Gotta Automate 'Em All: Pokémon help me do my work

I finally caved and bought a Switch 2. If I'm being honest with myself, it was only a matter of time before I fell victim to the fomo as videos of Pokopia started taking over my feeds. It's very quickly become one of my favorite games and most of my nights are now spent trying to make some progress on the storyline. So how does that relate to my work at all?

I wear a lot of hats: creating content, doing research, giving product feedback, fostering community, analyzing data, writing sample apps, and so many other things that don't always fit neatly into a category. Over the last year I've been integrating AI into my workflow to help me do all of these things. If you're in a similar spot, you may have a handful of specialized agents or prompts you call on manually. It works, but I knew I needed to unlock the next level so that instead of manually selecting which agent to use, an orchestrator would automatically route the request for me.

Building out my personal orchestration system meant rethinking agent delegation: what role each agent plays, when to call it, and what tools they need. I could've named them "researcher" and "blog writer" and called it a day. Instead, in an effort to bring some whimsy into everyday life I had some fun assigning a Pokémon to each of my agents based on what tasks they do.

The System: 1 Orchestrator, 8 agents

🧠 Orchestrator → Metagross

Role: Routes incoming tasks to the right, specialized agent.

Metagross has four brains joined by a complex neural network. No single brain does the work, but instead the four coordinate. As a Psychic type, they don't wait for instructions. They read the request and know which agent is the right one for the job.

🔬 Researcher → Alakazam

Role: Web search, technical deep-dives, knowledge gathering.

My researcher agent is Alakazam, which felt like an obvious fit since they have an IQ of 5,000 and continuously multiplying brain cells.

✅ Validator → Slowking

Role: Fact-checks claims, solves problems methodically.

Slowking fact-checks claims, challenges assumptions, and flags anything that doesn't hold up. The Pokédex describes him as: "Incredible intellect and intuition. Whatever the situation, he remains calm and collected." He doesn't rush. He reads the output, checks the claims, and stays methodical.

📅 Personal Assistant → Indeedee

Role: Manages calendar, email, scheduling, work priorities.

Indeedee can sense emotions through their horns and act as a valet, looking after their trainer's every need. It only felt right that they handle my calendar, email, scheduling, and prioritization.

📝 Friction Log → Rotom-Dex

Role: Documents developer experience observations.

Rotom with a Pokédex becomes a self-learning cataloger, documenting everything it encounters. Therefore, Rotom-Dex helps me write friction logs, which are structured walkthroughs of a developer experience: every rough edge, every confusing error, every moment of delight, documented as it happens.

💼 Meeting Prep → Porygon

Role: Synthesizes data into briefing packages.

Porygon helps me prep before I attend meetings. They can check my calendar, look up attendees, pull recent conversation history, and produce a clean briefing: who's in the room, what we've talked about before, and suggested talking points. Porygon is the first Pokémon created using modern technology, made entirely of programming code, built to move through cyberspace and process data. They were literally built to synthesize information and package it.

🔄 Repurposer → Ditto

Role: Transforms content across platforms.

Ditto rearranges their cellular structure to transform into any Pokémon they see. They don't create from scratch, but rather transform what already exists. Therefore, Ditto helps me repurpose content. I might give them a blog post and they help me transform it into a LinkedIn post or vice-versa.

📊 Analytics → Magnezone

Role: Connects to and helps analyze data on community engagement.

Magnezone has access to community engagement data from multiple platforms and helps me extract the insights I need to make decisions. Magnezone floats through the sky receiving and transmitting signals through their antenna, constantly monitoring their territory. They're always collecting.

🌟 Generalist → Mew

Role: Fallback agent, handles edge cases.

My generalist agent is Mew, the fallback for everything that doesn't fit neatly into another category. Mew contains the genetic composition of all Pokémon and can learn nearly any move in the game. When a task is too niche for a specialist, or I just need a second brain on something, Mew handles it. Every system needs a catch-all.

Example

Me:What should my top 3 priorities be today?"

Metagross routes to Indeedee.
Indeedee checks my calendar, scans recent emails, checks tickets assigned to me.

Metagross: You have a GTM release meeting at 2pm. Want me to prep you for it?

a. GTM release meeting at 2pm — you're presenting, need to prep
b. Developer feedback doc due EOD for the SDK team
c. Conference CFP deadline tomorrow

Me: Yes

Metagross routes to Porygon.
Porygon pulls the meeting details, looks up attendees, checks recent conversation history, and produces a briefing:

Metagross: Briefing ready. Want Alakazam to research any of the topics, or Slowking to validate the numbers?
a. 6 attendees: product, engineering, marketing leads
b. Last sync was 2 weeks ago & open action item was finalizing the developer messaging
c. Suggested talking points: SDK adoption numbers, developer onboarding flow, launch timeline

Me: I need to research what our community sentiment is about the SDK and compile into a report I can present at the meeting. Validate before compiling report. Open a ticket for tracking this.

Metagross routes to Magnezone who pulls community signal data.
Metagross routes output to Slowking who validates Magnezone's data.
Metagross routes to Mew to compile validated data into a report.
Meagross routes to Indeedee to create ticket.

Metagross: Report ready and ticket created. Slowking flagged the LinkedIn count — want to remove it or add a note about it?

Under the Hood

I use Kiro CLI, Amazon's agentic AI tool for my daily work, but you can definitely take this and adapt it to your tool of choice.

Each agent has a focused system prompt. Not a generic "you are a helpful assistant" prompt. A scoped, specific prompt that defines exactly what the agent does, how it responds, and what it ignores.

For example, Slowking's prompt doesn't say "check for accuracy". It says: "independently verify every factual claim, search for counter-evidence, flag sycophancy, check that links actually exist, and challenge my assumptions directly - not just the content."

Agents only get the context and tools they need. Any agents that produce writing have access to my style guide and bio. Only Indeedee has access to my calendar and emails. Only Magnezone has access to Common Room data. Scoped access keep each agent focused and reduce the risk of something going sideways.

Each agent has its own model. I can assign different models based on what the agent does. A deep research task might warrant a more capable model, whereas a quick scheduling task doesn't need it.

The orchestrator can chain agents. Magnezone doesn't just route to one agent and stop. It can pass output from one agent into another, running them in sequence when the task requires it.

Agents are set to auto-approve. By default, every tool call requires manual approval. I configured each agent's allowedTools so that trusted tools run automatically without interrupting the flow. Alakazam can search the web without asking. Indeedee can read my calendar waiting for me to confirm it's allowed pausing. The system runs end-to-end without me clicking approve on every step.

The system learns. Each agent has its own lessons file. When I correct an agent, it logs the lesson and applies it automatically at the start of every future session. I don't have to repeat myself.

Build Your Own

Try this prompt in whatever AI tool you use (Claude, ChatGPT, Kiro, etc.):

Review the list of tasks I do regularly and group them into distinct roles (e.g. researcher, validator, scheduler, content creator)

For each role, write a focused system prompt that defines exactly what that agent does, how it responds, and what it ignores

Write an orchestrator prompt that routes incoming requests to the right agent based on intent — include a delegation rules list with trigger phrases for each agent

Suggest which tools or integrations each agent should have access to

Suggest a model for each agent based on the complexity of its tasks

Add a self-improvement loop to each agent: when corrected, log a concrete lesson to a lessons file; at the start of every session, read that file and apply every lesson immediately

As always with AI, you should review the output and make any changes you see fit.