openakari: Skill for Claude Code

/synthesize

You are synthesizing accumulated findings to surface patterns, contradictions, and insights that individual log entries or analyses miss on their own. The argument specifies the scope: a project name, a time range, a topic, or specific file paths.

Pre-flight audit

Before writing synthesis output, run the synthesis pre-flight audit (docs/sops/synthesis-preflight-audit.md). Enumerate upstream sources, flag provisional data, and spot-check key numerical claims that will be cited. This prevents the most common synthesis failure: propagating contaminated or stale numbers from prior sessions.

Gather material

Based on the scope argument:

• If a project name: read the project README (especially Log and Open questions), any files in the project directory, and related decision records.
• If a time range: scan logs across all active projects for entries in that range.
• If a topic: grep across projects, decisions, and docs for relevant material.
• If file paths: read those files directly.

Also check decisions/ for relevant recorded choices and docs/ for framework documents.

Analyze across CI layers

For the gathered material, identify:

1. Cross-layer causal chains — Findings that connect across CI layers. (e.g., "The evaluation gap [L4] exists because the interface [L3] can't present 3D interactively to LLMs, which limits what the model [L1] can judge.")
2. Convergent signals — Multiple independent findings pointing to the same conclusion. What do they converge on?
3. Contradictions — Findings that conflict with each other. Which is better grounded? What would resolve the disagreement?
4. Gaps — What questions remain unasked? What CI layers are underrepresented in the findings? What experiments would fill the gaps?
5. Gravity candidates — Recurring patterns that should be formalized. What manual work could become automated tooling? What tooling could become model capability?

Output format

## Synthesis: <scope>

### Material reviewed
<bulleted list of files/entries consulted>

### Cross-layer chains
<numbered findings, each tracing a connection across 2+ CI layers>

### Convergent signals
<what multiple findings agree on — with specific references>

### Contradictions
<conflicting findings and what would resolve them>

### Gaps
<what's missing — specific questions or unexamined CI layers>

### Gravity candidates
<patterns that should move downward — from manual to tool to model>

### Implications
<1-3 concrete recommendations for what to do next, referencing specific projects or actions>