look-back

canon look-back

Adapted from Mervin Praison's Codex Meta-Prompt (mer.vin, May 25, 2026). See CREDITS.md for the full attribution.

Mine recent work for repeated workflows and propose the smallest reusable asset: a skill, subagent, automation proposal, or extension to something that already exists.

This workflow is intentionally conservative. It should produce a shortlist before creating files, and it should skip weak candidates.

User Surface

Primary command:

/canon:look-back

Natural triggers:

• "look back over my recent work"
• "find repeated workflows"
• "what should become a skill"
• "mine my last 30 days for automations"
• "turn repeated work into reusable assets"

Evidence Order

Inspect available evidence in this order:

1. Recent canon files: MEMORY.md, ERRORS.md, and CLAUDE.md.
2. Recent Codex or Claude session summaries if locally available.
3. Existing skills, custom agents, slash commands, and automations.
4. Chronicle if enabled, discovery only.
5. Git history when useful for dates and repeated task shape.

If a source is missing, state that it was skipped and continue.

Candidate Rules

A candidate can be packaged only when it satisfies all of these:

• It occurred at least twice, or is clearly likely to recur and costly to repeat.
• It has stable inputs.
• It has a repeatable procedure.
• It has a clear output or stopping condition.
• It materially improves speed, quality, consistency, or reliability.
• It is not already adequately covered.

Default to "skip" when evidence is weak.

Smallest Form Decision Table

Form	Use When	Output Location
Skill	Reusable workflow or playbook	skills/<slug>/SKILL.md
Subagent	Bounded specialist role or investigation task	agents/<slug>.md or documented prompt
Automation	Recurring check, reminder, report, or monitor	Proposal only in v1
Extend existing	Existing skill covers 70%+ of workflow	Patch existing skill after confirmation
Skip	One-off, sensitive, vague, or poorly evidenced	Report only

Workflow

1. Read evidence in the order above.
2. Inventory existing reusable assets before proposing new ones.
3. Build a compact shortlist with evidence, dates, and recommended form.
4. Ask before editing existing non-canon files.
5. Prefer extending an existing asset over creating a duplicate.
6. Create or extend only high-confidence assets after the shortlist is accepted.
7. Finish with created items, skipped items, and needs-more-evidence items.

Shortlist Format

## Shortlist
- **<candidate>** — <skill | subagent | automation | extend existing | skip>
  - Evidence: <sources and dates>
  - Why this form: <one sentence>
  - Confidence: <high | medium | low>

## Recommended Next Step
<One small action, or "no asset yet">

Hard Rules

• Shortlist before creating.
• Extend before duplicate.
• Ask before editing existing non-canon files.
• Never create speculative broad skills.
• Never include secrets, raw transcripts, or private local paths in generated assets unless the user explicitly asks.