claudboard-generate

Generate — Artifact Generation from Analysis

Reads the convention catalog and generates Claude Code artifacts: CLAUDE.md, rules, and full-scope skills.

Model tier: Sonnet. Generation is template-fill on structured catalog input — no cross-document synthesis or pattern discovery. See monorepo-generation spec for rationale.

Invocation

/generate [path]

Path defaults to the current working directory.

---

Phase 1: Load Analysis Input

Recommendation: Run this skill in a fresh Claude Code session. Analysis fills context with discovery data not needed for generation — a clean session gives better output.

1a. Detect mode and locate input

Check for .claudboard/catalog.json first. If present, proceed to 1b-catalog. If absent, check for .claude/reports/claudboard-analysis.md (legacy migration path, 1c).

All modes (single-project, monorepo, workspace) use the same catalog-driven path. Workspace detection is not special-cased: the catalog's mode and umbrella_root fields carry all the information needed. Do NOT branch on mode for generation logic — use umbrella_root (or repo if umbrella_root absent) as the write root for all artifacts.

1b. Load catalog (primary path)

Read .claudboard/catalog.json. Assert schema_version == "1":

if catalog.schema_version != "1":
    error: "Catalog schema mismatch: got <observed>, expected 1.
            Re-run /analyse to produce a fresh catalog, or check
            that claudboard SKILL.md and catalog-schema.json are
            at the same version."

Check generated_at — if older than 24 hours, warn: "Catalog is N days old. Consider re-running /analyse for fresh results."

Derive umbrella_root from catalog (umbrella_root field if present, else repo). All artifact writes use <umbrella_root>/.claude/ and <umbrella_root>/CLAUDE.md as the base.

Derive mode (UX text only), stacks, conventions, patterns, proposed artifacts, and adaptive depth from catalog fields. Display a summary of proposed_artifacts.

1c. Legacy migration (no catalog present)

When .claudboard/catalog.json is absent AND .claude/reports/claudboard-analysis.md is present:

1. Parse the global legacy report and any per-service cloudboard-analysis-<svc>.md files best-effort.
2. Extract proposed_artifacts, conventions, pattern exemplars (best_example file references), and stacks from the reports' sections.
3. Set from_audit: true if per-service reports are present, else false.
4. Write .claudboard/catalog.json with schema_version: "1" and synthesised fields.
5. Log: "Migrated legacy analysis reports to .claudboard/catalog.json"
6. Proceed with catalog-driven generation from 1b.

Legacy report files are NOT moved, deleted, or modified. Subsequent runs see the catalog and skip migration.

If parsing fails (missing "Proposed Artifacts" section or incompatible format):

• Exit with: "Migration failed: could not parse . Run /analyse to produce a fresh .claudboard/catalog.json."

1d. No input found

If neither catalog nor legacy reports exist:

• Single-project / monorepo: "No catalog or legacy reports found. Run /analyse first to produce .claudboard/catalog.json."
• Workspace mode (detected by structure): "No catalog or legacy reports found. Run /analyse from the workspace root to produce .claudboard/catalog.json. If the workspace is not bootstrapped, run /claudboard-workspace-init first."
• Stop here. Do not run analysis inline.

---

Phase 2: Pre-write summary

Show the Proposed Artifacts from catalog.proposed_artifacts. Preserve any dedup decisions encoded in the catalog — don't re-evaluate.

## Artifacts to Generate

Write root: <umbrella_root>/.claude/  (mode: <catalog.mode>)

**Umbrella CLAUDE.md** — [create/update] — [outline]

**Per-service CLAUDE.md (N files):**
- `<service-dir>/CLAUDE.md` — ≤30 lines, service-specific, references umbrella ecosystem.md

**Rules (N files):**
- `<name>.md` (paths: `<catalog.stacks[id].applicable_paths>`) — <depth: catalog.adaptive_depth[stack-id]>

**Skills (M files):**
- `<skill-name>/` — <SKILL.md + references/ + scripts/>
  exemplar: catalog.patterns[<id>].exemplar_path
  [Pattern A dispatcher if per-service variations detected]

[If existing .claude/ found:]
**Already covered (skipping):**
- `<existing artifact>` — no gaps

**NOT generated (umbrella-only rule):**
- No writes to <service>/.claude/skills/, <service>/.claude/rules/, or <service>/.claude/memories/

I'll proceed now — interrupt with Esc to abort or adjust.

