything more. A powered-off Droplet keeps billing, which is why I deleted it instead of stopping it.
-
Nothing billable is left: doctl shows no Droplets, volumes, snapshots or reserved IPs on the
leafpeep2account. -
Photos are saved: before deleting it I copied all 54 photos from the Droplet to
leaf-peep/backups/droplet-uploads/(13 MB). That folder is ignored by git, and I committed that change under your name only. - The database is unaffected. It lives on Tiger Cloud, so all photo records, walkers and likes are still there.
- What's offline: https://leafpeep.souvikbiswas.com no longer loads. The landing page on GitHub Pages, the repo, the DEV post and the YouTube video are all still up.
Optional cleanup you can do later:
- Delete the leafpeep DNS record (it now points at a deleted server) in Cloudflare.
- Revoke the DigitalOcean inference key (MODEL_KEY) in the DO console. It only costs money when it's used, so this isn't urgent.
If you ever want the app back, ./deploy/up.sh recreates the Droplet. You'd then copy the photos from the backup into it.
but this doesn't load: https://leafpeep.souvikbiswas.com
beacuse we stopped digital ocean I though it was just for the ai aprt?
No, the Droplet was more than the AI part. It ran the whole live app:
- the website itself (the map, camera, leaderboard and admin pages);
- the server behind it, handling uploads, likes and search;
- the stored photos.
The AI scoring was the only piece that ran separately, on DigitalOcean's pay-per-use inference. The database lives on Tiger Cloud, so all the data is safe. When you said you didn't need to keep the webpage active, I took that as fine to take the site offline, but I should have spelled out that deleting the Droplet means exactly that.
Your options:
- Leave it offline (free). The GitHub Pages landing page, the repo, the DEV post and the video all still work. I can change the "Open the app" button and the README link so people don't land on a dead page.
-
Bring it back as it was (about $12/month). I run
./deploy/up.shand copy the 54 backed-up photos in, which takes about 10 minutes. You'd destroy it again whenever you like; billing is by the hour, so a few days costs well under $1. - Bring it back for less (about $6/month). Same thing on the smaller 1 GB Droplet. The content-checking models are tight in that much memory, so it could be unstable. I …
</div>
<button class="agent-session-expand-btn" type="button">Show more</button>
</div>
</div>
</div>
<div class="agent-session-message agent-session-user">
<div class="agent-session-role-badge agent-session-role-user">
You
</div>
<div class="agent-session-content">
<div class="agent-session-text">
<p>ok fine keep it offline</p>
now, a new week hackathon is happening: https://dev.to/challenges/hacktoberfest-week2-2026-10-12
let's participate in this with a competely new project
read all the details from tje webpage
I think the prizes and the categories are the same
let me know what ideas are you thinking of?
</div>
</div>
</div>
<div class="agent-session-message agent-session-assistant">
<div class="agent-session-role-badge agent-session-role-assistant">
Agent
</div>
<div class="agent-session-content">
<div class="agent-session-tool-call">
<button class="agent-session-tool-toggle" aria-expanded="false" type="button">
<span class="agent-session-chevron">▸</span>
<span class="tool-name tool-name-toolsearch">ToolSearch</span>
<span class="tool-input-preview">select:mcp__devrelay-gateway__get_challenges,mcp__devrelay-gateway__get_chall...</span>
</button>
<div class="agent-session-tool-detail" style="display:none">
<div class="tool-section">
<div class="tool-section-label">Input</div>
<pre class="agent-session-pre"><code>select:mcp__devrelay-gateway__get_challenges,mcp__devrelay-gateway__get_challenge_details</code></pre>
</div>
<div class="tool-section">
<div class="tool-section-label">Output</div>
</div>
</div>
</div>
<div class="agent-session-tool-call">
<button class="agent-session-tool-toggle" aria-expanded="false" type="button">
<span class="agent-session-chevron">▸</span>
<span class="tool-name tool-name-mcpdevrelaygatewaygetchallenges">mcp__devrelay-gateway__get_challenges</span>
<span class="tool-input-preview"></span>
</button>
<div class="agent-session-tool-detail" style="display:none">
<div class="tool-section">
<div class="tool-section-label">Input</div>
</div>
<div class="tool-section">
<div class="tool-section-label">Output</div>
<pre class="agent-session-pre"><code>{"Ok":[{"id":82,"title":"Hacktoberfest Open-Source AI Challenge: Week 4","slug":"hacktoberfest-week4</code></pre>
</div>
</div>
</div>
<div data-collapsible>
<div class="agent-session-text agent-session-text-collapse">
<p>I read the full details through DEV's challenge API. First, the timing:</p>
Week 2 closes tonight at 11:59 PM PDT (06:59 UTC Oct 19). Assuming you're on India time, that's about 17 hours from now. It's doable, but only with a tight, small project. Week 3 starts tomorrow (Oct 19, 16:00 UTC) and gives a full week, but its theme isn't announced yet.
Week 2 in short. The theme is Side Hustle: build something with open-weight models or open-source AI that someone would actually pay for. The open part has to be what makes it work. The post should explain why open matters here: cost, privacy, fine-tuning for a niche, or avoiding lock-in. There are bonus points for landing a first user and telling them whether that person paid, would pay, or what they asked for next.
- Judging: writing counts most, then fit with the theme, creativity, technical execution, and use of a partner's tech.
- Prizes: the same as last week. $250 overall, six featured categories at $200 (Gemma, DigitalOcean, Render, TabPFN, Tinker, Arduino), and eleven partner categories at $100, now including a .tech domain. One win per person per challenge.
