Validate

Validator

Run MCS quality checks on any product in workspace/ and return actionable, scored reports.

When to use: After building or modifying a product, before publishing, or anytime you want a quality snapshot.

When NOT to use: On products outside workspace/. Do not use to validate the Engine itself.

Activation Protocol

0. Shared preamble: Load references/quality/activation-preamble.md — context assembly, persona adaptation, deterministic routing rules.
1. Detect product type: read .meta.yaml → product.type and state.phase
   • Missing → infer from file structure (SKILL.md, AGENT.md, SQUAD.md, hooks.json, etc.)
   • Cannot determine → ask: "What product type is this?" 1b. Mode selection (Express vs Guided). Read creator.yaml → preferences.workflow_style. Resolve the flow mode:
   • --express flag OR workflow_style == "autonomous" → Express mode. Skip the coaching explanations after each stage, suppress the remediation menu, and deliver a single verdict block at the end (pass/warn/fail + fix instructions in a compact list). Persona tone still holds; only the conversational scaffolding is trimmed.
   • workflow_style == "guided" or missing → Guided mode (default). Walk each stage with the full coaching voice and propose remediation interactively after failing stages.
2. Maintain creator persona: Read creator.yaml → adapt to profile.type and technical_level
3. Load voice identity: Load references/quality/engine-voice-core.md. Load the full references/quality/engine-voice.md only for peak moments (first-pass milestone celebration, confronting failure verdict) — see UX Stack below. 3b. Exemplar load: Load references/quality/exemplar-outputs.md sections E6 and E7 only — the validation pass and failure exemplars. Your verdict MUST carry the same visual structure: Frame for pass (with tier badge), rail format for failure (with numbered fixes + estimated score after). Adapt to creator context — never copy verbatim.
4. Load DNA requirements: product-dna/{type}.yaml 4b. Load architectural DNA: Read structural-dna.md. The 10 architectural principles and the Tier 1 DNA patterns (D1-D4, D13, D14) are the canonical audit baseline — Stages 3 and 5 grep the product against them, and any violation surfaces as coaching.
5. Load product spec: references/product-specs/{type}-spec.md 5b. Load entity ontology (squad/system/agent/workflow/minds): If type ∈ {squad, system, agent, minds, workflow}, read references/entity-ontology.md. This substrate drives semantic validation:
   • §HERITAGE: verify the product inherits correct DNA from its lineage (squad must pass all agent DNA + D9/D10/D12/D18)
   • §COMPOSITION: verify only allowed compositions (squad→agents+minds+skills+workflows; system→everything; workflow→skills only)
   • §AGENT_ROLES: if .meta.yaml has agent_role, verify tool boundaries match the role
   • §SQUAD_ANATOMY: verify all 8 mandatory components exist and have content
   • §WORKFLOW_VS_SQUAD: verify workflows don't contain agents and squads don't use fixed-sequence-only routing
   • §HEURISTICS: surface coaching if product shows signs of wrong type (skill >800 lines → suggest agent)
   • For type=system ONLY: §SYSTEM_ENGINES — verify declared gears have concrete implementations (not just prose), verify counterpart couplings are declared, verify critical chain (E4→E5→E6→E7) is complete if perception gear is active
   • §INTELLIGENCE_PIPELINE — verify baseline delta (is this better than Claude vanilla?), verify substance (does this carry domain intelligence or is it just formatted instructions?)
6. Load config: config.yaml → scoring weights, thresholds, placeholder patterns
7. Load gates: quality-gates.yaml → state transition rules 7b. Load proactives: Load references/engine-proactive.md — wire #1 (pipeline guidance: after validate passes, guide to /test then /package), #19 (error recovery: on validation failure, propose specific fixes), #20 (test mandate: if MCS-2+ and not tested, block /package suggestion).
8. CLI contract: Load references/cli-contract.md for Stage 6 (CLI Preflight). Severity map:
   • Warning: validate --json — CLI validation is advisory during /validate (blocking only during /publish)
   • Warning: doctor --json — health check is advisory, score < 8.0 triggers suggestion
   • Stage 6 detail protocol: references/validation-stages/stage-6-cli-preflight.md

Commands

/validate                    → Auto-detect, run MCS-1
/validate --level=2          → MCS-2 (includes MCS-1)
/validate --level=3          → MCS-3 (includes MCS-1+2, PRO only)
/validate --fix              → MCS-1 + auto-remediate fixable issues
/validate --report           → MCS-1 + output detailed report file
/validate --batch            → Validate ALL products sequentially

Core Instructions

STAGE EXECUTION

Execute stages in order. Blocking stages stop on failure. Non-blocking stages report but continue.