---

Phase 3: Generate Artifacts

Write files to <umbrella_root>/.claude/ and <umbrella_root>/CLAUDE.md. Never write to <umbrella>/<service>/.claude/ — per-service .claude/ directories are outside the generation scope under all modes. Never modify existing source files.

Load these references as needed:

• ../claudboard/references/claude-md-template.md → for CLAUDE.md structure
• ../claudboard/references/rule-templates.md → for rule file templates
• ../claudboard/references/skill-generation.md → for full-scope skill generation and Pattern A

3a. Generate or update umbrella CLAUDE.md

Follow ../claudboard/references/claude-md-template.md exactly.

Single-project: Use the standard template. Target 60-120 lines.

Monorepo / Workspace: Use the monorepo variant template from claude-md-template.md. Target 80-150 lines. Include:

• Services table (name, stack, directory, purpose) — use "Umbrella CLAUDE.md Services Section" from claude-md-template.md
• Shared libraries table (name, directory, consumers)
• Per-service build/test commands grouped by service
• Global conventions (branch, commit, CI/CD)
• Critical Rules section covering cross-service concerns
• Note: "Cross-service topology: .claude/memories/ecosystem.md (auto-loaded)"

If CLAUDE.md exists: Read it, add only gaps — new commands, new architecture notes, new rules/skills table entries, new critical rules. Do not rewrite.

If new: Generate from scratch using detected values for every field.

Write to: <umbrella_root>/CLAUDE.md (repo root for monorepo/single-project; workspace root for workspace)

3a-svc. Generate per-service CLAUDE.md (monorepo and workspace only)

For EACH detected service (not library), generate a short per-service CLAUDE.md using the "Per-Service CLAUDE.md Template" from ../claudboard/references/claude-md-template.md.

Rules:

• ≤30 lines total.
• References the umbrella ecosystem.md and rules — never duplicates their content.
• Contains only service-specific facts: entry point, build/test commands, deployment target, known quirks.
• Write to: <umbrella_root>/<service-dir>/CLAUDE.md
• If file exists: read it, add only what's missing. Do not rewrite existing content.

3b. Generate rules

For each rule in catalog.proposed_artifacts where type == "rule", follow ../claudboard/references/rule-templates.md.

Adaptive depth: from catalog.adaptive_depth[<stack-id>] (or proposed_artifact.depth_signal):

• full → 80-120 lines with real code examples
• medium → 50-70 lines
• skeleton → 30-50 lines with TODO markers

paths: scoping: use proposed_artifact.paths (derived from catalog.stacks[id].applicable_paths). In monorepos with multiple services sharing a stack, a SINGLE rule file covers all services in that stack — do NOT generate one rule file per service.

Global rules (no paths:): rules/ci-cd.md, rules/gitops.md, etc. — CI/CD and cross-service conventions.

Write to: <umbrella_root>/.claude/rules/<name>.md. If file exists: add only what's missing; never overwrite.

3c. Generate full-scope skills

For each skill in catalog.proposed_artifacts where type == "skill", follow ../claudboard/references/skill-generation.md.

Umbrella-only rule: All generated skills live under <umbrella_root>/.claude/skills/. NEVER write into <umbrella>/<service>/.claude/skills/ under any mode.

Every generated skill includes:

• SKILL.md (100-250 lines): architecture diagram specific to this project, step-by-step workflow, code examples from catalog.patterns[<id>].exemplar_path
• references/ dir: at minimum one template from detected conventions; annotated example from the exemplar file
• scripts/scaffold.sh if the skill creates 3+ boilerplate files with predictable naming

Adaptive depth: from catalog.adaptive_depth or proposed_artifact.depth_signal. Full for clean codebases, skeleton+ask for inconsistent patterns.

Monorepo / Workspace — Pattern A dispatcher skills:

When the catalog's proposed_artifacts contains a skill where services/repos in the same stack have meaningfully different procedural content, generate a Pattern A dispatcher skill:

<umbrella_root>/.claude/skills/<concern>/
├── SKILL.md              # dispatcher — enumerates exact valid service names
└── references/
    ├── <service-a>.md    # per-service content
    ├── <service-b>.md    # per-service content
    └── ...

