Learning To Rule Or Skill

learning-to-rule-or-skill

When to use

Use this skill when:

• A repeated mistake appears across multiple tasks
• A successful pattern should be reused in the future
• A new constraint or workflow should be captured permanently
• Reviewing post-task learnings or retrospectives
• Deciding whether a learning belongs in a rule or a skill
• After completing a task — reflecting on what worked or caused friction
• Mining the audit log (agents/runtime/state/audit/<YYYY-MM>.jsonl, audit-log-v1) surfaced a repeated phase pattern via extract_audit_patterns.ts — the pattern's count ≥ 2 already satisfies the repetition gate
• A comparative analysis handed over a ### SEED — block from its ## Seeds section (/analyze:repo § 6). This is an accepted intake shape: the block already carries the proposed name, kind, target template, one-line description, and the Extends instead? answer, so step 4's overlap search starts from a stated candidate rather than a blank page — it does not skip that search, and a seed whose Extends instead? line is missing is incomplete intake, not a shortcut. The repetition gate is satisfied differently here: the evidence is the anchored ADOPT/ADAPT verdict behind the seed rather than a recurrence count — a mechanism confirmed once against a named own-repo defect is not the one-off case the gate exists to filter.

Do not use this skill when:

• The learning is one-off and unlikely to repeat
• The issue is too vague to turn into guidance
• The content already exists in an equivalent rule or skill

Goal

• Turn repeated learnings into reusable project guidance
• Decide correctly between rule, skill, or update to an existing one
• Prevent the same mistakes from happening again
• Keep the system small, clear, and non-duplicative

Preconditions

• A concrete learning, mistake, or successful pattern exists
• The learning can be stated clearly in one or two sentences
• Existing rules and skills can be checked for overlap

Decision matrix

Decision hints

• Same issue appeared at least twice → strongly consider codifying
• One-off or too narrow → do not create anything yet
• Standard tool knowledge (jq, docker, git basics) → Nothing — the model knows this
• If unsure between skill and guideline: does it need step-by-step decisions? → Skill. Just conventions? → Guideline
• If unsure between guideline and pattern: is it "how to write code in X" (prose convention)? → Guideline. Is it "I have problem P — here's the proven recipe + a verification step + how reliable it is"? → Pattern
• When a related workflow fires, surface a matching src/patterns/ recipe (never auto-apply) — the human decides whether to use it

Procedure

0. Promotion Gate (mandatory)

Before proceeding, the learning MUST pass all gates:

If ANY gate fails → stop. Do not create or update anything. Memory-entry rejection criteria live in /memory:add (see § "When to skip").

1. State the learning clearly

Write as a concrete sentence.

Good:

• Nested triple backticks break copyability in generated markdown
• Route inspection is more reliable via JSON and jq than text parsing

Bad:

• Markdown is annoying
• Routing was confusing

2. Identify the pattern type

Classify by type:

• Constraint
• Workflow
• Anti-pattern
• Quality check
• Environment-specific convention

Tag with feedback category (for tracking/aggregation):

• skill-weakness — skill gave wrong or incomplete guidance
• rule-weakness — rule was too vague or missed a case
• routing-issue — wrong skill was selected
• assumption-issue — agent made bad assumptions
• verification-gap — verification step was missing or weak
• optimization-overreach — optimize command suggested harmful change

3. Decide the target

Choose one:

• New rule (always-true constraint)
• Update existing rule
• New skill (step-by-step workflow)
• Update existing skill
• New guideline (coding convention / reference)
• Update existing guideline
• Nothing (baseline knowledge, standard tool usage, one-off)

4. Check for overlap — search protocol (mandatory)

A grep that returns zero hits is not proof of no overlap. Knowledge in this package is distributed across five surfaces — skills/, rules/, docs/guidelines/, commands/, src/patterns/. Skip any of them and recall drops. Run all the steps before declaring "no overlap":

Step 1 — list all five surfaces. Directory taxonomy is free evidence:

ls src/skills/ \
   src/rules/ \
   docs/guidelines/ \
   dist/agent-src/commands/ \
   src/patterns/

Sub-directories matter — docs/guidelines/php/patterns/, docs/guidelines/agent-infra/, etc. carry topic taxonomies a flat file scan misses. Always descend one level.

