Create

Scaffolder

Generate complete, MCS-1-valid project structure for any product type with guidance comments baked in.

When to use: Starting a new product — any of the 13 types.

When NOT to use: When the product already exists in workspace/ and you want to add content (use /fill). Do not use to overwrite an existing scaffold unless --force is specified.

---

Activation Protocol

1. Shared preamble: Load references/quality/activation-preamble.md — context assembly, persona adaptation, deterministic routing rules.
2. Read creator.yaml — load defaults. Check schema_version:

• **` aggregation (§COMPOSITION_PRINCIPLES)

1. Load ${CLAUDE_SKILL_DIR}/references/discovery-questions.md → category questions.
2. Load references/product-specs/{category}-spec.md + templates/{category}/.
3. Load workspace/domain-map.md if exists → prefill scaffold sections.

10b. Link scout report if used — scout_source is already recorded in intent_declaration.scout_source by Section 0 Step 11. No additional action needed here; the canonical home for scout_source is intent_declaration, not a parallel field. 10c2. CLI contract: Load references/cli-contract.md for unified error handling. This skill has minimal CLI surface — one command at Step 1.5. Severity map:

• Silent-skip: search --category {type} --sort downloads --limit 3 — marketplace scan for competitive context. Skip without warning on failure. First-product guard: skip entirely if is_first_product (already documented at Step 1.5)
• All queries: append --json 2>/dev/null, 15s timeout

10c. Portfolio intelligence (back-reference from /validate): Read STATE.yaml → workspace.products[] AND STATE.yaml → mcs_results. Group by domain. If domain has existing products:

• Show: "Your {domain} portfolio: {slugs}. This will be product #{N}."
• If N >= bundle threshold (config.yaml → intelligence.portfolio.bundle_suggestion_threshold): suggest bundling.
• If complementary type opportunity exists: surface composition suggestion.
• MCS target calibration: If domain has 0 products → default MCS-1 (ship fast, test market). If domain has 2+ validated products → default MCS-2 (quality matters, audience exists).
• Record intelligence.domain and intelligence.market_position in .meta.yaml.

1. Generate scaffold in workspace/{product-slug}/ with DNA patterns + WHY comments. The scaffold's type + structure come from the matched_cell.canonical_form decided at Section 0 Step 10.

Ontology-aware scaffold verification (requires entity-ontology.md loaded at step 7b): For type=squad: verify scaffold contains all 8 anatomy components from entity-ontology.md §SQUAD_ANATOMY:

1. Agent roster (agents/) — each specialist as separate .md file with role-appropriate frontmatter
2. Routing table (config/routing-table.md) — IF intent → agent mapping with judgment points
3. Handoff protocols (config/handoff-protocol.md) — structured envelope format (from, to, task, state, constraints)
4. Workflows (workflows/) — multi-agent sequences with gates and loops
5. Skills-as-instruments (skills/) — reusable fragments agents invoke
6. Checklists — deterministic verification section in SQUAD.md
7. Templates — output format standards in kernel/
8. Escalation/quality gates — confidence thresholds and loop limits in SQUAD.md

For type=agent or type=minds: apply §AGENTROLES — the role selected during discovery determines frontmatter tool boundaries (allowed-tools/denied-tools) and handoff format expectations. Generate role-appropriate frontmatter. For type=workflow: verify scaffold distinguishes from squad per §WORKFLOWVSSQUAD — workflows use skills (not agents), fixed sequence (not LLM routing), deterministic gates (not judgment). If scaffold looks like a squad, surface heuristic coaching. For type=system: apply §HERITAGE — system inherits all squad DNA + adds CLAUDE.md fragment, hooks, output-style, STATE.yaml. Scaffold must include provisions for all composed types per §COMPOSITION. For type=system: apply §SYSTEMENGINES — scaffold a gears_declaration section in the system's primary file. Ask the creator: "Which of these organism capabilities do you want active?" Present the 13 gears grouped by function:

• Boundary: E0 (clausura) — always required
• Metabolism: E1 (token budget), E2 (constitution re-injection)
• Memory: E3 (triple memory), E10 (persistent state)
• Perception-Reflex: E4 (hooks perception), E5 (cognitive reflex)
• Judgment-Coordination: E6 (isolated judgment), E7 (coordination network)
• Identity: E8 (external voice)
• Integrity: E9 (recursive validation), E11 (longitudinal self-observation — advanced, MCS-3)
• Lifecycle: E12 (reversible install/uninstall)

