Agent architect

Agent Architect — meta-agent skill

Owner-bounded. This skill drafts artifacts; it does not ship them. Adoption into canonical paths requires an Owner-signed sentinel (ADR-010).

When to load this skill

• The Owner runs /architect "<domain-brief>" (e.g. /architect "RegTech for European banking with PSD2 + AML focus")
• The Owner wants a draft of a new squad without committing to the squad's adoption
• The Owner is exploring whether a domain warrants a dedicated squad vs. extending an existing squad's pitfalls

What the meta-agent emits

A bundle of 5 draft files under .claude/plans/PLAN-NNN/architect/round-1/:

1. team.draft.md — proposed team-personas.md content (≥5 fictional personas with VETO scopes per ADR-009)
2. pitfalls.draft.yaml — proposed pitfalls catalog (≥10 entries)
3. skill-selection.draft.md — list of 3+ proposed skills with descriptions and rationale (one bullet per skill). Each proposed skill's frontmatter sketch includes a paths: activation-glob list BY DEFAULT (PLAN-135 K1 — non-empty glob strings matching the files whose edit should surface the skill, e.g. paths: ["src/payments/**", "**/billing/**"]; the vibe coder never has to know the skill exists). Heavy analytic / audit-class skills additionally declare context: fork so they run isolated instead of polluting the main window. Contract + lint: SPEC/v1/skill-frontmatter.schema.md, lint-skills.py LINT-FM-40/41.
4. personas.draft.md — focused persona discussion (alternative to team.draft.md if the Owner wants prose-first)
5. rationale.md — why this domain warrants a squad, what foundational profile it expects, what it does NOT cover

Plus a sentinel placeholder:

6. approved.md.template — empty sentinel template the Owner copies to approved.md after review (with the actual Approved-By: line filled in)

What the meta-agent must NOT do

• Never write to canonical paths. The hook check_canonical_edit.py blocks Edit/Write/MultiEdit against team.md, frontend-team.md, pitfalls-catalog.yaml, or any skills/**/SKILL.md unless a sibling approved.md sentinel exists with a valid Approved-By: line.
• Never spawn another Architect. The recursion guard in check_agent_spawn.py blocks any spawn whose persona matches Agent Architect if CEO_ARCHITECT_ACTIVE=1 is set in the env.
• Never use real-person names. Personas are fictional composites per ADR-009 §positioning invariants. Real names are a hard reject during bundle validation.
• Never propose a paid tier. ADR-009 §positioning forbids paid tier marketing in squads.
• Never fewer than 3 VETO holders if the domain has ≥3 critical risk axes. Default to 3+ VETOes when in doubt.

Drafting checklist

For each domain brief, the Architect produces drafts that:

• Name 5+ fictional composite personas with backgrounds + mantras
• Identify 3+ VETO holders with explicit block triggers
• Propose 3+ skills with frontmatter (name, description, trigger, paths — activation globs emitted by default per PLAN-135 K1; context: fork on heavy analytic skills)
• Author 10+ pitfalls in YAML format matching domain risks
• Outline 2+ task chains (workflows) for common operations
• Cite the recommended foundational profile (e.g. "core,fintech")
• Document explicitly what the squad does NOT cover
• List references (industry papers, regulator docs, prior incidents)

Output discipline

• All drafts are markdown / YAML, no executable code.
• Drafts are concise: each artifact is ≤ 500 lines.
• Bundle directory is named architect/round-1/ (anticipating potential round-2 revisions).
• rationale.md is the README of the bundle — the Owner reads it first.

Acceptance test (bundle validator)

.claude/scripts/architect-bundle-validate.py <bundle-dir> runs:

1. All 5 draft files present and parseable
2. Bundle dir matches .claude/plans/PLAN-*/architect/round-*/
3. No file in the bundle uses real-person names (heuristic: matches [A-Z][a-z]+ [A-Z][a-z]+ against a deny list of public real names — kept small and Owner-curated)
4. No skill description mentions paid tiers
5. Personas count ≥ 5, pitfalls count ≥ 10, skills count ≥ 3

Exit 0 = pass. Exit 1 = fail (with reasons printed).

Adoption flow (post-draft)

1. Owner reviews the bundle in .claude/plans/PLAN-NNN/architect/round-1/.
2. Owner edits to taste; commits drafts to the PR.
3. Owner copies approved.md.template to approved.md, fills in the Approved-By: line with their handle and the bundle's commit SHA.
4. Owner (or a subsequent agent) migrates the drafts into canonical paths. The sentinel hook ALLOWS the canonical edits because the sentinel is now valid.
5. After migration, the bundle dir is archived; the canonical paths are reviewable in the next squad-validation CI run.

References

• ADR-010 (canonical-edit sentinel)
• ADR-009 (squad bundle contract)
• .claude/scripts/architect-bundle-validate.py
• .claude/hooks/check_canonical_edit.py
• .claude/hooks/check_agent_spawn.py recursion guard