orchestration/gencreator-stack

gencreator-stack Skill — Cross-Brand Stack Catalog + Drift Sentinel

Every repo is an agent surface. Every brand is a constellation of surfaces. This skill is how you keep the map honest.

Activation

Auto-activates when the user's question spans more than one repo or asks about tool assignment across the ecosystem. Pair with starlight-architect (Leadership tier). For single-repo decisions, defer to engineering:architecture (single ADR). For new brand spawning, compose with integration/domain-stack-architecture (vertical-tier scaffolder).

Why this exists

Frank operates 6+ active repos across 4+ brands. Each surface has a coding agent (Claude Code / Codex / Cursor / Gemini CLI), a reasoning surface (Claude Project / ChatGPT Project / Gemini Gem / Custom GPT), a research browser (Arc Space / Chrome+CIC / Comet), and a stack of native MCPs (Vercel / Supabase / Notion / Linear / Figma / GitHub / Gmail / Calendar). Without a manifest, three failure modes compound:

1. Surface drift — a repo's CLAUDE.md claims a tool that's no longer wired. Claude Code in that repo runs with stale context.
2. MCP overlap waste — same connector configured twice in two repos with conflicting auth scopes. Quiet rate-limit failures.
3. Cognitive overhead — Frank has to remember which AI runs which brand. Cognitive load grows non-linearly with brand count.

The skill makes the manifest first-class: one STACK.md per repo, rolled up into one constellation file at substrate level. Drift detection runs weekly so entropy never compounds past 7 days.

Verbs

audit — walk every repo, emit the constellation

python skills/orchestration/gencreator-stack/scripts/stack.py audit

Walks the inventory at skills/orchestration/gencreator-stack/assets/inventory.json, reads each repo's CLAUDE.md + STACK.md (if present), and writes context/stack-constellation.md — the single source of truth for "what runs where". Output is sorted by brand, then by surface tier (substrate → reasoning → coding → research → ops). Repos missing STACK.md show up as gaps with a scaffold hint.

scaffold — drop STACK.md template into a target repo

python skills/orchestration/gencreator-stack/scripts/stack.py scaffold --repo C:\Users\frank\Arcanea

Copies assets/STACK.template.md into the target repo (refuses to overwrite — Frank-ack required). The template is pre-filled with the canonical five-tier matrix and prompts for repo-specific bindings.

assign — bind tools to tiers via guided prompt

python skills/orchestration/gencreator-stack/scripts/stack.py assign --repo C:\Users\frank\FrankX

Reads existing STACK.md, asks Frank (via stdin or AskUserQuestion when running in Cowork) to confirm or change tool bindings for each of the five tiers. Writes back to STACK.md with timestamped change log. No manual editing of the markdown table required.

diff — surface drift between intent and reality

python skills/orchestration/gencreator-stack/scripts/stack.py diff

This is the high-leverage verb. For each repo with a STACK.md, the script compares the manifest against actual filesystem state:

• .mcp.json — does every MCP claimed in STACK.md actually appear here?
• .claude/settings.json and .claude/CLAUDE.md — does the claimed Claude Code config exist?
• .cursor/ or .cursorrules — present iff Cursor is claimed as a coding agent
• .codex/ / .gemini/ / .antigravity/ — present iff those agents are claimed
• Stale references — STACK.md claims vercel-mcp but no Vercel project exists in .vercel/ or no vercel.json at root → suspected drift

Outputs context/stack-drift-<ISO8601>.md with green / yellow / red status per repo. Yellow = manifest says X, reality says X' (close but different). Red = manifest says X, reality has no trace of X.

Tier model — the canonical five

Every STACK.md carries this matrix. Tiers are normative — do not invent a sixth without a board pass.

Tier	Role	Canonical picks (pick one per repo)
Substrate	Persistent context, memory, attestation	Starlight Intelligence System (always)
Reasoning	Long-form thinking, research synthesis, project-scoped chat	Claude Project · ChatGPT Project · Gemini Gem · Perplexity Space · Grok
Coding	Code execution agents (CLI / IDE)	Claude Code · Codex CLI · Gemini CLI · Cursor · Antigravity · GitHub Copilot CLI · Aider
Research browser	Read + agentic web	Arc Space · Chrome + Claude-in-Chrome · Comet · Opera Workspaces
Ops MCPs	Native connectors	Notion · Linear · Vercel · Supabase · GitHub · Gmail · Calendar · Figma · Slack · Drive · Bunq

Rationale for the picks lives in references/tier-rationale.md. Read it before overriding the defaults.

Inventory

The repo list lives in assets/inventory.json. Hand-maintained. Each entry has: brand, path, role (production / development / private / docs / oss), and whether STACK.md is expected. Add new repos here, not inline.

Composes with

• integration/repo-bridge — when a STACK move requires cross-repo coordination
• integration/ecosystem-sync — when STACK changes need to propagate to ecosystem manifest
• intelligence/decision-framework — when assign faces a contested binding
• vision/voice-anti-slop — when generating STACK.md prose (no AI-slop in machine-readable manifests)

When NOT to use this skill

• Single-repo ADRs → engineering:architecture
• Spawning a new domain vertical → integration/domain-stack-architecture
• Touching the substrate's own STACK.md (the universal IS taxonomy) → that's substrate-tier, requires /starlight-board

File outputs

Verb	Writes
audit	context/stack-constellation.md (idempotent — same input, same output)
scaffold	<target-repo>/STACK.md (refuses overwrite)
assign	<target-repo>/STACK.md (updates in place, appends change log)
diff	context/stack-drift-<ISO8601>.md (timestamped, one per run)

Scheduled cadence

Weekly diff runs Sun 08:30 (before /vault-atlas at 08:45). Surfaces drift before Frank starts the week. If drift count > 3 OR any red status, the scheduler escalates a single Memory Bus atom of kind stack-drift-alert so it surfaces on Monday's cockpit.

---

Built on SIP · gencreator-stack skill · v1.0.0