- To enter: use the Side Hustle submission temp …
her's the post template:
<pasted_content id="dd41">
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
<!-- What does it do, and who would pay for it? What problem does it solve for them? -->
Demo
<!-- Share a deployed link or a video demo. -->
Code
<!-- Show us the code! You can embed a GitHub repo directly into your post. -->
How I Built It
<!-- Which open-source AI did you use (open-weight models, agent harnesses, frameworks, local inference), and how is your project built around it? -->
Why Does Open Innovation Matter?
<!-- Why does open innovation matter for what you built? What did it make possible that a closed API wouldn't? -->
My Agent Session
<!-- Optional, but judges love it. Save your session with DevRelay and embed it with the agent_s …
Going with the shoot culler, and you'll be the first user. I'll check the local setup first.
I'll build it as Keeper, a local app you start with a folder of photos. Everything runs on the laptop with Gemma 4 through Ollama.
The user sent a new message while you were working:
what categories would this qualify in>
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
[Request interrupted by user]
first allow do an assessment of which categories would be the easiest to win based on all th submission yet and from the last few weeks ones
based on that we might decide to change the prject idea
[Request interrupted by user for tool use]
have youn stared the assessment yet??
I counted mentions of each partner in the tags, titles and summaries of all 2,904 #hf26challenge posts:
| Category (prize) | Weekend | Week 1 | Week 2 so far |
|---|---|---|---|
| Gemma ($200) | 107 | 194 | 151 |
| TabPFN ($200) | 9 | 16 | 19 |
| GitHub Copilot ($100) | 6 | 2 | 18 |
| Render ($200) | 4 | 5 | 10 |
| Sentry ($100) | 2 | 0 | 6 |
| ElevenLabs, Entire ($100) | 6 / 2 | 4 / 4 | 4 / 4 |
| Tinker ($200) | 1 | 2 | 3 |
| DigitalOcean ($200) | 0 | 1 | 2 |
| Tiger Data ($100) | 0 | 0 | 2 |
| Temporal, MongoDB, SerpApi, .tech ($100) | 1 / 3 / 0 / 0 | 5 / 3 / 1 / 0 | 2 / 2 / 2 / 2 |
| Arduino ($200) | 0 | 1 | 1 |
Week 2 has 852 posts so far, against 1,210 for Week 1.
Caveat: this undercounts, because most people list categories only in the post body, which I didn't download. The proportions are what matter, and they're very lopsided.
What this means:
- Gemma is crowded. About one in six posts mentions it, so it's the hardest category to win despite paying $200.
- DigitalOcean, Tinker and Arduino pay $200 and have almost no competition. Last week you won DigitalOcean against probably one or two other entries, likely inc …
prioritize the ones those qualify for this week first
Here are the Week 2 categories ranked by how winnable they are tonight. That weighs how many entries mention each so far, the prize, and whether we can use it properly in about 16 hours.
| # | Category | Prize | Week 2 entries so far | Can we do it tonight? |
|---|---|---|---|---|
| 1 | DigitalOcean | $200 | 2 | ✅ Yes. Same setup as your win: Gemma on DigitalOcean's inference plus a Droplet |
| 2 | Tiger Data | $100 | 2 | ✅ Yes. Reuse the Leaf Peep account and setup |
| 3 | .tech domain | $100 | 2 | ⚠️ Possibly. It needs a .tech domain, which may come free through hacktoberfest.com/my. Worth checking |
| 4 | SerpApi | $100 | 2 | ⚠️ Only if search fits the idea. A listing writer could use it to look up competing apps' keywords |
| 5 | Arduino | $200 | 1 | ❌ Only if you have a board at home |
| 6 | Tinker | $200 | 3 | ❌ Fine-tuning takes too long for tonight |
| 7 | Temporal / MongoDB / Mastra / Backboard | $100 | 1–2 each | ⚠️ Possible, but each adds new setup work |
| 8 | ElevenLabs / Entire | $100 | 4 each | ⚠️ Moderate competition |
| 9 | Sentry / Render | $100 / $200 | 6 / 10 | ⚠️ More crowded |
| 10 | GitHub Copilot / TabPFN | $100 / $200 | 18 / 19 | ❌ … |
can you actually check all the project body for the category they are participating for the ones those qualify for this week>
I'll download the full text of the 852 Week 2 posts, fairly slowly so DEV doesn't block us. Then I'll read each post's "Prize Categories" section.
[Request interrupted by user for tool use]
also I do have an ardunoi UNO q board at home
and what about these?
You stopped the download, so I'll use only what's already on disk: 476 Week 2 posts whose full text was saved by the earlier attempt. No more requests to DEV.
I worked only from posts already saved on disk, with no new downloads. That gave 207 posts, of which 163 are Week 2 submissions and 80 list prize categories:
| Category | Prize | Entries in sample | Estimated across all Week 2 (×5) |
|---|---|---|---|
| DigitalOcean | $200 | 0 | ~0–5 |
| MongoDB, SerpApi | $100 | 0 | ~0–5 |
| Arduino | $200 | 1 | ~5 |
| Tiger Data, Sentry, Temporal, Mastra, Backboard | $100 | 1 each | ~5 |
| ElevenLabs, Entire | $100 | 2–3 | ~10–15 |
| .tech domain | $100 | 4 | ~20 |
| TabPFN | $200 | 5 | ~25 |
| Tinker | $200 | 5 | ~25 |
| GitHub Copilot | $100 | 5 | ~25 |
| Render | $200 | 7 | ~35 |
| Gemma | $200 | 36 | ~190 |
Arduino UNO Q is your best shot. It's a $200 featured category with almost no competition, and you have the board. Hardware also makes the post and demo memorable, which helps with writing quality and creativity, the criteria judges weight most. The UNO Q runs Linux next to its microcontroller, so it can run a small open model on the board itself. That's a strong "why open" story: no cloud, no monthly fee and private by default, which is exactly what this week's prompt asks about.