Step 2 — grep with both vocabularies. Search for solution-words and problem-words. Solution-only grep is confirmation bias — the existing artifact may name the symptom, not the cure.

grep -rl -E "<solution-words>|<problem-words>" src/ docs/guidelines/

Step 3 — taxonomy scan. For any topic with a likely sub-folder (patterns/, php/, laravel/, agent-infra/), ls that folder before reading any file. Filename alone often answers the overlap question.

Step 4 — sample, do not just list. On any keyword overlap from steps 2–3, open and skim the 3 nearest matches — read § headings, the "When to use" / "Overview" block, and the examples list. Listing filenames is not enough; semantic overlap hides behind unrelated keywords.

Only after all four steps return clean → declare "no overlap" and proceed. Citation in the proposal: "Reviewed before drafting: " — this is the audit trail § 0's Non-duplication gate verifies against.

→ When the parent task is "create a new artifact", artifact-drafting-protocol Phase B (Research) requires this same protocol — single source of truth.

5. Draft the content

If rule:

• Short, durable constraint
• General, clear, always applicable

If skill:

• Focused workflow
• Include: When to use, Procedure, Output format, Gotchas, Do NOT

6. Validate usefulness

• Will this improve future outputs?
• Specific enough to act on?
• Different from existing guidance?
• Likely to matter more than once?

7. Keep only the smallest effective change

Prefer:

• Update over duplicate
• Small focused skill over broad skill
• Short rule over long rule

8. Write the proposal (if scope ≠ nothing)

The output of this skill is a curated proposal under agents/proposals/<proposal_id>.md, using the template at .augment/templates/agents/proposal.example.md (shipped by the package). This is the input to the five-stage pipeline (capture → classify → propose → gate → upstream); see self-improvement-pipeline.

Mandatory fields the draft MUST fill:

• proposal_id — stable kebab-case slug, unique in this repo
• type — rule | skill | command | guideline
• scope — project (stays in agents/overrides/) or package (contributed upstream via upstream-contribute)
• source_learning — path to the agents/learnings/<date>-<slug>.md file this proposal was captured from
• evidence — at least two independent references (PR, issue, incident, review-comment, test-failure, or audit-log line ids per audit-log-v1); entries that all resolve to the same PR or the same audit-log run_id are rejected by the gate (independence floor)
• Proposed artefact (§4) — the full draft body, no TODO / TBD
• Success signal (§7) — one metric, one baseline, one target, one evaluation date

Run ./agent-config proposal:check agents/proposals/<id>.md before handing to upstream-contribute. The gate is hard: non-zero exit = the proposal does not move to stage gated.

9. Degrade the source to a pointer (when promoted from a knowledge page)

If the learning came from an agents/knowledge/procedures/skill-candidates.md entry or any other agents/knowledge/ page (see § "Knowledge-candidate learnings" below), promotion is not complete until the source is degraded to a pointer — this is what prevents double-maintenance between the knowledge page and the new artifact:

./scripts-run src/scripts/degrade_to_pointer \
    --source "agents/knowledge/procedures/skill-candidates.md#<topic>" \
    --artifact "<promoted-artifact-path>" \
    --date "<YYYY-MM-DD>"

This rewrites the source entry to Promoted to <artifact> on <date>; see <path>. and regenerates agents/knowledge/INDEX.md. Skip this step only when the learning's source was NOT a knowledge page (e.g. a direct user request, an audit-log pattern, or a one-off observation).

Output format

For the decision step (what this skill prints to the user):

1. Learning summary
2. Decision: rule, skill, update, or no action
3. Rationale in one to three lines
4. If decision ≠ "no action": path of the written proposal (agents/proposals/<proposal_id>.md) and gate status (./agent-config proposal:check exit 0 = ready for review)

The proposal file itself follows proposal.example.md verbatim — all ten sections, YAML frontmatter complete, draft body in §4.