Load detailed stage protocols from references/ on demand. Each stage is a separate file — load only the stage(s) you need for the current --level.

Stage 0 runs first when .meta.yaml contains an intent_declaration block. It is advisory — surfaces coherence drift as coaching, never blocks. Skips silently for legacy products that lack the declaration, with a one-line advisory note.

Stage 0 failure behavior: If Stage 0 checks detect errors (enum_membership, type_legality, codex_consistency), they are reported as coaching items in the verdict — they appear in the output but do NOT affect the overall score or the pass/fail verdict. Stage 0 results are written to .meta.yaml → stage_0_results for downstream consumers but never prevent progression to Stage 1.

Stage 7 sub-stages (7b cognitive fidelity, 7c baseline delta, 7d composition check) are inside the stage 7 file — they only fire under their prerequisites and never need a separate load.

Index: ${CLAUDE_SKILL_DIR}/references/validation-stages/_index.md lists all stages with file paths and routing rules.

Stage routing by level:

• --level=1 (MCS-1): Stages 0, 1, 2, 3, 6, 9
• --level=2 (MCS-2): Stages 0, 1, 2, 3, 4, 5 (squad only), 6, 7, 7b-7d, 8, 9
• --level=3 (MCS-3): Stages 0, 1, 2, 3, 4, 5, 6, 7, 7b-7d, 8, 9

Squad-specific validation (applies at MCS-2+): When product.type == "squad", the following additional checks run regardless of edition:

• Stage 5 DNA Tier 3 patterns D9, D10, D18 are MANDATORY at MCS-2 for squads (not PRO-gated). These are the coordination DNA — without them, a squad is just agents in a folder.
  • D9 (Orchestrate Don't Execute): Verify orchestrator agent's allowed-tools includes Agent but excludes Write/Edit/NotebookEdit. Orchestrator routes, never executes.
  • D10 (Handoff Specification): Verify config/handoff-protocol.md exists and contains a handoff envelope format (structured fields: from, to, task, state, constraints).
  • D18 (Subagent Isolation): Verify each agent in agents/*.md is a distinct file with its own identity. No agent definition is inlined in SQUAD.md.
• Per-agent validation: Iterate every agents/*.md file and verify individually:
  • D1: Has activation protocol or identity section
  • D2: Has anti-patterns section with ≥3 items
  • D4: Has quality gate with ≥2 verifiable criteria
  • D14: Has when-not-to-use or degradation section
• Routing completeness: Verify config/routing-table.md covers all agents listed in SQUAD.md roster. Every agent in roster must appear in at least one routing rule.
• Task registry coherence: If tasks/task-registry.yaml exists, verify each task references an agent that exists in agents/.
• Chain registry coherence: If chains/chain-registry.yaml exists, verify each chain references only agents that exist in agents/.
• Heritage coherence (entity-ontology.md §HERITAGE): Verify squad inherits DNA from agent lineage — each specialist agent in agents/*.md must independently pass agent-level DNA (D1, D2, D4, D6, D8, D11, D14, D15). The family-skill inheritance chain enforced at validation time.
• Composition coherence (entity-ontology.md §COMPOSITION): Verify squad only composes allowed entities: agents(2+), minds(0+), skills(0+), workflows(0+). If squad references hooks or claude-md fragments, flag coaching: "Squads don't contain hooks or claude-md directly — that's system-level. Consider promoting to system."
• Role coherence (entity-ontology.md §AGENT_ROLES): If tasks/task-registry.yaml exists, verify each task's assigned_agent has a ROLE consistent with the task type:
  • Write/create tasks → assigned to EXECUTOR or TRANSFORMER agents
  • Analyze/report tasks → assigned to SPECIALIST or VALIDATOR agents
  • Route/coordinate tasks → assigned to ORCHESTRATOR or ROUTER agents
  • Advise/reason tasks → assigned to ADVISOR agents Mismatch is coaching (not blocking): "Task '{task_id}' is a write task assigned to {agent} which has SPECIALIST role (read-only). Consider reassigning."
• Anatomy completeness (entity-ontology.md §SQUAD_ANATOMY): Verify all 8 squad anatomy components have content beyond WHY comments:
  1. agents/ has ≥2 files with substantive content
  2. config/routing-table.md has routing rules (not just placeholder)
  3. config/handoff-protocol.md has envelope format defined
  4. workflows/ has ≥1 workflow with steps
  5. skills/ directory exists (advisory — some squads don't need shared skills)
  6. SQUAD.md has quality/checklist section with ≥2 items
  7. kernel/ has output format standards
  8. SQUAD.md has escalation section with confidence thresholds Missing anatomy = coaching warning with specific fix instruction.
• Type fitness heuristic (entity-ontology.md §HEURISTICS): Apply promotion/demotion checks:
  • Squad with only 1 specialist agent → "This squad has only 1 specialist. Consider demoting to agent."
  • Agent primary file >800 lines → "This agent has {N} lines. Consider splitting into a squad."
  • Workflow with LLM-judgment routing → "This workflow has contextual routing. Consider promoting to squad."
• Agent role-tool coherence (entity-ontology.md §AGENT_ROLES): For each agent, if .meta.yaml → agent_role is populated:
  • EXECUTOR: must NOT deny Write/Edit/Bash
  • SPECIALIST/VALIDATOR: should deny Write/Edit/Bash (read-only)
  • ORCHESTRATOR: must have Agent tool, must deny Write/Edit/NotebookEdit
  • ADVISOR: must deny Write/Edit/Bash/NotebookEdit Mismatch → coaching: "Agent role is {role} but tool boundary doesn't match. Expected: {expected}."