The othe …
let's just start with withoit wasting time:
The idea that fits these best: the App Store listing writer. It's a hosted subscription tool, so DigitalOcean is central rather than bolted on. Tiger Data keeps each app's listing history, SerpApi pulls competing apps' keywords, and it could live on a .tech domain. The culler is mainly a laptop app, so it only meaningfully uses Gemma, the crowded category.
Starting the build. It's called StoreScribe for now, and we can rename it later. I'll reuse the parts of Leaf Peep that worked: Gemma through DigitalOcean (with Ollama locally), Tiger Cloud and the deploy scripts.
I have what I need. Writing the backend now.
The user sent a new message while you were working:
what quick ideas do you have for Ardunio Q which you think might win?
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
I've paused StoreScribe. The setup is done but no app code is written yet, so switching now costs nothing.
The UNO Q is an Arduino and a small Linux computer in one: it has an 8×13 LED matrix, runs Debian and can run a small open model (Gemma 3 1B or 270M) on the board itself. Here are ideas that only need the board, plus your laptop or phone for setup:
-
Billable hours box for freelancers (my pick). A little desk clock for side-hustlers who bill by the hour.
- You pick a client in a tiny web page served from the board and start a timer. The LED matrix shows the client and running time, so you don't need another browser tab.
- When you stop, you jot a quick note like "fixed login bug, call with Priya". The open model on the board turns the notes into clean invoice lines, then makes the week's invoice as a PDF or CSV.
- Who pays: freelancers already pay $9–12/month for tools like Toggl and Harvest. This is a one-off device with no subscription.
- Why open: client names and rates never leave the desk, and it keeps working with the Wi-Fi off. That's an exact match for the prompt's "keep your customers' data private" and "avoid lock-in".
- **Feasibilit …
I think doing a hardware project in limited time qwould be a hazzle, continue with the one we started
Continuing with StoreScribe. I'll write the model layer first: one wrapper that works with both Ollama locally and DigitalOcean's inference in production.
Next are the store rules: character limits, Apple's 100-byte keywords field, and Play policy words. Code will enforce these instead of trusting the model.
The schema is live on Tiger Cloud: three hypertables (listing_versions, keyword_ranks, usage) and a continuous aggregate for the usage meter. Next is the writing pipeline.
Let me test the pipeline with local Gemma before building the server around it:
Backend committed (under your name only). Now the front end, starting with the page structure:
Now the styling, with light and dark themes:
The user sent a new message while you were working:
make sure to creat a lanbdinbg page fo the app a well
and do end to end testing of the project
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Noted: I'll add a landing page at / and move the tool to /app, then test the whole thing end to end. First the app's logic:
Now the landing page at /. The tool moves to /app:
The user sent a new message while you were working:
Instead of we habing to give the following:
<pasted_content id="dd41">
What you give it
In a single form:
- App name (required, 30 characters max)
- One-line pitch (required), e.g. "Split bills with roommates and settle up without awkward reminders"
- Details (optional): features, what makes it different, tone of voice
- Who it's for and category (optional)
- Search terms (optional): up to 3, like "split bills, rent tracker". If you leave this empty, Gemma suggests 3 itself.
- Screenshots (optional, up to 5). Gemma reads them to list what the app actually does. They're shrunk for the model and never saved to disk or the database.
- Stores: App Store, Google Play, or both
- Languages: English plus up to 5 more (Japanese, German, Spanish, Hindi and others)
</pasted_content id="dd41">
what if there's also a cli version of this tool t …
All routes are up. Now a real test in Chrome:
Now the app itself: filling the form and running a real generation in the browser.
The listing came back after about 3 minutes on the local model. Checking the rest of the page and the Japanese version:
The user sent a new message while you were working:
make sure you pay attensiton to the ui and ux of the app so that it's optimized for wide desktop and tablet and mobile screen sizes and do several passes until ebverything looks perfect
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Noted. Once the CLI is wired up, I'll do several UI passes at wide desktop, tablet and phone widths. Next, the server accepts repo context from the CLI:
Testing the CLI on a real app repo from your machine:
That's a good brief from AeroTail, your plane-spotting app. Running the CLI fully offline on it:
Now the UI passes. I'll restart the server in watch mode so changes show up immediately, then check desktop, tablet and phone widths.
