Kn Extract

Extracting Knowledge (with Compounding)

Announce: "Using kn-extract to extract knowledge."

Core principle: EXTRACT PATTERNS + DECISIONS + FAILURES → COMPOUND LEARNINGS.

Inputs

• Usually a completed task ID
• Sometimes a code change, repeated pattern, or recurring support issue
• Optional: --compound flag for full 3-category analysis
• Optional: --consolidate flag to review and consolidate all existing learnings

Mode Detection

Check $ARGUMENTS:

• Contains --consolidate → Go to "Consolidation Mode" section
• Otherwise → Continue with normal extraction flow

Extraction Rules

• Extract patterns, decisions, AND failures — not just code patterns
• Prefer updating an existing doc over creating a duplicate
• Link the extracted knowledge back to the source task or source doc
• Only create a template if the pattern is genuinely reusable for generation
• Promote critical learnings to learnings/critical-patterns for future kn-init sessions

Step 1: Identify Source

mcp__knowns__get_task({ "taskId": "$ARGUMENTS" })

Look for three categories:

Step 2: Search for Existing Docs

mcp__knowns__search({ "query": "<pattern/topic>", "type": "doc" })

Don't duplicate. Update existing docs when possible.

Step 3: Three-Category Analysis

3a. Patterns

Identify reusable patterns:

• Code patterns: new utilities, abstractions worth standardizing
• Architecture patterns: structural decisions that worked
• Process patterns: workflow approaches that saved time

3b. Decisions

Identify significant decisions:

• GOOD_CALL: decisions that proved correct or saved time
• BAD_CALL: decisions that required rework
• SURPRISE: things that turned out differently than expected
• TRADEOFF: conscious choices where alternatives were considered

3c. Failures

Identify failures and wasted effort:

• Bugs introduced and root causes
• Wrong assumptions that required backtracking
• Missing prerequisites discovered mid-execution
• Test gaps that allowed regressions

Step 4: Create/Update Documentation

For patterns → pattern doc (same as before)

mcp__knowns__create_doc({
  "title": "Pattern: <Name>",
  "description": "Reusable pattern for <purpose>",
  "tags": ["pattern", "<domain>"],
  "folder": "patterns"
})

For decisions + failures → learning doc

mcp__knowns__create_doc({
  "title": "Learning: <feature/domain>",
  "description": "Learnings from <task/feature>",
  "tags": ["learning", "<domain>"],
  "folder": "learnings",
  "content": "<see template below>"
})

Learning doc template:

## Patterns

### <Pattern Name>
- **What:** <description>
- **When to use:** <applicable conditions>
- **Source:** @task-<id>

## Decisions

### <Decision>
- **Chose:** <what was chosen>
- **Over:** <what was rejected>
- **Tag:** GOOD_CALL / BAD_CALL / SURPRISE / TRADEOFF
- **Outcome:** <how it played out>
- **Recommendation:** <for future work>

## Failures

### <Failure>
- **What went wrong:** <description>
- **Root cause:** <not just symptom>
- **Time lost:** <estimate>
- **Prevention:** <what to do differently>

Step 5: Save to Memory

For each extracted pattern or decision worth quick recall, save a concise memory entry alongside the doc:

mcp__knowns__add_memory({
  "title": "<pattern/decision name>",
  "content": "<2-3 sentence summary>. Full reference: @doc/<path>",
  "layer": "project",
  "category": "<pattern|decision|convention|failure>",
  "tags": ["<domain>"]
})

Memory = fast agent recall in future sessions. Doc = full structured reference. Do NOT duplicate the entire doc content — store a summary and link to the doc. Skip this step if the extraction produced nothing generalizable.

Step 6: Create Template (if code-generatable)

mcp__knowns__create_template({
  "name": "<pattern-name>",
  "description": "Generate <what>",
  "doc": "patterns/<pattern-name>"
})

Step 7: Promote Critical Learnings

For any finding that meets ALL criteria:

• Affects more than one future feature
• Would cause ≥30 minutes wasted effort if unknown
• Is generalizable, not implementation-specific

Check if critical-patterns doc exists:

mcp__knowns__search({ "query": "critical patterns", "type": "doc", "tag": "learning" })

If exists — append:

mcp__knowns__update_doc({
  "path": "learnings/critical-patterns",
  "appendContent": "\n\n## [Date] <Learning Title>\n**Category:** pattern / decision / failure\n**Source:** @task-<id>\n**Tags:** [tag1, tag2]\n\n<2-4 sentence summary and what to do differently>\n\n**Full entry:** @doc/learnings/<slug>"
})

If not exists — create:

mcp__knowns__create_doc({
  "title": "Critical Patterns",
  "description": "Promoted learnings that save the most time. Read at session start.",
  "folder": "learnings",
  "tags": ["learning", "critical"],
  "content": "# Critical Patterns\n\nPromoted learnings from completed work. Read this at the start of every session via `kn-init`. These are lessons that cost the most to learn and save the most by knowing.\n\n---"
})

Calibration: Do NOT promote everything. If critical-patterns grows past 20-30 entries it becomes noise. Only promote learnings that would have saved ≥30 minutes if known in advance.

