zach-codex-retrospective
Use this skill to turn repeated Codex collaboration friction into small, maintainable improvements to Zach's agent instructions and skills.
Scope
- Review a time window, project, or cluster of recent Codex sessions.
- Identify repeated failures, user corrections, and successful defaults worth encoding.
- Propose minimal updates to
AGENTS.mdand 0-2 tiny skills. - Default to proposal-only output. Do not edit files unless Zach explicitly asks to apply the proposals.
This skill is not for broad rewrites, generic productivity advice, or creating large workflow frameworks from weak evidence.
Inputs
| Input | Default | Meaning |
|---|---|---|
TIME_WINDOW |
Last 14 days | Session or repo history to review |
FOCUS |
None | Optional project, workflow, or pain point |
TARGET_REPO |
Current repo | Repo whose AGENTS.md and skills may be affected |
APPLY_MODE |
proposal-only |
Use apply only when Zach explicitly asks |
Workflow
[1] Frame scope
-> confirm or infer the time window, focus, repo, and apply mode
[2] Gather evidence
-> inspect current AGENTS.md, README, relevant skills, recent git history,
and available session context without inventing missing history
[3] Find patterns
-> separate repeated friction, repeated success, and one-off noise
[4] Apply gates
-> use references/change-gates.md before proposing AGENTS.md or skill changes
[5] Output proposal
-> use references/retrospective-output-template.md exactly
[6] Apply only if requested
-> make narrow edits, preserve existing voice, update symlinks/README when
adding skills, then verify with git diff and repo checks
Decision Rules
- Prefer a one-line
AGENTS.mdupdate over a new skill when the lesson is a general collaboration preference. - Prefer a project-level
AGENTS.mdupdate when the lesson is tied to one repo's architecture, commands, release process, or data model. - Extract a tiny skill only when the procedure is reusable, non-obvious, and would have saved meaningful time in at least two sessions or one high-cost failure.
- If evidence is weak, say so and propose no changes.
- Scheduled automation must stay proposal-only and should optimize for low noise over completeness.
Output
For retrospective output, read and follow:
references/retrospective-output-template.mdreferences/change-gates.md
Verification
When applying changes in a skills repo:
- Run
git diff --check. - Run
.githooks/pre-commitif present. - Confirm every new skill has a README entry and flat symlinks under both
.agent/skills/and.claude/skills/when the repo uses that convention.
Source: Zach677/Zach-Skills — distributed by TomeVault.