The hidden checkboxes inside the language and store chips aren't anchored to their chip, so they stick out and cause the sideways scroll. Fixing that:
The user sent a new message while you were working:
once the ui passes are done for consistency, make sure to take screenshots and screenrecording and create a nice video with proper elevenlabs voiceover and review it as well
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The CSS fix didn't show because the browser cached the stylesheet for 7 days, and the cache-busting value only changes when the server restarts. I'll turn caching off in development:
No more sideways scroll. But reloading should reopen the last app (Pawlog), and it's showing the empty state, so I'll check the console:
The user sent a new message while you were working:
can you open the dir in vs code
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
also keep commit the chages as you go
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Opened storescribe in a new VS Code window, and I'm committing as I go (4 commits so far, all under your name). Now reviewing the screenshots:
The user sent a new message while you were working:
and puysh to this new github repo: https://github.com/sbis04/storescribe.git
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
No overflow at any size now. Before pushing to GitHub, I'll add the MIT licence and a basic README, and confirm no secrets are tracked:
Pushed to https://github.com/sbis04/storescribe. No .env or secrets are in the repo. Reviewing the second pass:
hi, new week now and new chanllenge
https://dev.to/challenges/hacktoberfest-week3-2026-10-19
do the same research we did for the last challenge, check all the elligible submission, which category would be best to target for better chance of winning, etc
I got 479 of 481; the other 2 were deleted. Reading every "Prize Categories" section:
Before recommending, I'll check whether the Week 2 winners are out, since that matters for us:
Week 2 winners are out. StoreScribe didn't place: DigitalOcean went to "Panga", SerpApi to an eBay recall study and Tiger Data to "Perforvisor". The winners still tell us a lot, though.
Week 3: Slow and Steady 🦥
The theme: build something with open models that rewards laziness: relaxation, an automation for a task you dread, anything that makes life easier. The bonus this week is showing the lazy payoff: how much time or effort it actually saves. Judging is the same as before, with writing weighted most.
Deadline: plan for Oct 24, 11:59 PM PDT, which is 12:29 PM tomorrow (Oct 25) your time, about 28 hours away. The challenge record says that, but the text says Oct 25, which could give a day more.
Categories: three new $100 partner categories: Auth0 for AI Agents, Browserbase and Ultramarine (Paper Instruments). That makes 21 winners in total, and you can still only win once.
What everyone has entered
I read all 481 posts tagged #hf26challenge since Week 3 opened. 436 are Week 3 submissions, and 217 of those list prize categories:
| Category | Prize | Entries | |
|---|---|---|---|
| Arduino | $200 | 0 | Nobody has entered yet |
| **Digi … |
this woudlk quality for which prizes?
- Screenshot sloth. Your Desktop/Screenshots folder is full. A tool reads each screenshot with Gemma's vision, renames it to what it shows, files it into folders and flags duplicates. The payoff is concrete: "sorted 1,240 screenshots in 11 minutes; doing it by hand would take about 3 hours", measured on your own folder. It works offline with Ollama. A hosted "watch folder" with DigitalOcean is optional.
this woudlk quality for which prizes?
- Screenshot sloth. Your Desktop/Screenshots folder is full. A tool reads each screenshot with Gemma's vision, renames it to what it shows, files it into folders and flags duplicates. The payoff is concrete: "sorted 1,240 screenshots in 11 minutes; doing it by hand would take about 3 hours", measured on your own folder. It works offline with Ollama. A hosted "watch folder" with DigitalOcean is optional.
Before answering, I'll check the hub's descriptions of the less familiar categories (Entire, Ultramarine, Browserbase, Auth0), since what each partner wants decides whether this idea can qualify:
Here's how the screenshot idea maps to the categories, using the hub's description of what each partner wants:
Natural fits, where the project would use them for real:
| Category | Prize | Entries so far | Why it fits |
|---|---|---|---|
| Tiger Data | $100 | 2 | The hub asks for "embeddings with pgvector, hybrid keyword and vector search". That's a killer feature here: "find the screenshot of my flight booking" searches by what's in the screenshots. Only the text descriptions and embeddings get stored, never the images. |
| Gemma | $200 | 130 | Gemma 4's vision reads every screenshot. It's a perfect fit, but very crowded. |
| Temporal | $100 | ~11 | The hub: "survives failures, retries flaky tool calls, picks up where it left off". Sorting 1,000+ screenshots on a laptop that goes to sleep is exactly that, and it suits the sloth theme: slow and steady, it never loses its place. |
| ElevenLabs | $100 | 7 | The hub explicitly counts "generate narration for your demo", which we do anyway. |
| .tech domain | $100 | 6 | Put the landing page on a free .tech domain (claim it at hacktoberfest.com/my/promos). |
Possible, but check first:
- Entire ( …
i think this is a better idea:
- Overnight standup. While you sleep, a scheduled DigitalOcean job reads your commits, PRs and review comments, and in the morning you get your standup and weekly update ready to paste. It uses DigitalOcean heavily, Tiger Data can hold the history of "minutes saved per day", and you're the real first user.
can you shed more light on the idea and prizes we are targetting?
🦥 Idea: "Sloth Standup", a standup that writes itself while you sleep
The problem: every morning you reconstruct yesterday from memory, scrolling through commits, PRs and review threads to write two lines for standup. On Fridays it's worse, because the weekly update covers five days.
What it does:
1. At 3 AM, a scheduled job on DigitalOcean wakes up and reads your last 24 hours on GitHub: commits, PRs opened, merged and reviewed, review comments given and received, and issues closed.
2. Gemma 4 turns that into your standup: Yesterday / Today / Blockers, in your tone. On Fridays it adds a weekly update for your manager.
3. Every bullet links to its source (the commit or PR it came from). Code checks that each bullet cites a real event, so it never invents work. Judges have rewarded this kind of honesty in past winners ("asks before it…", "says I'm not sure").
4. In the morning you open one page, read it, tweak a word and click Copy. Optionally it posts straight to a Slack channel.
5. "What did I do on…?" You can search your own history: "when did I fix the calendar crash?" finds the day, the PR and the standup it went into.