Core rules

• Capture repeated patterns, not random observations
• Prefer updating existing guidance over creating duplicates
• Rules are for durable constraints
• Skills are for repeatable workflows
• Keep new guidance as small and focused as possible

Gotchas

• Model tends to create new files when a small update is enough
• Model tends to turn vague frustrations into bad guidance
• Model may write documentation instead of reusable instructions
• Over-capturing weak learnings creates noise and reduces quality

Do NOT

• Do NOT create a rule or skill for one-off problems
• Do NOT duplicate existing guidance
• Do NOT create broad "catch-all" skills
• Do NOT write vague learnings without a concrete behavioral consequence

Auto-trigger keywords

• learning
• retrospective
• repeated mistake
• recurring issue
• create rule from learning
• create skill from learning
• codify this
• capture this pattern
• after task
• what did we learn
• post-mortem

Anti-patterns

• Creating a skill for every minor annoyance
• Rule that says "be careful" without concrete constraint
• Skill that duplicates an existing one with slightly different wording
• Capturing a learning before it has repeated

Examples

Learning: "Nested triple backticks broke markdown copyability twice this week." Decision: Update existing markdown rule + add markdown-safe skill if none exists.

Learning: "Route checks keep failing when done via text parsing." Decision: Create focused skill for Laravel route inspection via JSON and jq.

Learning: "I forgot to run PHPStan once." Decision: No action — one-off, already covered by verify-before-complete rule.

Learning: "We re-invented a per-format PDF extractor in three different analysis skills." Decision: Update the affected skills to dispatch to markitdown instead of writing new extractors. Non-text ingestion (PDF / DOCX / XLSX / PPTX / image / audio) goes through the upstream markitdown-mcp server first; only write a custom extractor if markitdown cannot handle the format and the gap is documented in its skill body.

Audit-derived learnings (optional source)

When the input is a pattern surfaced by extract_audit_patterns.ts mining agents/runtime/state/audit/<YYYY-MM>.jsonl (audit-log-v1):

1. Treat the script's pattern record as the State the learning step input (§1) — pattern.summary is the one-sentence statement, pattern.line_ids is the evidence.
2. The repetition gate is already satisfied for count ≥ 2. Skip to §3 (decide the target) — overlap check (§4) and proposal draft (§8) remain mandatory.
3. Independence floor still applies: two line ids from the same run_id count as one piece of evidence. The mining script already de-duplicates by run_id; the gate trusts that output.
4. Audit-derived proposals MUST set source_learning: agents/runtime/state/audit/<YYYY-MM>.jsonl#<line_ids> and link the mining-script run id, so the human reviewer can reproduce the pattern from the raw audit log.

Knowledge-candidate learnings (optional source)

When the input is an entry from agents/knowledge/procedures/skill-candidates.md that reached candidate status (≥ 3 mentions across distinct sessions — see update_skill_candidates.ts and the memory-consolidation recurrence step):

1. Treat the candidate's topic as the State the learning step input (§1). The repetition gate (§0) is already satisfied by the ≥ 3-mention threshold — skip straight to §3 (decide the target). Overlap check (§4) and proposal draft (§8) remain mandatory.
2. Ground the draft in the original session logs, never the consolidated summary alone. The candidate entry's Sessions: list names the chat-history session ids the topic recurred in — read those from the host's native transcript view, or reconstruct them manually from the retained continuity records, before drafting. A skill written purely from the recurrence-counter summary tends to be plausible but wrong; the original transcripts carry the detail the summary dropped.
3. Set source_learning to the knowledge-candidate entry (agents/knowledge/procedures/skill-candidates.md#<topic>).
4. §9 (degrade the source) is mandatory for this source type — the whole point of the candidate counter is that it stops counting once promoted.

Environment notes

Prefer updating existing rule/skill when possible. Create new files only when the learning introduces a clearly distinct pattern.