/review

Review accumulated lessons in .autocontext/lessons.json. Use AskUserQuestion to walk through lessons.

Initialization

The script first reads .autocontext/lessons.json and separates lessons into two groups:

• Active lessons — deleted: false or not marked deleted
• Tombstoned lessons — deleted: true (lessons that were intentionally removed)

If .autocontext/cache/curated-pending.json exists (from previous curation sessions with pending items), those are presented first for final approval.

Review order

Active lessons are sorted by confidence score (lowest first). Low-confidence lessons need the most attention — these are either new, frequently contradicted, or validated by few developers.

Lesson presentation

Lessons are presented in batches of 3-4 using AskUserQuestion. For each lesson, the review shows:

• Text — the lesson content
• Category — lesson type (bug-fix, architecture, workflow, pattern, gotcha, etc.)
• Confidence — current score (0.0 to 1.0)
• Validated count — how many sessions/developers have confirmed it
• Created by — which developer/session discovered it
• Age — how long ago it was created

Example format:

Lesson: "Always run git pull before pushing to main"
Category: workflow
Confidence: 0.9 | Validated: 12 times | Created by: alice | Age: 45 days

Review actions

For each lesson, you choose one action:

Options:

• Approve — increase confidence by +0.2 (confirms lesson is valuable)
• Edit — modify lesson text, category, or tags
• Delete — tombstone the lesson (marks as deleted: true)
• Supersede — replace this lesson with a newer/better version
• Skip — leave unchanged and move to next

Approve

Bumps confidence by 0.2. Use this when you confirm a lesson is still valid and useful.

Edit

Opens the lesson for modification. You can update:

• Lesson text (the knowledge itself)
• Category (bug-fix, architecture, workflow, pattern, gotcha, integration, test-strategy)
• Tags (optional metadata)

Edited lessons are automatically marked as reviewed by you.

Delete

Marks the lesson as tombstoned (deleted: true). The original content is preserved in .autocontext/archive/superseded.json for historical reference, but the lesson won't load during future sessions.

Use this for lessons that are outdated, wrong, or no longer relevant.

Supersede

Replace the current lesson with a completely new version. You provide the new lesson text. The old lesson is tombstoned and the new one is added as a fresh, high-confidence lesson.

Use this when a lesson is correct in spirit but the specific guidance has changed (e.g., "Use React class components" → "Use React functional components with hooks").

Skip

Leaves the lesson unchanged and moves to the next one. Use this when you're not sure or want to review later.

Promote to global (skill-tagged lessons only)

When a lesson has a skill field (non-null) and skill_learning.enabled is true in config, a sixth action is available:

• Promote to global — Copy this lesson to the global skill lesson store using the store module. Update the original lesson's scope to "skill" in lessons.json. Report: "Promoted to global store for [skill name]."

The global store path comes from config: skill_learning.global_store (default ~/.claude/skill-lessons/).

To perform the promotion, run:

python3 -c "
import sys
sys.path.insert(0, '${CLAUDE_PLUGIN_ROOT}/scripts/skill-evolution')
from store import promote_lesson, ensure_store
import json

ensure_store()
lesson = json.loads('''LESSON_JSON_HERE''')
result = promote_lesson(lesson, 'SKILL_NAME_HERE', 'PROJECT_NAME_HERE')
print('promoted' if result else 'already_exists')
"

Only show this action when skill_learning.enabled is true in config and the lesson has a non-null skill field.

Tombstoned lessons

After reviewing all active lessons, you'll be asked about any tombstoned lessons:

Question: There are N tombstoned lessons. What would you like to do?

Options:

• Remove all tombstones permanently — deletes archived lessons (cannot be undone)
• Let me review them individually — presents each tombstone for potential restoration or permanent removal
• Keep them — preserves tombstones for historical reference

If you choose to review individually, each tombstoned lesson is presented with options to:

• Restore — mark as deleted: false and add back to active lessons
• Permanently remove — delete from archive (cannot be undone)
• Skip — leave in archive

Playbook regeneration

After all reviews are complete, the script automatically regenerates .autocontext/playbook.md using:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/generate-playbook.py .autocontext/lessons.json .autocontext/playbook.md

The playbook reflects your curated lessons and is the human-readable summary of project knowledge.

Review summary

At the end, you'll see a report:

Review complete:
- N lessons reviewed
- M lessons approved (confidence increased)
- K lessons deleted (tombstoned)
- J lessons edited
- Playbook regenerated with X active lessons

This summary confirms what changed and gives you confidence that the curation was applied.

Tips

• Low-confidence lessons first — the review order (lowest confidence first) helps you focus on lessons that need validation
• Be liberal with approval — if a lesson is still correct, approve it. This builds signal about what's truly useful.
• Supersede instead of delete — if a lesson is mostly right but needs an update, supersede it rather than deleting. This preserves the learning path.
• Regular curation — run /autocontext-review weekly or after major changes to keep lessons fresh and accurate