Scaffold

Scaffold

Design a workflow system, then build it from primitives.

Core Principle

The workflow is the product. CLAUDE.md, rules, hooks, skills, agents, and tools are primitives — building blocks that compose into a coherent workflow system. This plugin doesn't generate config files in isolation. It designs a workflow first, then determines which primitives implement each part.

Workflow (the goal)
├── Agents (who does the work)
│   └── Skills (what they know how to do)
│       └── Tools (what they use)
├── Hooks (automatic triggers)
│   └── Rules (constraints they enforce)
└── CLAUDE.md (shared context for everything)

Pipeline

Phase 1: Scan the System

Explore the target repo to understand the existing system:

• Languages, frameworks, package manager (manifest files)
• Build, test, lint commands (Makefile, scripts, CI config)
• Project structure (monorepo vs single, workspace layout)
• Existing Claude Code config (.claude/, CLAUDE.md, settings)
• Existing AI tool configs (AGENTS.md, .cursor/rules, etc.)
• Formatter and linter configuration
• Git conventions (commit style, branch naming from history)
• Documentation, architecture docs
• CONTEXT.md and docs/adr/ if present
• Environment setup (services, env vars, dependencies)
• CI/CD pipeline structure
• Deployment targets and process

Synthesize findings into a system map — not raw output, but a concise model of how the repo's development system works today.

Phase 2: Grill — Design the Workflow

Run grill-with-docs to design the target workflow system. The grill isn't about config preferences — it's about understanding the workflow the user needs.

Decision Persistence

Write each resolved decision to .claude/scaffold-decisions.md immediately. Do not rely on conversation context alone — compaction will degrade it.

Format:

## Decisions

### D1: [Decision Title]
- **Resolved**: [what was decided]
- **Primitives**: [which primitives this maps to, e.g., Hook + Rule + CLAUDE.md]
- **Rationale**: [why, in one line]

### D2: [Decision Title]
...

Create the file lazily on first resolved decision. Generators read this file as primary input — conversation context is supplemental.

Grill Rules

1. One question at a time. Never batch.
2. Recommend an answer. Based on scan findings.
3. Explore before asking. If code answers it, don't ask.
4. Challenge the glossary. Check terms against CONTEXT.md.
5. Sharpen fuzzy language. Propose precise terms.
6. Cross-reference with code. Verify claims against the codebase.
7. Update CONTEXT.md inline. As terms resolve, write immediately.
8. Persist decisions inline. Write to .claude/scaffold-decisions.md as each decision resolves.
9. ADRs sparingly. Only when ALL THREE: hard to reverse + surprising + real trade-off.

Grill Topics (workflow-oriented)

1. System purpose — What does this repo do? Who uses it? What's the development lifecycle?
2. Current workflow — How does work flow from idea to production today? Where are the bottlenecks?
3. Desired workflow — What should the ideal workflow look like? What's missing?
4. Actors — Who (or what agent) performs each step? Where does Claude fit?
5. Automation boundaries — What should be automatic (hooks) vs on-demand (skills) vs orchestrated (workflows)?
6. Quality gates — What checks must pass? When? Who enforces them?
7. Domain knowledge — What does Claude need to know to participate effectively? What does it get wrong?
8. Failure modes — What goes wrong? How should errors be handled? Escalation paths?

Decision Tracking

Each resolved decision maps to one or more primitives:

Decision: "All PRs must pass type checking before review"
→ Hook: pre-commit type check
→ Rule: type-safety.md with path scope
→ CLAUDE.md: mention typecheck command

Decision: "Deploy process is: build → test → stage → smoke test → promote"
→ Workflow: deploy.js with pipeline() stages
→ Skill: /deploy for manual trigger
→ Agent: deploy-verifier for smoke tests

Phase 3: Compose the System

Existing File Protection

Before invoking each generator, check for existing files it would create. For each existing file:

1. Read the existing file — understand what's already there
2. Diff, don't overwrite — propose additions/removals with one-line reasons
3. Ask before replacing — "Apply these changes?" with options: Apply all / Let me pick / Skip
4. Respect --append — pass --append to generate-claudemd if CLAUDE.md already exists

Generators that find existing .claude/rules/, .claude/skills/, or .claude/agents/ must complement, not duplicate. Scan existing artifacts first.

Primitive Mapping

Map resolved decisions to primitives. This is systems thinking — each primitive has a role:

Primitive	Role in System	When to Use
CLAUDE.md	Shared context — the system's "memory"	Knowledge every primitive needs
Rules	Constraints — the system's "laws"	Standards that must be enforced
Skills	Capabilities — the system's "procedures"	On-demand multi-step workflows
Hooks	Reflexes — the system's "autonomic responses"	Automatic reactions to events
Agents	Specialists — the system's "team members"	Focused roles with specific expertise
Tools	Instruments — the system's "equipment"	External integrations and utilities
Workflows	Orchestration — the system's "nervous system"	Multi-step automated pipelines

Generate in dependency order:

1. CLAUDE.md        (shared context — everything references this)
2. Rules            (constraints — agents and hooks enforce these)
3. Skills           (capabilities — agents use these)
4. Hooks            (reflexes — automatic enforcement of rules)
5. Agents           (specialists — use skills, follow rules)
6. Tools            (instruments — agents and skills use these)
7. Workflows        (orchestration — compose agents, skills, tools into pipelines)

Invoke each generator: /generate-claudemd, /generate-rules, etc.

Skip generators when the grill determined they're not needed.

Phase 4: Verify the System

After generation:

1. Reconcile CLAUDE.md — Re-read CLAUDE.md and update its Workflow section to match actual generated artifacts. CLAUDE.md was generated first (step 1) but summarizes things defined in steps 2-7. Ensure the workflow description, agent names, and skill references match what was actually produced.
2. Consistency check — Read .claude/scaffold-decisions.md and verify each decision maps to at least one generated primitive. Check that hooks reference rules that exist, workflows invoke agents that are defined, and agents reference skills that exist.
3. Completeness check — Does every grill decision in the decisions file map to at least one primitive? List any gaps.
4. Coherence check — Walk through the workflow end-to-end using the generated primitives. Do they compose into the workflow the user described?
5. Security check — No hardcoded secrets, no sensitive data in generated files.
6. Present the system — Show the full primitive tree with one-line descriptions. Explain how the workflow flows through the primitives.

Invocation

/scaffold — Full pipeline: scan → grill → compose → verify.

$ARGUMENTS:

• --scan-only — Phase 1 only, report system map
• --quick-grill — Ask 3-5 high-signal questions instead of full interrogation (fast, less accurate). Warns that output will be generic and may not reflect actual workflow.
• --generators=claudemd,rules,hooks — Run only specified generators
• --dry-run — Show the workflow design without writing files
• --workflow=deploy — Focus scaffold on a specific workflow

Context Management

Long-running skill. Protect context:

• Subagents for deep file reads (>100 lines)
• Synthesize, don't dump raw output
• Track decisions as structured records, not prose
• Each generator is a separate skill invocation — keeps context focused