Stage 9 (Voice Coherence) is the last stage at every level. Advisory — never blocks publish. Stage 9 audits the product against the myClaude voice contract (P10 Touch Integrity anchor).

UX STACK (load before rendering output)

1. references/ux-experience-system.md §1 Context Assembly + §2.3 Moment Awareness (pass vs fail) + §3.3 Score Trajectory
2. references/ux-vocabulary.md — translate tiers and terms
3. references/quality/engine-voice.md — Brand DNA + sfumato constraints

Cognitive rendering: /validate output adapts to creator journey. First validation pass = milestone moment (§4.1). Nth pass at 100% = skip celebration, surface next challenge ("Quality mastered. Distribution next?"). Failure = direct, prioritized, no sugar-coating. Score trajectory = show only when 2+ data points tell a meaningful story. Expert creators get technical details. Beginners get human explanation of what the score means.

SCORING & OUTPUT

Vocabulary rule: All creator-facing output uses ux-vocabulary.md tier names (Verified / Premium / Elite) instead of MCS-1/2/3. Stage names (Stage 0-9) are internal — creator sees "structure check", "content depth check", "expertise check", etc. DNA pattern IDs (D1-D20) never appear in creator output — describe the issue in plain language. Exception: developer/hybrid creators with technical_level=expert MAY see MCS-N labels with the vocabulary translation in parentheses.

Read ${CLAUDE_SKILL_DIR}/references/validation-scoring.md for:

• Hard veto rules (pre-scoring)
• Scoring formula: OVERALL = (DNA×0.50) + (STRUCTURAL×0.30) + (INTEGRITY×0.20)
• Verdict logic (READY / NEEDS WORK / NOT READY)
• Output format template
• Token budget report
• Guided iteration (draft fixes for failures)
• Auto-fix rules (--fix flag)
• Batch validation (--batch)
• Publish pre-flight
• Persona-aware rendering
• State update templates (.meta.yaml + STATE.yaml)

Quality Gate

The Validator skill itself passes if:

• It correctly identifies product type from .meta.yaml or file structure
• It runs all checks appropriate for the requested MCS level
• Every failed check includes a specific, actionable fix instruction
• Score calculation matches formula: passed / total × 100
• .meta.yaml is updated after every validation run
• --fix never modifies content that the creator wrote

Anti-Patterns

1. Validating without reading product-dna — Always load type-specific DNA before checking patterns
2. Skipping blocking stages — Stage 1-2 must pass before Stage 3+
3. Generic fix instructions — Every failure needs a specific, actionable fix
4. Modifying creator content in --fix — Only structural/formatting fixes
5. Running Stage 8 for MCS-1 — Value intelligence adds noise at MCS-1, skip it
6. Blocking cascade hiding downstream issues — When Stage 1-2 fails, downstream stages don't run. But the creator loses visibility into DNA/content issues that exist independently. After blocking failure, emit an advisory note: "Structural issues block full validation. After fixing these, re-run to see content-level results." Never silently hide what wasn't checked.

Compact Instructions

When context is compressed, preserve:

• Product slug, type, and current MCS level being validated
• Stage progress (which stages passed/failed)
• Overall score and verdict (READY/NEEDS WORK/NOT READY)
• Top 3 failures with fix instructions
• Value intelligence score if Stage 8 ran
• Whether --fix or --batch mode is active
• UX rule: Celebrate work not person. Show score trajectory if 2+ data points. Use ux-vocabulary.md tiers (Verified/Premium/Elite not MCS-1/2/3) in user-facing output.