**The laz …
yerah, this part is what I'm worried about:
- Whose GitHub? Your personal sbis04 repos are simple. Your FlutterFlow work repos would make a far more realistic first-user story, but they're company code. Sending PR titles and review comments to a DigitalOcean server and model may need your employer's OK. A middle path: run on work repos locally with Ollama for your own grading and numbers, and keep the hosted demo on your personal repos.
That's the right instinct. The good news is that you don't need your work repos at all, because your personal GitHub is already full of real activity: Leaf Peep and StoreScribe were both built in the last three weeks. Let me check how much there is:
Here's what your GitHub actually shows for October:
| Where | Activity | Use it? |
|---|---|---|
Your personal repos (leaf-peep, storescribe, souvik-biswas-website, tv_gallery, memories_gemma, chrono_snap…) |
About 210 commits in 3 weeks | ✅ Yes. It's your own code, mostly already public, and it's real work. |
| FlutterFlow/flutterflow (work) | 18 PRs, ~15 reviews, ~12 review comments, pushes | ❌ Leave it out entirely. |
My recommendation: build, measure and demo only on your personal repos. That's enough on its own:
- Real history to backfill: Leaf Peep (about 57 commits), StoreScribe (about 21) and your website (about 89) give about 15 working days of real standups to generate and grade. That's an honest first-user story: "I ran it over the three weeks I spent building my last two hackathon projects."
- Nothing from work goes to DigitalOcean, to any model or into the post. No permission questions, nothing to redact.
- It's still you as the user. These are your side projects, and "a standup for your side-project week" fits the lazy theme well.
One trade-off: your personal repos have commits but no PRs or reviews, since you …
any other ideas that can qualitry for the digital ocean category and matches this week's theme?
any other ideas that can qualitry for the digital ocean category and matches this week's theme?
Yes, here are four more. Each uses several DigitalOcean products, so the DigitalOcean entry would be meaningful rather than "it runs on a Droplet". Each fits "rewards laziness" with a payoff we can measure, and uses only your personal data. All are buildable in the roughly 27 hours left.
What DigitalOcean offers (from the hub: "host your app or agent, run an open-weight model on a GPU Droplet, or build an agent on the Gradient AI Platform"):
- Functions with scheduled triggers, for nightly jobs that cost about $0;
- serverless inference for Gemma 4;
- App Platform for the web page, free for static sites;
- Spaces for storing files such as audio;
- the Gradient AI agent platform.
1. 🗞️ Reading-list sloth, "Read it later, finally"
You save articles you never read. Every night a DigitalOcean Function takes what you saved that day and Gemma writes a 5-minute morning brief: three lines per article, plus a "worth reading in full?" verdict with the reason. Optionally an ElevenLabs voice reads the brief, so you can listen while making coffee.
- Your real data: your DEV reading list, which I can read through DevRelay, plus any links you drop in.
- * …
let's do this: the reading-list sloth
can we also use elevenlabs to read an article for the user like streaming?
let's do this: the reading-list sloth
can we also use elevenlabs to read an article for the user like streaming?
Yes, and it's a great fit. ElevenLabs has a streaming voice API: audio starts playing within about a second while the rest is still being generated, so "Read it to me" on any article starts almost instantly. It also adds real product depth to the ElevenLabs entry.
One catch I'll design around is ElevenLabs credits. The free plan has about 10,000 characters a month, and one long article can use 10,000 on its own. So:
- The morning brief (short summaries) uses ElevenLabs, at about 1,500 characters a day.
- "Read the full article" streams with ElevenLabs's cheaper Flash voice. Each article's audio is cached, so replaying it is free.
- Testing and recording use a free open voice model running locally (Kokoro), so no credits are burned while building. It also makes a nice "why open" point.
- Hacktoberfest participants can claim more ElevenLabs credits at hacktoberfest.com/my if you need them.
First, how big your reading pile actually is:
Yes, and it's a great fit. ElevenLabs has a streaming voice API: audio starts playing within about a second while the rest is still being generated, so "Read it to me" on any article starts almost instantly. It also adds real product depth to the ElevenLabs entry.
One catch I'll design around is ElevenLabs credits. The free plan has about 10,000 characters a month, and one long article can use 10,000 on its own. So:
- The morning brief (short summaries) uses ElevenLabs, at about 1,500 characters a day.
- "Read the full article" streams with ElevenLabs's cheaper Flash voice. Each article's audio is cached, so replaying it is free.
- Testing and recording use a free open voice model running locally (Kokoro), so no credits are burned while building. It also makes a nice "why open" point.
- Hacktoberfest participants can claim more ElevenLabs credits at hacktoberfest.com/my if you need them.
First, how big your reading pile actually is:
The user sent a new message while you were working:
start on it
make sure you do end to end testing and complete the entire implenenmation,
use the leaf peep elevenlabs key
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
just use elevenlabs for the demo
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
when recording
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
otherwise please don't spend any credits while bulding aor when not requried
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Same embedding model locally and on DigitalOcean (bge-m3), so search behaves identically in dev and production. Scaffolding Sloth Reads:
Schema first. It's Tiger Data (TimescaleDB) with pgvector for search, and hypertables for nightly runs and minutes saved:
Now the database helpers and the model layer. It's the same pattern as StoreScribe, plus embeddings:
Testing the core on your real DEV reading list (17 posts), with local Gemma:
The chunks come out at a steady ~1,700 characters each. Now the server:
The chunks come out at a steady ~1,700 characters each. Now the server:
All three date fixes are in. Now the page:
All three date fixes are in. Now the page:
Now the styling: a calm, warm palette with light and dark themes:
Now the page logic: the player, the cards, search, filters and owner mode.
