ai-crystallize

AI Crystallize

Autonomous companion to the human crystallization path. Each run synthesizes converged thinking topics into standalone notes written to a dedicated, segregated folder - never Brain/02-Permanent/.

Why this exists: Converged topics pile up faster than a human can crystallize them. This playbook captures the AI's synthesis of each converged topic so the conclusion is not lost - while preserving the one boundary the second brain must never cross autonomously: nothing enters the endorsed permanent knowledge base without an explicit endorsement act by the user.

The boundary, stated plainly:

• This playbook writes ONLY to Brain/05-Meta/AI Crystallizations/, tagged provenance: ai-inferred, excluded from semantic indexing.
• It NEVER writes to Brain/02-Permanent/, Brain/AI Extracted Notes/, or Brain/05-Meta/Reports/.
• It NEVER changes a topic's status in the Thinking Registry (the topic stays converged), so /manage-thinking-topics crystallize remains fully available to the user for the real, endorsed crystallization.
• Its output is a candidate for human review, not a permanent note.

Design principle

Automate the synthesis, segregate the artifact, gate the endorsement. The AI may draft what it concluded; only the human may promote a conclusion into the knowledge they will reason FROM.

State Dependencies

Prerequisites

• Thinking Registry exists at Brain/05-Meta/Thinking/THINKING-REGISTRY.md
• Output folder Brain/05-Meta/AI Crystallizations/ exists (create if missing)
• At least one topic with status: converged (otherwise clean exit)

Configuration

• Per-run cap: process at most 5 un-crystallized converged topics per run (reliability + 45-minute budget). Remaining topics are picked up on the next run. The cap is a soft limit; if fewer than 5 remain, process all.

---

Process

Step 1: Get Date and Load Converged Topics

date '+%Y-%m-%d'

Read Brain/05-Meta/Thinking/THINKING-REGISTRY.md. Collect every topic listed in the Converged (Ready for Crystallization) table (or any topic whose thinking-file frontmatter is status: converged).

If a specific [topic-slug] was passed as argument, restrict to that single topic (verify it is converged; if not, log and exit).

If there are no converged topics, write a brief changelog entry and exit cleanly.

Step 2: Deduplicate (Idempotency)

For each converged topic, check whether Brain/05-Meta/AI Crystallizations/[topic-slug].md already exists.

• Exists -> already AI-crystallized; skip.
• Does not exist -> eligible for this run.

This file-existence check is the dedup mechanism. The playbook never mutates the Thinking Registry or thinking files, so it is fully idempotent and safe to retry.

Step 3: Select the Batch

From the eligible (un-crystallized) converged topics, take up to the per-run cap (5), oldest-converged first where the converged date is available; otherwise registry order. Convergence-date proxy: thinking files carry no converged_on frontmatter field, so use the frontmatter updated: date (the converging run stamps it) as the oldest-first sort key - cheap to extract in one pass with grep -m1 '^updated:' across the Thinking folder. Note in the run log any eligible topics deferred to a later run.

Step 4: Extract the Converged Conclusion

For each selected topic, read Brain/05-Meta/Thinking/[topic-slug].md and extract the conclusion. Prefer the canonical block, fall back gracefully:

1. Look for ## Final Synthesis - the canonical block appended by the incubation loop at convergence. If present, this is authoritative. Extract: Conclusion, Confidence + leading hypothesis, Key evidence, Convergence path, Remaining uncertainty, Related topics. Note: legacy topics instead carry an inline **Final Synthesis:** / **Final converged position:** line (not a ## header) near the end - a grep '## Final Synthesis' will miss these. Match the lowercase inline form too, or just go straight to the tail per step 2.
2. If absent (legacy topics), read only the last ~120 lines of the file (some are 5,000+ lines). Extract: central question (frontmatter topic), the trailing "Final Synthesis" / "Final converged position" / "Current Best Answer", the last "Updated Hypotheses" table, and any "Open Questions". The convergence-assessment block (run count, leading-hypothesis streak, confidence delta) sits just above and gives the Convergence Path.
3. Read the full file only if the last 200 lines contain no conclusion.

Do not run KB search or web search - this playbook synthesizes from the already-converged reasoning, not new evidence. Keep it fast and self-contained.

