AI-Layer Starter Kit
AI-Layer Starter Kit
Version: v1.0 (2026-05-26) · Source: WO-0010 / RD-0033 Owners: Pablo (Production Line Architect), Marcel (Methodology), Elena (Enterprise Architect) Thèse : « la valeur est le harnais, pas le modèle » — foundry, pas hébergeur
Standard harness embedded in every digital talent shipped by the factory. Closes 5 factory gaps identified from the RD-0007→RD-0016 video intelligence batch (cf. RD-0033 R&D evaluation, score 4.2/5).
What's inside
| # | Component | Purpose | Closes Gap |
|---|---|---|---|
| 1 | CLAUDE-root.md + CLAUDE-domain-template.md |
Layered CLAUDE.md (lean root + one per role/domain) | #1 — per-role CLAUDE.md |
| 2 | hooks/stop-self-improve.* |
Stop-hook auto-improving — captures lessons at end of session | #2 — auto-improving stop-hook |
| 3 | skills/SCOPING.md + example |
Convention: skills scoped via paths: frontmatter |
#3 — globally loaded skills |
| 4 | agents/explorer.md |
Read-only subagent for safe code exploration | #4 — formalized explorer |
| 5 | mcp/codebase-search/ |
MCP server with where_is / find_references / outline |
#5 — AST search (fusion with RD-0031) |
How a delivered talent uses it
The kit is copied (not symlinked) into every freshly generated talent during assembly Phase D. The talent then ships standalone — the client runs it autonomously, no factory dependency (foundry model, cf. memory delivery-model-foundry-not-hosted).
See USAGE.md for the full integration flow.
Anti-doublon
- Composant 5 (codebase-search MCP) fusionne avec RD-0031 (distribution toolkit-catalog). Pablo synchronise les deux ordres avant le build du composant 5.
- Ne pas re-scoper RD-0021 (Claude Managed Agents) — adjacent mais distinct.
- Understand-Anything (RD-0007) = source optionnelle du
knowledge-graph.json, jamais une dépendance dure du kit.
Anti-derive (règle simplicité)
V1 = 5 composants strictement. Si un 6ᵉ émerge, il devient v2 — pas une dérive de v1. Cf. WO-0010 « Out of scope ».