Now the page logic: the player, the cards, search, filters and owner mode.
Now a first look at desktop and phone sizes, in light and dark:
The count is 3 rather than 1, which suggests the insert matched more than once. Checking:
The insert also matched two await load(); lines inside the form handler. Restoring those two lines so only the real block at the end remains:
The Chrome extension just disconnected. Rather than wait, I'll run the same end-to-end test in headless Chrome with Puppeteer: real clicks and typing, checking each step and capturing screenshots.
The first four checks pass, but the new card never appeared. I'll check the API directly:
The server did mark it done, so the problem is in the page. I'll reproduce just that step and look at what the Done tab shows:
The DigitalOcean config is committed:
- App Platform builds from GitHub on every push.
- Functions wake the sloth hourly from 1 to 5 AM your time, then write the brief at 6 AM.
- Gemma 4 and bge-m3 run on DigitalOcean's inference.
Here's where things stand, and what I need from you.
Done and tested (local Gemma and the free voice; no ElevenLabs credits or paid inference used):
- Overnight reader: reads each saved item and writes a summary, three takeaways and a read/skim/skip verdict with a reason tailored to you. It skipped the expired 2022 hackathon announcements and flagged your Leaf Peep post as worth reading.
- Morning brief: about one minute of listening for 34 minutes of reading.
- Time saved: 46 minutes logged in Tiger Data, shown in a daily chart.
- Streaming voice: read-aloud in chunks, so generation stops when you stop listening. Everything is cached, and visitors can only replay cached audio, so they can never spend your credits.
- Search: finds articles by meaning and by words (pgvector plus full text on Tiger).
- Saving links: a link box, the phone share sheet once installed on your Home Screen, and DEV reading-list sync.
…
her's the githyb repo, it's private for now: https://github.com/sbis04/sloth-reads.git
The user sent a new message while you were working:
I have this many credits on elevenlad that expires on nov 10th I think
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
so should be fine to just use elevenlabs now
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
how does the listen mode work now?
I was thinking mayeb it would be better ux to bring up the artile on screen inside the app, and maybe have a live text highlight while reading an article? is that possible
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The app is live on App Platform: https://sloth-reads-llcsa.ondigitalocean.app. I'll check it and then answer your question.
The user sent a new message while you were working:
also the listen seesm to be disabled?
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Now the reader logic: rendering words, syncing the highlight to the audio clock, follow-along scrolling and tap-to-jump.
The user sent a new message while you were working:
feel free to publiush the sub domian changes to souvikbiswas.com
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
You're happy for me to publish the subdomain DNS. I'll check how souvikbiswas.com's DNS is managed from here:
The user sent a new message while you were working:
can you give the instructions what I nee dto add to cloudflare
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Adding the domain to the DigitalOcean app first, so it can issue the HTTPS certificate as soon as your record exists:
The DigitalOcean side is ready. In Cloudflare, add one record:
- Go to dash.cloudflare.com, open souvikbiswas.com, then DNS → Records → Add record.
- Fill in:
-
Type:
CNAME -
Name:
slothreads -
Target:
sloth-reads-llcsa.ondigitalocean.app - Proxy status: off, so the cloud is grey and it says "DNS only". DigitalOcean issues the HTTPS certificate itself, and Cloudflare's proxy would get in the way.
- TTL: Auto
-
Type:
- Click Save.
While you're in DNS records, tidy two leftovers:
- Keep the storescribe A record ([REDACTED]). StoreScribe is still live for judging.
- Delete the leafpeep A record if it's still there. It points to the Droplet we deleted.
Tell me when it's saved. DigitalOcean usually issues the certificate within a few minutes, and I'll check that https://slothreads.souvikbiswas.com works.
Meanwhile I'm fixing the last reader check. Everything else passes, including the live word highlight and seekable replays, but tapping a paragraph to jump there doesn't move playback yet.
done
https://slothreads.souvikbiswas.com is live with HTTPS (DigitalOcean is still finishing its domain checks, but it already serves). I'll note the new records in memory, then debug tap-to-jump:
All 9 reader checks pass. A look at the reader mid-playback:
The reader is live on https://slothreads.souvikbiswas.com. Now the DigitalOcean Functions that wake the sloth overnight:
The whole chain works: DigitalOcean Function → App Platform → Tiger Data, and the run is logged ("schedule", 84 ms). Committing, then checking the reader at every screen size:
Sloth Reads now runs on DigitalOcean, with the reader view you asked for at https://slothreads.souvikbiswas.com.
Read & listen opens the article inside the app in a clean reading view:
- each word lights up as it's spoken, using ElevenLabs' character timings;
- the view follows along, and pauses following if you scroll by hand;
- tap any word to jump there;
- closing the reader keeps the audio playing, and tapping the mini player reopens it.
All 9 reader checks and the earlier 18 end-to-end checks pass. I also checked it at desktop, tablet and phone sizes, in both themes, with no layout problems.
What runs where on DigitalOcean:
- App Platform runs the app (from the free Container Registry), at slothreads.souvikbiswas.com with HTTPS.
- Functions fire hourly from 1 to 5 AM your time to read a few saved articles, then at 6 AM to write the morning brief. I invoked one by hand: Function → App → Tiger works, and the run was logged.
- Serverless inference runs Gemma 4 31B for reading and bge-m3 for search.
- Total about $5 a month, prorated: roughly $1.20 for a week of judging.
Spending so far: about 1,000 ElevenLabs credits (one brief), and …
why disavbled still?