Follow the Pattern A contract from ../claudboard/references/skill-generation.md → "Pattern A". SKILL.md must enumerate the exact set of valid service names (closed-set lookup). Reference filenames must exactly match those names.

Scope per-stack skills to catalog.stacks[id].applicable_paths; reference only that stack's conventions and exemplars.

Write to: <umbrella_root>/.claude/skills/<name>/. If skill exists: add only missing components.

3d. Completion report

## Generation Complete

Generated:
- CLAUDE.md [created/updated — N lines]
- Rules: [list with adaptive depth used]
- Skills: [list, each with: SKILL.md ✓, references/ ✓/✗, scripts/ ✓/✗]

Next steps:
- Review generated artifacts and adjust to team preferences
- Run `/refresh` as the project evolves to keep artifacts current
- Flesh out any TODO sections in skeleton rules or skills
- Try the suggested validation tasks below to verify artifacts work

Conditional next step (include only when analysis report indicates Jira+ADO MCPs are configured):

Ready to generate a tailored feature-workflow skill? Run /claudboard-workflow to create a full ticket→branch→BDD→plan→implement→PR workflow for this project.

3e. Validation suggestions

After presenting the completion report, propose 3-5 representative tasks the user can try in a fresh Claude Code session to verify the generated artifacts actually help. Select tasks based on what was detected:

Task selection (pick 3-5 from this priority list based on what was generated):

1. If skills were generated: "Create a new [entity/component/endpoint] using the generated [skill-name] skill" — tests whether the skill workflow produces correct code
2. If rules cover conventions: "Modify [file from god-class list or hotspot] — does Claude follow the [DI/logging/error-handling] conventions?" — tests rule loading
3. If CLAUDE.md has build commands: "Build and test the project" — tests whether the commands table is accurate
4. If infra rules generated: "Add a new [Helm value/pipeline step/Dockerfile]" — tests infrastructure context
5. If testing rules generated: "Write tests for [existing untested file]" — tests test convention rules

Output format:

## Suggested Validation Tasks

Try these in a fresh Claude Code session to verify the artifacts work:

1. [Task description] — validates: [artifact name]
2. [Task description] — validates: [artifact name]
3. [Task description] — validates: [artifact name]

If Claude stumbles on any of these, run `/refresh` to update the artifacts.

---

Error Handling

Condition	Behavior
Target path doesn't exist	Report error with path and stop
No catalog or legacy reports found	Tell user: "No catalog or legacy reports found. Run /analyse first to produce .claudboard/catalog.json." Stop.
Catalog schema mismatch	Exit with error naming observed and expected version; instruct user to re-run /analyse
Catalog older than 24 hours	Warn but proceed if user confirms
Legacy migration parsing fails	Exit with error naming offending file; instruct user to run /analyse
Catalog contains "ASK USER" markers	Pause and ask user before generating those artifacts

Constraints

• Never modify source code — only write to <umbrella_root>/.claude/, <umbrella_root>/CLAUDE.md, and <umbrella_root>/<service>/CLAUDE.md.
• No per-service .claude/ writes — NEVER write to <service>/.claude/skills/, <service>/.claude/rules/, or <service>/.claude/memories/ under any mode. Per-service content reaches the umbrella via rules with paths: globs, ecosystem.md, or a Pattern A dispatcher skill.
• Merge, don't replace — when .claude/ artifacts exist, append or update sections, never delete existing content.
• Secrets: Never include secret values in generated artifacts.
• Per-service audit reports (.claudboard/audits/*.md) are NOT consumed by /generate even when present — only the catalog is read.
• mode field is UX-only — never branch on mode for generation logic; use umbrella_root for write paths.

Reference Files

File	When to load
../claudboard/references/catalog-schema.json	Phase 1b — schema validation on catalog read
../claudboard/references/catalog-format.md	Phase 1b-1c — catalog field semantics and migration rules
../claudboard/references/claude-md-template.md	Phase 3a — CLAUDE.md generation
../claudboard/references/rule-templates.md	Phase 3b — rule file generation
../claudboard/references/skill-generation.md	Phase 3c — full-scope skill generation