eval-loop
Use this after a task fails or produces wrong output. Documents the failure in growth-os/eval/corrections.md so future agents don't repeat it. Three-part format—what went wrong, root cause, fix. Terse
SKILL.mdskills/shared/eval-loop/SKILL.md
GitHubEval Loop
When a task fails, document it in growth-os/eval/corrections.md so the next agent doesn't repeat the mistake.
When to use
- Outreach message bounced or got negative reply
- Copy failed commodity-gate after multiple rewrites
- Agent hallucinated customer data
- Wrong pricing, wrong brand voice, wrong exclude list
- Any task that failed and might be repeated
Format for corrections.md
Three-part terse bullet format:
## [Date] [Task description]
**What went wrong**: [one sentence, specific outcome]
**Root cause**: [why it failed—missing context, wrong assumption, unclear instruction]
**Fix**: [what to do differently—new check, new source, new gate]
Example
## 2026-09-02 Brighton cafe cold email bounced
**What went wrong**: Email to James Groom at JustAir sent as a cold pitch
**Root cause**: Agent didn't check growth-os brands/chrisschouk/excludes.md which lists James as existing contact
**Fix**: Before any chrisscho.uk outreach, read brands/chrisschouk/excludes.md and brands/chrisschouk/contacts.md
Writing style
- Terse bullet points
- Specific outcome, not generic description
- Root cause must be actionable (not "AI made a mistake")
- Fix must be a concrete step (not "be more careful")
- UK spelling
- No AI fluff
Where corrections.md lives
growth-os/eval/corrections.md in the private growth-os repo.
If growth-os isn't cloned, note the correction and tell Chris to add it manually.
Related
- growth-os-read: how to read marketing memory
- commodity-gate: quality gate before publishing