You're seeing the site as a visitor there. I built it so visitors can only replay audio that already exists, to protect your credits, and nothing has been voiced on the live site yet. That's too strict, though, because judges will see the same disabled button. I'll change it:
- "Read & listen" always opens the reader, so anyone can read the clean article view, even with no audio.
- Visitors can generate voice too, within a daily budget. I'll cap visitor voice at 12,000 characters a day in total (about 4 articles), so a busy day can't drain your credits. Since audio is generated only while someone listens, the real cost is usually lower. Over the cap, the reader still opens, with a "the voice is resting until tomorrow" note.
- As the owner, you stay unlimited.
The user sent a new message while you were working:
openb the project in vs coed
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
added dev api key
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Both look right: the top search result is Decifer, and the dark phone reader highlights the spoken word ("read"). Writing the README:
Both look right: the top search result is Decifer, and the dark phone reader highlights the spoken word ("read"). Writing the README:
The user sent a new message while you were working:
I think the payback length doesn't make sense here
instead what if we show a nice voice sound graph and stil keep the time there
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
also can we give user optin to swicth the voices?
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Five voices work with this key (Eric, George, Lily, Brian and Matilda); Sarah isn't available. Wiring the picker through the voice module, the API and the player:
The user sent a new message while you were working:
also is it okay to rerender like this without the dev.to controls, isn't it bad for dev.to if they are not gettig the read count?
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
if it's possible to like bring up the normal dev.to artivle page, and add the highlight, that woudl be nicer?
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Now the page: a "Read on DEV ↗" button for other authors' posts when you're a visitor, a ❤️ nudge in the reader, and the voice picker.
The user sent a new message while you were working:
no need to buidl an extension
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Still small, and suspiciously uniform. Headless Chrome may not feed real audio into the analyser, so I'll log the raw levels to tell a code problem from a test-environment one:
That's a real voice waveform now: the bars move with the speech, and with cached audio the time shows "0:05 / 1:03". I'll run the full regression (both test suites) before shipping:
The user sent a new message while you were working:
actualy allow the read and listen on all ecen for visitors, but make sure there's a redict lin k t the original article rthere
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Visitors can now open every post, and every card has its "Original ↗" link. Checking how the reader's byline and the thanks note look:
Now shipping all of this to the live site and rerunning the full regression locally while it deploys:
The user sent a new message while you were working:
once doen, let's compelte the remaining items for the real submirrins
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
use one of the female voice for the demo, but feel free to showcase amultiple voice/ voice switching as well
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
one thing I have noticed, if I switch a voice mid reading, it starts from the start, cna we fix that, it should always resume from the same place
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The user sent a new message while you were working:
also maybe we save upto which point an article is read so that it resumes even if the user comes back to it later
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Now the player side: starting from any word, remembering your place, resuming on voice change, and the "Resume · 40%" cards:
Switching voice never started the new audio. I suspect today's visitor voice budget (12,000 characters) is used up by all my testing, since this test runs as a visitor. Checking:
Still stuck after the switch even as owner, so this is a real bug. I'll capture what the player is doing at that moment:
Committed and pushed: saved progress, voice switching that keeps your place, tapping ahead to start from that word, and the clear "voice isn't available" message.
Waiting on you: raise the Leaf Peep API key's credit limit in ElevenLabs (Developers → API Keys → edit Leaf Peep → credit limit up or unlimited). It's capped at 9,000 and has 109 left, which blocks:
- re-running the resume test (the code is done; the test just needs a working voice);
- new voice on the live site (cached audio still plays);
- recording the demo with a female voice and voice switching.
As soon as it's raised I'll rerun the tests, deploy to the live site and record the video. Meanwhile I'll draft the DEV post.
Cost so far today: about 12,400 characters voiced, roughly 9,000 credits (most of it this testing). DigitalOcean inference on the live site has been near zero.
added a new api key
what are left now?
The new key works, and the live site is updated with everything (image 6). Visitors have their voice again: your listening and my testing no longer count against their daily budget.
Built and tested (all live at slothreads.souvikbiswas.com):
- Overnight reading with verdicts, the morning brief, and minutes saved shown in a chart.
- The reader with the live word highlight and the voice waveform.
- Five voices. Switching keeps your place, and each article resumes where you left off ("Resume · N%").
- Tap any word to jump there.
- Search by meaning.
- "Original ↗" links and the ❤️ nudge for authors.
- On DigitalOcean: App Platform, Functions schedules, the free Container Registry and inference, at your domain.
Still to do:
| # | Item | Who | Notes |
|---|---|---|---|
| 1 | Bookmark 15–20 DEV posts you'd like to read | You | Tonight's 1–5 AM run reads them and your 6 AM brief writes itself. That's the real "lazy payoff" for the post. |
| 2 | Make the repo public | You | It's private, and judges need to see the code: gh repo edit sbis04/sloth-reads --visibility public --accept-visibility-change-consequences, or GitHub → Settings. |
| 3 | **Pre-voice the brief and yo … |
continue
Starting with item 3. First I'll make Matilda (female) the default voice for new listeners, then check how much text pre-voicing would cost:
The user sent a new message while you were working:
is the live site updated
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
Much better: the full title shows and the controls are comfortable. The waveform still gets squeezed when the total time is shown, so on phones I'll show just the current time:
Recording now. It saves a real post on camera, the Tinker fine-tuning one, which suits your interests:
While that records, the narration in Lily's voice:
About 72 seconds of narration. Now the Sloth Reads cards and window frame, in the app's own look:
Narration and cards are ready. While the recording finishes, I'll look for the Week 3 submission template on the challenge page so the post matches it exactly:
README pushed. I'll also deploy the latest UI (the two-row phone player) to the live site while the recording runs:
The recording stopped at the morning-brief step: the highlight never appeared. Debugging that step on its own:
The full walkthrough recorded (102 s). Next I'll pull the app's real audio for the moments where it speaks, so the video plays exactly what was heard:
The user sent a new message while you were working:
have made the repo public
here's the submission template:
<pasted_content id="dd41">
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
What I Built
<!-- What does it do, and what effort does it save? Who gets to be a little lazier because of it? -->
Demo
<!-- Share a deployed link or a video demo. -->
Code
<!-- Show us the code! You can embed a GitHub repo directly into your post. -->
How I Built It
<!-- Which open-source AI did you use (open-weight models, agent harnesses, frameworks, local inference), and how is your project built around it? -->
Why Does Open Innovation Matter?
