Crystallize — distil session patterns into reusable skills
Invoke when the CTO says /crystallize, "crystallize", "extract knowledge", or
"what have we learned?". Also auto-suggested when session count is a multiple
of 10 (the session-end hook checks .great_cto/.last-crystallize).
The knowledge-extractor agent (Opus) does the heavy lifting. This skill
orchestrates the workflow and emits the final report.
Session-end hint integration: The session-end hook checks
.great_cto/.last-crystallize and suggests running /crystallize when the
session count exceeds last_sessions + 10. Run this skill after ≥10 sessions
to keep extracted skills current.
Step 1 — Gather raw material
# Count sessions
SESSION_COUNT=$(ls .great_cto/logs/session-*-end.md 2>/dev/null | wc -l | tr -d ' ')
echo "Sessions: $SESSION_COUNT"
# Read lessons
cat .great_cto/lessons.md 2>/dev/null || echo "(no lessons yet)"
# Read cross-project decisions
cat ~/.great_cto/decisions.md 2>/dev/null | head -200 || echo "(none)"
# Find patterns that appear in ≥3 sessions
grep -h "^## pattern:" .great_cto/logs/session-*-end.md 2>/dev/null | sort | uniq -c | sort -rn | head -20
# Recent git log for context
git log --oneline --since="30 days ago" | head -30
If SESSION_COUNT is 0, tell the CTO: "No session logs found in
.great_cto/logs/. Run at least 10 sessions before crystallizing." Exit.
If SESSION_COUNT < 10, tell the CTO: "Only {N} sessions found. Patterns
are more reliable after ≥10 sessions. Proceed anyway? [yes/no]" Wait for
confirmation before continuing.
Step 2 — Cluster patterns (via knowledge-extractor agent)
Spawn the knowledge-extractor agent with the gathered data as context:
Agent: knowledge-extractor
Task: |
Read .great_cto/lessons.md and all files in .great_cto/logs/.
Cluster lesson entries by pattern slug.
For each cluster with ≥3 occurrences, write a draft skill file to
skills/{domain}/SKILL.md (status: draft in frontmatter).
If a skill for that domain already exists, append a new ## section instead
of replacing the file.
Infer domain from the pattern slug and its archetype tags.
Return a structured summary: clusters found, drafts written, already-covered.
Wait for the agent to complete before proceeding to Step 3.
Step 3 — Emit promotion report
After the agent completes, print:
CRYSTALLIZE REPORT
════════════════════════════════════════
Sessions analysed: {SESSION_COUNT}
Lessons found: {LESSON_COUNT}
Clusters: {CLUSTER_COUNT}
Draft skills: {DRAFT_COUNT} (in skills/{domain}/SKILL.md)
Already covered: {COVERED_COUNT} (pattern already in existing skill)
════════════════════════════════════════
Draft files:
{list of paths and brief description per draft}
Next: review drafts, remove `status: draft` when satisfied.
Run /crystallize again after 10 more sessions.
════════════════════════════════════════
Step 4 — Write .last-crystallize marker
After emitting the report, write the marker file:
SESSION_COUNT=$(ls .great_cto/logs/session-*-end.md 2>/dev/null | wc -l | tr -d ' ')
DRAFT_COUNT={P} # from agent output
mkdir -p .great_cto
node -e "
const fs = require('fs');
fs.writeFileSync('.great_cto/.last-crystallize', JSON.stringify({
ts: new Date().toISOString(),
sessions: parseInt('$SESSION_COUNT') || 0,
drafts: parseInt('$DRAFT_COUNT') || 0
}) + '\n');
"
Step 5 — Auto-run cadence suggestion
If SESSION_COUNT is a multiple of 10 (and > 0), append to the report:
Auto-suggestion: you've completed {SESSION_COUNT} sessions. Consider running
`/crystallize` every 10 sessions to keep skills current.