consolidate-learnings

Consolidate — Cross-Feature Learning Promotion

Per-feature LEARNINGS files are write-only until something promotes their recurring signal to a level future features actually read. That is this skill. You read the LEARNINGS accumulated since the last consolidation pass and promote patterns (not one-offs) into project-level context.

Scope: Read LEARNINGS files; promote durable patterns into docs/_constraints/ and docs/_decisions/; propose skill-prompt changes as a reviewable artifact. You do NOT autonomously edit skill files, and you do NOT promote a signal seen in only one feature unless it is unambiguously a standing constraint.

Cadence: Manually invoked. A SessionStart nudge hook reminds when ≥5 un-consolidated LEARNINGS exist, but you may be run any time.

---

How to Invoke

/pdlc:consolidate-learnings

No argument — operates across the whole repo's docs/.

---

Git Workflow

1. Before starting: check out or create chore-consolidate-learnings-{date}. Pull latest.
2. After completing: commit promoted constraints/decisions, the proposal artifact, and the updated log; push. Open for human review — do not merge skill-prompt changes yourself.

---

Consolidation Process

1. Find the boundary. Read docs/_decisions/.consolidation-log.md (create it if absent). Every docs/*/LEARNINGS-*.md with a Date Completed after the last logged pass is in scope.
2. Read them all. Collect §2 Cross-Feature Patterns, §4 Process Learnings, and §5 Open Items from each.
3. Cluster. Group items pointing at the same underlying concern across features.
4. Distinguish pattern from coincidence. Two features mentioning the same concern might be a pattern or might be coincidence. Promote only when the signal is durable: it recurs across ≥2 unrelated features, or a single occurrence states a standing invariant (security, data integrity, regulatory) that obviously generalizes. When in doubt, leave it in the proposal for a human to judge — do not promote.
5. Route each promoted item:
   • Domain invariant future REQs must respect → append to docs/_constraints/DOMAIN-CONSTRAINTS.md (read by pm-author).
   • Architectural decision now project-level → docs/_decisions/DECISIONS-{topic}.md (read by se-author).
   • Process learning about a skill prompt / review checklist / workflow phase → propose, never apply (next section).
6. Record the pass in .consolidation-log.md: date, which LEARNINGS files were consumed, what was promoted, what was deferred.

---

Skill-Update Proposals (human-in-the-loop)

When a learning says a skill prompt itself should change, you do not edit the skill. Write docs/_decisions/CONSOLIDATION-PROPOSAL-{date}.md:

# Consolidation Proposal — {date}

| Source LEARNINGS | Target skill | Proposed change | Rationale |
|---|---|---|---|
| {features} | pdlc/skills/{skill}/SKILL.md | {concrete edit} | {why, and the pattern that motivates it} |

A human reviews and applies (or rejects) each row. This is deliberate: agents proposing changes to the prompts that govern agents must pass through human judgment.

---

Output Formats

docs/_constraints/DOMAIN-CONSTRAINTS.md entries:

## DC-{NN}: {Constraint title}
**Constraint:** {what must hold for all future features in this domain}
**Origin:** promoted {date} from LEARNINGS of {features}
**Applies to:** {REQ authoring / TSPEC / testing / all}

docs/_decisions/DECISIONS-{topic}.md reuses the per-feature DECISIONS format (DEC-{topic}-{NN}), with an Origin line citing the source features.

---

Quality Checklist

• Every LEARNINGS file since the last logged pass was read
• Nothing promoted on a single coincidental mention (pattern ≥2, or clear standing invariant)
• Skill-prompt changes are proposed in CONSOLIDATION-PROPOSAL, never auto-applied
• .consolidation-log.md updated with date, consumed files, promoted + deferred items
• Promotions cite their origin features

---

Communication Style

• Lead with what you promoted and why it cleared the pattern-vs-coincidence bar.
• Be explicit about what you deferred and what a human needs to decide.
• Tables for promotions and proposals.