Step 5: Draft the AI Crystallization Note

Write to Brain/05-Meta/AI Crystallizations/[topic-slug].md:

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
created_by: [model-name]
updated_by: [model-name]
agent_version: 01.25
provenance: ai-inferred
source_topic: "[[topic-slug]]"
status: ai-crystallized
---

> 🤖 **AI Crystallization - not endorsed knowledge.** Synthesized autonomously from a converged thinking topic. Carries `provenance: ai-inferred` and lives outside `Brain/02-Permanent/`. This is a **candidate for human review**, not a permanent note. To promote it, the user runs `/manage-thinking-topics crystallize [topic-slug]`.

# [Topic Title]

**Central Question:** [exact question from the thinking file]

## Conclusion
[1-2 paragraphs - the substantive answer, in neutral third-person AI-synthesis voice. NOT first-person your voice - this is ai-inferred, not his endorsed thinking.]

## Confidence
[X% on H[N] (and any co-leading hypotheses)]

## Key Evidence
- [[Note A]] - [why it mattered]
- [[Note B]] - [why it mattered]
- [[Note C]] - [why it mattered]

## Convergence Path
[Standard | Framework exhaustion] after [N] runs

## Remaining Uncertainty
[1 paragraph - what is known to be unknown; what could change this]

## Source
Crystallized from [[topic-slug]] (`Brain/05-Meta/Thinking/[topic-slug].md`), converged [date].
[Related topics, if any: [[other-slug]]]

Voice discipline: write the Conclusion in neutral analytical voice, never first-person as the user. This artifact is ai-inferred; it must never read as if the user authored or endorsed it. Per CLAUDE.md's guarded boundary, only an explicit endorsement act by the user (via manual crystallization) may set provenance to endorsed or originated.

Step 6: Update the Crystallization Index

Append a row to Brain/05-Meta/AI Crystallizations/INDEX.md (create the file with a header if missing):

| [topic-slug] | [one-line conclusion summary] | [confidence] | YYYY-MM-DD |

Step 7: Write Session Changelog

Write to Brain/05-Meta/Changelogs/CHANGELOG - AI Crystallize YYYY-MM-DD.md:

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
created_by: [model-name]
updated_by: [model-name]
agent_version: 01.25
---

# AI Crystallize Session: YYYY-MM-DD

## Crystallized This Run
| Topic | Conclusion Summary | Confidence | Status |
|-------|-------------------|------------|--------|
| [slug] | [summary] | X% | ai-crystallized (awaiting human review) |

## Deferred (over per-run cap)
[List of eligible converged topics not processed this run, or "None"]

## Note
These are AI-inferred candidates in `Brain/05-Meta/AI Crystallizations/`. The source topics remain `status: converged` and available for the user's manual crystallization via `/manage-thinking-topics crystallize [slug]`.

---

What This Playbook Must Never Do

• ❌ Write to Brain/02-Permanent/, Brain/AI Extracted Notes/, or Brain/05-Meta/Reports/
• ❌ Set status: crystallized on a topic (that marks human, endorsed crystallization)
• ❌ Set provenance to endorsed or originated
• ❌ Write any note in the user's first-person voice
• ❌ Mutate the Thinking Registry or thinking files

If any of these would be required, stop and log - the topic belongs to the human path.

---

Error Handling

---

Completion Checklist

• Registry read - converged topics identified
• Each converged topic deduped against the AI Crystallizations folder
• Batch selected within per-run cap; deferred topics logged
• Each processed topic: conclusion extracted (Final Synthesis preferred)
• Each processed topic: note written to Brain/05-Meta/AI Crystallizations/ with provenance: ai-inferred
• No write touched 02-Permanent/, AI Extracted Notes/, or Reports/
• No topic status changed in the registry
• Index updated
• Session changelog written

---

Relationship to Other Skills

---

Self-Improvement

After completing this skill's primary task, consider tactical improvements:

• Review execution: Were there friction points, unclear steps, or inefficiencies?
• Identify improvements: Could extraction, the note template, or the cap be sharper?
• Scope check: Only execution changes - NEVER relax the segregation boundary (the never-touch-permanent rule is load-bearing, not tactical)
• Apply improvement (if identified):
  • Edit this SKILL.md with the specific improvement
  • Keep changes minimal and focused
• Version control:
  • git add .claude/skills/ai-crystallize/SKILL.md
  • git commit -m "refactor(ai-crystallize): <brief improvement>"