<!-- Why does open innovation matter for what you built? What did it make possible that a closed API wouldn't? -->
My Agent Ses …
I'll also make the final scene scroll properly down to the "night shift" log, then re-record:
This take worked: the tap jumped the article from 4.8 s to 49.4 s, exactly as intended. Checking the resume and night-shift scenes:
The cut is 1:43. Reviewing it: loudness first, then a frame every 4 seconds:
One gap: restarting the same article from the top should also clear the "already read" shading. Adding that, then running all four test suites before re-recording:
The replay still fails, so I'll reproduce exactly that sequence (play to the end, close, play again) and inspect it:
All 9 reader checks pass again. Committing, then re-recording the video with the smoother voice switch:
The cut is 1:47. Checking the voice-switch section frame by frame (37–48 s), which should now be smooth:
Timings shifted a little, since Gemma read the post faster this take. Updating the assembly to match and rebuilding:
Nearly identical timings. Applying the small offset changes, fetching George's exact clip, and building the final cut:
The cut works overall. One gap: the resume scene (around 1:02) shows the top of the home page, not the card with "Resume · 9%". I'll check the raw recording for that moment:
ok fine keep it offline
now, a new week hackathon is happening: https://dev.to/challenges/hacktoberfest-week2-2026-10-12
let's participate in this with a competely new project
read all the details from tje webpage
I think the prizes and the categories are the same
let me know what ideas are you thinking of?
[Request interrupted by user for tool use]
also I do have an ardunoi UNO q board at home
and what about these?
The user sent a new message while you were working:
once the ui passes are done for consistency, make sure to take screenshots and screenrecording and create a nice video with proper elevenlabs voiceover and review it as well
This is how Claude Code surfaces messages the user sends mid-turn — within the running turn, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.
The card does show "Resume · 9%". The resume scene just starts a little early and ends during the scroll back up. Shifting it to 76.5–79.0 s:
The resume scene now shows the Leaf Peep card's "Resume · 9%" with its progress line. Copying the video into the repo and making the thumbnail:
The thumbnail works. I'll refresh the README screenshots so they show the new player and paragraph-style brief, then commit:
The video is reviewed and ready for you to upload. I've opened it in Finder.
File: ~/MobileDev/sloth-reads/docs/sloth-reads-demo.mp4 (1:47, 1080p)
Thumbnail: ~/MobileDev/sloth-reads/docs/img/thumbnail.jpg
Review checks:
- Narration: Lily, a female voice. All 13 lines fall inside their scenes.
- App audio: you hear Matilda read the brief, then George continuing the same sentence after the voice switch, then Matilda again in the article reader, including the tap-to-jump. Those parts are labelled "Real app audio".
- Other labels: sped-up parts (the waiting while Gemma reads) are marked.
- Live URL: in the browser frame throughout, plus on the title card and the outro.
- Loudness: −15.9 LUFS, the standard level for YouTube.
- All scenes checked: the pile, verdicts, saving a link, the brief, the voice switch, the reader, resume, search and the night shift.
For YouTube:
- Title: Sloth Reads: an open model reads your read-later pile overnight (Gemma 4)
- Description:
```
Save it. Sleep on it. Sloth Reads reads everything you saved overnight with Gemma 4, an open model, tells you what's worth your time, and reads the good one …
The Lazy Payoff
I pointed it at my own DEV reading list.
- 19 articles, 92 minutes of reading. The sloth read all of them.
- 5 worth reading, 10 to skim, 4 to skip. The 4 skips were hackathon announcements from 2022.
- 52 minutes saved without me opening a single one: the skips entirely, the skims down to a summary.
- The morning brief: 1 minute 11 seconds to hear what's in all of it.
My honest estimate: left to myself, I'd have opened maybe two of those 19 posts this month and felt guilty about the rest. Instead, the first brief took 1 minute 11 seconds to hear, I know which 5 are worth my time, and the rest are off my conscience. That's the kind of laziness I can get behind.
Prize Categories
- Best Use of DigitalOcean: Functions on a schedule run the night shift, App Platform and Container Registry host the app, and serverless inference runs Gemma 4 and bge-m3.
- Best Use of Gemma: Gemma 4 reads, judges and summarizes every article and writes the morning brief, on DigitalOcean in production and on Ollama locally.
- Best Use of Tiger Data: pgvector embeddings with hybrid search, hypertables for runs and minutes saved, and a continuous aggregate behind the chart.
- Best Use of ElevenLabs: streaming speech with character timestamps drives the live word highlight, with five voices, resume, and pay-only-for-what-you-hear streaming.