Step 8: Validate

CRITICAL: After creating doc/template, validate to catch broken refs:

mcp__knowns__validate({ "entity": "<doc-path>" })

If errors found, fix before continuing.

Step 9: Link Back to Task

mcp__knowns__update_task({
  "taskId": "$ARGUMENTS",
  "appendNotes": "📚 Extracted to @doc/<path>"
})

Consolidation Mode (Dream Lite)

When $ARGUMENTS contains --consolidate:

Announce: "Using kn-extract --consolidate to review and consolidate learnings."

Scan all existing learnings docs, merge duplicates, flag outdated entries, and promote new critical patterns. Run on-demand when the learnings folder feels messy or after a batch of completed work.

C-Step 1: Scan All Learnings

mcp__knowns__list_docs({ "tag": "learning" })

Read each learning doc:

mcp__knowns__get_doc({ "path": "<path>", "smart": true })

C-Step 2: Identify Issues

For each learning doc, check:

Duplicates

• Two docs covering the same root cause or pattern
• Same advice phrased differently across docs
• → Merge into one, delete the other

Outdated

• Fix/pattern references code that no longer exists
• Advice contradicts current architecture or conventions
• → Update or mark as outdated with date

Missing Promotions

• Learning that meets critical criteria (affects multiple features, saves ≥30 min) but isn't in critical-patterns
• → Propose promotion

Orphaned

• Learning that references a task or doc that no longer exists
• → Fix ref or note the context is lost

C-Step 3: Apply Changes

For each issue found, present to user:

Consolidation findings:

1. MERGE: "Learning: auth token" + "Learning: JWT refresh" → same root cause
   → Merge into "Learning: auth token handling"?

2. OUTDATED: "Learning: webpack config" — references webpack.config.js which was removed
   → Mark outdated or delete?

3. PROMOTE: "Learning: Go test race conditions" — saved 2h on 3 separate tasks
   → Promote to critical-patterns?

4. ORPHAN: "Learning: SSE reconnect" — references @task-abc123 which doesn't exist
   → Keep content, remove broken ref?

Apply all? (yes / review each / skip)

If "review each": present one at a time, apply user's choice. If "yes": apply all suggested changes.

C-Step 4: Report

Consolidation complete:
- Merged: X docs
- Updated: X docs
- Promoted: X to critical-patterns
- Orphans fixed: X
- Total learnings: X docs

Shared Output Contract

Required order for the final user-facing response:

1. Goal/result — what knowledge was extracted, updated, or intentionally not extracted.
2. Key details — docs created/updated, critical promotions, template status, validation.
3. Next action — recommend a concrete follow-up command only when a natural handoff exists.

For kn-extract, the key details should cover:

• what knowledge was extracted (patterns, decisions, failures)
• whether docs were created or updated
• whether critical learnings were promoted
• whether a template was created
• where the canonical knowledge now lives

When the extraction leads to a clear follow-up, include the best next command. If the correct outcome is a no-op or a completed doc update with no obvious continuation, stop after the result and key details.

No-Op Case

If the work is too specific to generalize, say so explicitly and do not force a new doc.

Do NOT fabricate findings. If the task ran smoothly with no surprises, write that. A short learning with 2 genuine entries is better than a long doc with invented ones.

What to Extract

Checklist

• Three categories analyzed (patterns, decisions, failures)
• Knowledge is generalizable
• Includes working example (for patterns)
• Links back to source
• Critical learnings promoted (if applicable)
• Template created (if applicable)
• Validated (no broken refs)

Red Flags

• Only extracting code patterns, ignoring decisions and failures
• Promoting everything as critical (noise kills the learning loop)
• Writing generic learnings ("test more carefully" is worthless)
• Fabricating findings when the task was straightforward
• Not checking existing docs before creating duplicates

Converted and distributed by TomeVault — claim your Tome and manage your conversions.