# Brain Synthesize

> Find repeated verified mechanisms across brain owners and automatically promote only high-confidence patterns into global or narrower gotcha, decision, code-style, or recurring-improvement records. Use after eligible structured writes and for explicit/scheduled brain-wide pattern passes. Do not use for generic summarization, single-node facts, cleanup, repair, or unsupported inference.

- Skill: `riggyz/brain-synthesize` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add riggyz/brain-synthesize`
- Raw SKILL.md: https://api.skillmd.com/api/skills/riggyz/brain-synthesize/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: riggyz (https://skillmd.com/u/riggyz)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/riggyz/brain-synthesize

---


# Brain Synthesize

Turn repeated experience into reusable memory without inflating copies or inventing policy.

## Resource Handoff

Invoke `brain` through the host's native skill loader and use the exact base directory it returns for foundation resources. Use the base directory returned for this `brain-synthesize` invocation for its reporter script. Do not list or search parent skill directories, read `SKILL.md` directly, or probe alternate config locations.

## Modes

- **Targeted:** cheap check around newly written structured records; invoked automatically by `brain-remember` and `brain-build`.
- **Full:** explicit or host-scheduled brain-wide prevalence pass.

Run the bundled read-only reporter first. Candidate scores rank review; they never override blockers.

```sh
BRAIN_SYNTHESIZE_DIR="<exact base directory returned by the brain-synthesize skill loader>"
npx tsx "$BRAIN_SYNTHESIZE_DIR/scripts/report-patterns.ts" report \
  --vault-path "<vault-path>" --strict
```

## Evidence Rules

- Count independent occurrences and lineages, not repeated wording.
- A workspace summary and a member project repeating one incident count once.
- Promoted summaries, vendored copies, and derived records are not fresh evidence.
- Legacy prose may produce report-only candidates but cannot auto-promote.
- Contradictions, ambiguous ownership, incomplete provenance, suspicious copies, or a human-authored target collision block automatic writes.
- Choose the narrowest valid scope: project, workspace, wiki, tool, then global.

## Lane Gates

- **Gotcha:** same causal trap, complete provenance, no contradiction, and at least three verified independent occurrences from three independent leaf owners.
- **Decision:** explicit authority covers the target scope. Repeated implementation never creates user intent.
- **Code style:** direct user/org authority or a canonical config explicitly designated for the target scope. Repository-local instructions always override.
- **Recurring improvement:** at least three independent occurrences, common remediation, brain ownership rather than product-tracker ownership, and automatic priority no higher than medium.

## Automatic Apply

When every static and semantic gate passes:

1. Recall bounded source records and any existing target.
2. Confirm mechanism, independence, scope, authority, novelty, and exceptions.
3. Build a plan containing candidate/revision fingerprints and exact source IDs.
4. Revalidate the plan against current vault hashes.
5. Load `brain-remember` and create/update only a synthesis-owned atomic record or stable improvement item.
6. Link the corresponding root/owner router, preserve every source record, and record provenance/fingerprints.
7. Rerun targeted reporting plus structural audit; the result must become promoted/no-op.

Fully automatic means no per-candidate user ceremony after explicit or lifecycle-authorized synthesis invocation. Explicit no-write instructions still win. Never overwrite unrelated human-authored records or delete/merge source evidence.

## Output

Report promoted, already-present, blocked, and skipped candidates concisely. A silent no-op is normal after a targeted check.