condense-learnings

---

name: condense-learnings description: Condense the learnings backlog: dedupe, normalize tags, archive stale entries, then caveman-compress the body of every remaining active learning. Validated against load-bearing invariants. Use when the user says "condense learnings", "clean up learnings", "deduplicate learnings", or "organize learnings". allowed-tools: [Bash, Read] model: sonnet effort: medium

Condense Learnings

Two-pass shrink of learnings/:

1. Curation — archive duplicates / stale entries, normalize tags.
2. Caveman compression — rewrite the body of each remaining active learning terse, validated against invariants. Sets compressed: true on success.

Workflow

1. Pass 1 — curation

Load all active learnings:

node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-list --json true

If empty or fewer than 2: skip Pass 1, go straight to Pass 2.

Read each learning's id, title, tags, created, and body. Identify:

• Duplicates — same root insight. Keep the one with fuller body or clearer title. Archive the other. Optionally enrich the winner from the loser's content.
• Stale / superseded — was valid once, now wrong, trivially obvious, or fully covered elsewhere. Archive.
• Tag normalization — error-handling vs errors vs error_handling → pick canonical, update all.
• Keep — non-obvious, still valid, actionable. When in doubt, keep.

Apply each action:

# Archive a learning (duplicate loser or stale entry):
node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-move {{ID}} --to archived

# Update the winner when merging (add missing insights from the loser):
cat <<'BODY' | node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-update {{ID}} --body -
{{MERGED_BODY}}
BODY

# Normalize tags:
node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-update {{ID}} --tags "{{canonical-tag1}},{{canonical-tag2}}"

Rebuild the index:

node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" index-rebuild --type learnings

2. Pass 2 — caveman compress each remaining active learning

Reload the active list (curation may have changed it):

node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-list --json true

For each learning:

1. Read the file at .backlog/learnings/{{id}}-{{slug}}/index.md with the Read tool. Skip if frontmatter has compressed: true.
2. Rewrite the body following the caveman rules below. Frontmatter (---\n…\n---) stays out of the rewrite — only the body below it.
3. Pipe to learning-compress:

cat <<'BODY' | node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-compress {{ID}} --body - --json true
{{COMPRESSED_BODY}}
BODY

Exit code 0 = success, 2 = invariant failure (JSON errors lists missing tokens). File untouched on failure.

4. On invariant failure — retry once with diagnostics in your reasoning (e.g. "validator says missing LRN-014 — preserve it byte-exact"). Second failure → log + skip. Do not delete or alter the file.

3. Caveman compression rules (Pass 2 body rewrite)

Drop

• Articles: a, an, the
• Filler: just, really, basically, actually, simply, essentially, generally
• Pleasantries: sure, certainly, of course, happy to, I'd recommend
• Hedging: it might be worth, you could consider, it would be good to
• Connective fluff: however, furthermore, additionally, in addition
• Redundant phrasing: in order to → to, make sure to → ensure
• "you should", "remember to" — state the rule directly

Preserve EXACTLY (validator enforces — byte-exact)

• Fenced code blocks (```) — every byte
• Inline code (`…`)
• URLs (https://…)
• IDs: TSK-\d{3,}, LRN-\d{3,}, PLN-\d{3,}, RPT-\d{3,}
• Dates: YYYY-MM-DD
• Version numbers: vX.Y.Z
• All markdown headings (same set, same order, exact heading text)

Compress

• Short synonyms: big not extensive, fix not implement a solution for, use not utilize
• Fragments OK
• Merge bullets that say the same thing
• Keep one example where multiple show the same pattern
• For learnings with **Why:** / **How to apply:** blocks: keep the labels, compress the prose after them

Pattern

Original: "When you are writing tests for the database layer, you should always make sure to use a real database connection rather than mocking it, because we got burned last quarter when mocked tests passed but the production migration failed."

Compressed: "DB-layer tests must use real connection, not mock. Past incident 2025-Q3: mocks passed, prod migration failed."

4. Report Summary

## Learnings Condensed

**Pass 1 — curation**
- Archived ({{N}}): LRN-XXX: {{title}} — {{reason}}
- Merged into ({{N}} winners updated): LRN-XXX ← absorbed LRN-YYY
- Tags normalized ({{N}}): `errors` → `error-handling` on LRN-XXX, LRN-YYY

**Pass 2 — caveman compress** ({{N}} touched):
- LRN-XXX: {{title}} — saved {{bytes}} bytes

**Skipped — already compressed** ({{N}})
**Failed validation** ({{N}}):
- LRN-XXX: {{first error}}

Total bytes saved: {{TOTAL}}

If nothing changed: All {{N}} learnings are already curated and compressed.

Notes

• Both passes are idempotent. Curation can run repeatedly without re-archiving the same entries; compressed: true blocks re-compress.
• Validation rejects rewrites that drop load-bearing tokens or reorder headings. Original preserved on failure.
• Pass 1 is the right place to reshape semantics; Pass 2 only compresses prose, never paraphrases away meaning.