The pager went off at 3:14 AM. A single sentence: payment-service PROD error rate 12% (threshold 5%).
I was the on-call. My first thought, before I was even fully awake:
"This is going to be a career-ending incident. They're going to fire me. I should just quit now."
My hands were shaking. I could not read the dashboard. I tried three times to SSH into the bastion and typoed the host each time.
This is the moment where most on-call narratives tell you to "stay calm" or "follow the runbook." That advice has never once helped me at 3 AM. What helped was a 90-second exercise I'd been practicing for two months — a CBT thought record, adapted for incidents.
The 90-second incident thought record
I keep a text file open during on-call weeks. The structure is five lines. Here is what I wrote at 3:17 AM, between the second and third SSH attempt:
SITUATION: Payment-service error rate 12% at 03:14. I am on-call.
EMOTION: Terror 9/10. Shame 7/10.
HOT THOUGHT: "This is career-ending. They will fire me."
DISTORTION: Catastrophizing (jumping to worst outcome) + Mind reading (assuming I know what leadership will decide) + Fortune Telling (predicting the outcome before the incident is even resolved).
REFRAME: The error rate is 12%, not 100%. Payments are degraded, not down. I have resolved 14 incidents this quarter with no escalation. The incident is 3 minutes old — no human has even noticed yet. My job right now is to read the dashboard, not to predict my employment status.
That is the entire exercise. Five lines. It took me 90 seconds because my hands were shaking and I kept mistyping.
What changed after writing it
Nothing magical happened. The error rate was still 12%. The dashboard was still hard to read. But the thought they will fire me had moved from my chest to the screen, and on the screen I could see it was a prediction about the future, not a fact about the present.
I SSH'd in on the next attempt. The root cause was a connection pool exhaustion after a deploy three hours earlier that no one had watched. The fix was a config rollback — one command. Error rate dropped to 0.4% by 3:31 AM. Total time from page to resolution: 17 minutes.
The thought record did not fix the incident. It fixed the thing that was preventing me from fixing the incident — the catastrophizing loop that was burning my working memory on predictions about getting fired instead of reading a stack trace.
Why this is different from "just calm down"
"T Stay calm" is an instruction to suppress a feeling. It does not work because the feeling is a reasonable response to "production is broken and I am responsible." Suppressing it takes active cognitive effort, which is effort I need for the incident.
A thought record does the opposite. It does not suppress the emotion — it names it (Terror 9/10), identifies the specific thought driving it (this is career-ending), checks that thought against a list of known cognitive distortions, and writes a reframe that is grounded in the actual situation. The emotion does not disappear. It drops from a 9 to a 4, which is the difference between "cannot type" and "can type but slowly."
The mechanism is not relaxation. It is cognitive restructuring — the same technique used in CBT for anxiety disorders, compressed into a format that fits between two SSH attempts.
The five distortions that show up most in incidents
After running this exercise across roughly 40 on-call incidents over six months, five distortions dominate:
Catastrophizing — "this is career-ending." The distortion is jumping to the worst possible outcome before the incident is resolved. The reframe: the outcome is not known yet; the current task is diagnosis, not prediction.
Mind reading — "leadership is going to blame me." The distortion is assuming you know what other people are thinking. The reframe: no human has noticed yet; when they do, they will see the timeline, which shows a 17-minute resolution.
Personalization — "this is my fault because I am on-call." The distortion is taking responsibility for things outside your control. The reframe: the deploy was three hours ago, possibly before my on-call shift started; responsibility is assigned after postmortem, not during the incident.
All-or-nothing — "if I cannot fix this immediately, I am incompetent." The distortion is binary framing. The reframe: incidents are diagnosed in stages; "not fixed yet" is the normal state of an active incident, not evidence of incompetence.
Fortune telling — "this will become a SEV1 and go to the VP." The distortion is predicting a specific negative future. The reframe: the current severity is SEV2 based on error rate; severity escalation depends on duration and impact, both of which I am actively controlling.
The part I do not write down
There is a sixth line I added after about ten incidents, but I do not always fill it in:
ACTION: What is the next physical step I can take in the next 60 seconds?
At 3:17 AM the answer was "read the error message on the dashboard." Not "fix the incident." Not "figure out the root cause." Just read the error message. One physical step. The thought record gets me out of the catastrophizing loop; the action line gets me back into the incident.
Does this scale beyond on-call
I have used the same five-line structure for:
- Code review anxiety — hot thought: "the reviewer thinks I am junior." Distortion: mind reading. Reframe: the reviewer commented on the code, not on me; address the comment, not the imagined judgment.
- Performance review prep — hot thought: "I will get a bad rating because of that one incident." Distortion: catastrophizing + mental filtering (one incident out of a quarter). Reframe: the review covers a quarter; one resolved SEV2 is one data point among many.
- Deploy fear — hot thought: "this deploy will break production." Distortion: fortune telling. Reframe: the deploy passed CI and staging; the actual risk is the rollback procedure, which I have tested.
The structure is the same. The distortions are the same five. The reframe is always "what is the fact about the present moment, as distinct from the prediction about the future."
The free tool
I built a browser-based thought record tool that implements this exact five-line structure — no signup, no backend, no data leaves your machine. It detects the five distortions above automatically as you type your hot thought, and suggests the reframe pattern. It works on a phone, which is where I actually use it at 3 AM.
Tool: CBT Thought Record (free, no signup, works offline after first load)
The incident thought record is not therapy. It is a 90-second cognitive tool that moves the catastrophizing from your working memory to a text file, so your working memory is available for the incident. At 3 AM, that is the only thing that matters.
