- Add CCGS Skill Testing Framework: self-contained QA layer with 72 skill specs, 49 agent specs, catalog.yaml, quality-rubric.md, templates, README, CLAUDE.md - Add /skill-improve: test-fix-retest loop covering static + category checks - Add 4 missing skills: /art-bible, /asset-spec, /day-one-patch, /security-audit - Add /skill-test category mode (Phase 2D) with quality rubric evaluation - Extend /skill-test audit to cover agent specs alongside skill specs - Update all skill-test and skill-improve path refs to CCGS Skill Testing Framework/ - Remove stale tests/skills/ directory (superseded by CCGS Skill Testing Framework) - Add director gate intensity modes (full/lean/solo) to gate-check and related skills Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
7.5 KiB
Skill Test Spec: /create-architecture
Skill Summary
/create-architecture guides the user through section-by-section authoring of a
technical architecture document. It uses a skeleton-first approach — the file is
created with all required section headers before any content is filled. Each
section is discussed, drafted, and written individually after user approval. If an
architecture document already exists, the skill offers retrofit mode to update
specific sections.
In full review mode, TD-ARCHITECTURE (technical-director) and LP-FEASIBILITY
(lead-programmer) spawn after the complete draft is finished. In lean or solo
mode, both gates are skipped. The skill writes to docs/architecture/architecture.md.
Static Assertions (Structural)
Verified automatically by /skill-test static — no fixture needed.
- Has required frontmatter fields:
name,description,argument-hint,user-invocable,allowed-tools - Has ≥2 phase headings
- Contains verdict keywords: APPROVED, NEEDS REVISION, MAJOR REVISION NEEDED
- Contains "May I write" collaborative protocol language (per-section approval)
- Has a next-step handoff at the end (
/architecture-reviewor/create-control-manifest) - Documents skeleton-first approach
- Documents gate behavior: TD-ARCHITECTURE + LP-FEASIBILITY in full mode; skipped in lean/solo
- Documents retrofit mode for existing architecture documents
Director Gate Checks
In full mode: TD-ARCHITECTURE (technical-director) and LP-FEASIBILITY
(lead-programmer) spawn in parallel after all sections are drafted and before
any final approval write.
In lean mode: both gates are skipped. Output notes:
"TD-ARCHITECTURE skipped — lean mode" and "LP-FEASIBILITY skipped — lean mode".
In solo mode: both gates are skipped with equivalent notes.
Test Cases
Case 1: Happy Path — New architecture doc, skeleton-first, full mode gates approve
Fixture:
- No existing
docs/architecture/architecture.md docs/architecture/contains Accepted ADRs for referenceproduction/session-state/review-mode.txtcontainsfull
Input: /create-architecture
Expected behavior:
- Skill creates skeleton
docs/architecture/architecture.mdwith all required section headers - For each section: drafts content, shows draft, asks "May I write [section]?", writes after approval
- After all sections are drafted: TD-ARCHITECTURE and LP-FEASIBILITY spawn in parallel
- Both gates return APPROVED
- Final "May I confirm architecture is complete?" asked
- Session state updated
Assertions:
- Skeleton file is created with all section headers before any content is written
- "May I write [section]?" asked per section during authoring
- TD-ARCHITECTURE and LP-FEASIBILITY spawn in parallel (not sequentially)
- Both gates complete before the final completion confirmation
- Verdict is APPROVED when both gates return APPROVED
- Next-step handoff to
/architecture-reviewor/create-control-manifestis present
Case 2: Failure Path — TD-ARCHITECTURE returns MAJOR REVISION
Fixture:
- Architecture doc is fully drafted (all sections)
production/session-state/review-mode.txtcontainsfull- TD-ARCHITECTURE gate returns MAJOR REVISION: "[specific structural issue]"
Input: /create-architecture
Expected behavior:
- All sections are drafted and written
- TD-ARCHITECTURE gate runs and returns MAJOR REVISION with specific feedback
- Skill surfaces the feedback to the user
- Architecture is NOT marked as finalized
- User is asked: revise the flagged sections, or accept the document as a draft
Assertions:
- Architecture is NOT marked finalized when TD-ARCHITECTURE returns MAJOR REVISION
- Gate feedback is shown to the user with specific issue descriptions
- User is given the option to revise specific sections
- Skill does NOT auto-finalize despite MAJOR REVISION feedback
Case 3: Lean Mode — Both gates skipped; architecture written with user approval only
Fixture:
- No existing architecture doc
production/session-state/review-mode.txtcontainslean
Input: /create-architecture
Expected behavior:
- Skeleton file is created
- All sections are authored and written per-section with user approval
- After completion: TD-ARCHITECTURE and LP-FEASIBILITY are skipped
- Output notes: "TD-ARCHITECTURE skipped — lean mode" and "LP-FEASIBILITY skipped — lean mode"
- Architecture is considered complete based on user approval alone
Assertions:
- Both gate skip notes appear in output
- Architecture document is written with only user approval in lean mode
- Skill does NOT block completion because gates were skipped
- Next-step handoff is still present
Case 4: Retrofit Mode — Existing architecture doc, user updates a section
Fixture:
docs/architecture/architecture.mdalready exists with all sections populated
Input: /create-architecture
Expected behavior:
- Skill detects existing architecture doc and reads its current content
- Skill offers retrofit mode: "Architecture doc already exists. Which section would you like to update?"
- User selects a section
- Skill authors only that section, asks "May I write [section]?"
- Only the selected section is updated — other sections unchanged
Assertions:
- Skill detects and reads the existing architecture doc before offering retrofit
- User is asked which section to update — not asked to rewrite the whole document
- Only the selected section is updated
- Other sections are not modified during a retrofit session
Case 5: Director Gate — Architecture references a Proposed ADR; flagged as risk
Fixture:
- Architecture doc is being authored
- One section references or depends on an ADR that has
Status: Proposed production/session-state/review-mode.txtcontainsfull
Input: /create-architecture
Expected behavior:
- Skill authors all sections
- During authoring, skill detects a reference to a Proposed ADR
- Skill flags: "Note: [section] references ADR-NNN which is Proposed — this is a risk until the ADR is accepted"
- Risk flag is embedded in the relevant section's content
- TD-ARCHITECTURE and LP-FEASIBILITY still run — they are informed of the Proposed ADR risk
Assertions:
- Proposed ADR reference is detected and flagged during section authoring
- Risk note is embedded in the architecture document section
- TD-ARCHITECTURE and LP-FEASIBILITY still spawn (the risk does not block the gates)
- Risk flag names the specific ADR number and title
Protocol Compliance
- Skeleton file created with all section headers before any content is written
- "May I write [section]?" asked per section during authoring
- TD-ARCHITECTURE and LP-FEASIBILITY spawn in parallel in full mode
- Skipped gates noted by name and mode in lean/solo output
- Proposed ADR references flagged as risks in the document
- Ends with next-step handoff:
/architecture-reviewor/create-control-manifest
Coverage Notes
- The required section list for architecture documents is defined in the skill
body and in the
/architecture-reviewskill — not re-enumerated here. - Engine version stamping in the architecture doc (parallel to ADR stamping) is part of the authoring workflow — tested implicitly via Case 1.
- The retrofit mode for updating multiple sections in one session follows the same per-section approval pattern — not independently tested for multi-section retrofits.