For each active gear, the scaffold includes a section with WHY comment and its counterpart declaration. 11b. Locale-adaptive clause substitution (W3.6 — MANDATORY for 6 certified types): For each template file written (SKILL.md, AGENT.md, SQUAD.md, CLAUDE.md for system, AGENT.md for minds, OUTPUT-STYLE.md), perform placeholder substitution for {{LOCALE_ADAPTIVE_CLAUSE}}: a. Read the canonical clause block from references/locale-adaptive-clause.md §2 (the fenced markdown block between >> and >>, inclusive). b. Read the localized header from config.yaml → routing.common.locale_adaptive_clause.localized_header_catalog.{creator.language}. If the creator's language is not in the catalog, use the fallback entry and emit an advisory note. c. Build the substitution block: \n\n{canonical_clause}. d. Replace the literal string {{LOCALE_ADAPTIVE_CLAUSE}} in each written template file with the substitution block. Use a literal-string replace, not a regex — the placeholder is unique and must not be interpreted. e. Verify post-substitution: the forged file must contain both marker strings (>> and >>). If either marker is missing after substitution, the forge fails + rollback. e2. Duplication guard: Before substitution in step (d), check if the target file ALREADY contains the marker /dev/null. Show top 3 with downloads + "What will yours do differently?" Coaching only — never blocking. First-product guard: If is_first_product, SKIP this step — a first-timer seeing competitor data before articulating their own idea causes second-guessing. Show marketplace context AFTER scaffold, not before. | | 2 | Discovery Questions | Load + ask from ${CLAUDE_SKILL_DIR}/references/discovery-questions.md. Ask conversationally. workflow_style=guided → AskUserQuestion; autonomous → plain text. | | 3 | Load Defaults | From creator.yaml: license, version (always 1.0.0), author, quality target. | | 4 | Generate Scaffold | Structure only — files, YAML frontmatter, section headers with WHY comments, template vars for metadata. Never generate substantive prose. Leave [To be filled — run /fill] for expertise sections. See router.md for per-category structures. | | 5 | MCS-1 Structural Validation | Silently verify: required files present, metadata fields populated, README.md follows templates/readme/README.md.template structure (Hero, Install, "Is this for me?", Quick Start, Features, How It Works, type-specific, Requirements, Compatibility, Language, License — trilingual EN/PT-BR/ES with anchor nav), no YAML/JSON syntax errors. Auto-fix and note corrections. | | 6 | Print Next Steps — Cognitive UX | Load full UX stack: references/ux-experience-system.md (§1 context assembly, §2.3 moment awareness for "first scaffold"), references/ux-vocabulary.md (type naming), references/quality/engine-voice.md (brand DNA). Output is REASONED, not templated — adapt based on creator context. |

Step 6 Cognitive Output Protocol:

Assemble context from creator.yaml + STATE.yaml, then REASON about output:

IF is_first_product:
  → Full journey map. Warm tone. "Your first product exists."
  → Show all 4 pipeline steps explicitly.
  → Use encouraging language: "I'll guide you through each step."

IF products_published >= 5:
  → Compact. "✦ {name} scaffolded. {N} files. /fill when ready."
  → Skip journey map (they know the pipeline).
  → Surface only non-obvious: portfolio position, domain intelligence.

IF scout_source exists:
  → Reference the research: "Scaffolded from your {domain} research. {gaps_found} gaps mapped."

IF same domain has existing products:
  → Surface composition: "This joins {existing_slugs} in your {domain} portfolio."

ALWAYS:
  → Use ux_type_name from ux-vocabulary.md (e.g., "Deep Intelligence" not "minds")
  → Brand frame: ✦ marker, MyClaude Studio box (for first product or guided mode)
  → Pipeline position: "scaffold → ▸ fill → validate → test → package → publish"
  → Clear next action: "/fill" (always the next step after create)
  → VOCABULARY GUARD: Before emitting ANY creator-facing text, grep output for internal terms (MCS-N, DNA, D1-D20, scaffold, forge, substrate, habitable cell, intent_declaration, Sparring Protocol). If creator.profile.type is NOT "developer" or "hybrid", replace per ux-vocabulary.md. If "developer" or "hybrid", internal terms MAY appear but MUST be accompanied by context.

Default frame (guided mode or first 3 products):

┌─ MyClaude Studio ───────────────────────────┐
│  ✦ {product_name} — {ux_type_name}          │
│                                              │
│  {cognitive_insight based on context}        │
│                                              │
│  ▸ scaffold → fill → validate → test → ship │
│  Next: /fill                                │
└──────────────────────────────────────────────┘

Compact frame (autonomous mode or 5+ products):

For developers:

✦ {product_name} scaffolded. {N} files. {cognitive_insight}. /fill

For non-developers (profile.type != developer):

✦ {product_name} is ready. {N} files created. {cognitive_insight}. Add your expertise next → /fill

---

.meta.yaml Template

Create workspace/{product-slug}/.meta.yaml:

# Product metadata — not distributed, stripped during /package
product:
  slug: "{product-slug}"
  type: "{category}"               # skill|agent|squad|workflow|design-system|claude-md|application|system|bundle|statusline|hooks|minds|output-style
  created: "{YYYY-MM-DD}"
  mcs_target: "{MCS-1|MCS-2|MCS-3}"
  # NOTE: For type=minds, cap mcs_target at MCS-2 (Tier 3 DNA patterns are mostly N/A).
  # Show: "Minds ceiling is MCS-2 (Tier 3 DNA patterns not applicable). Setting target to MCS-2." (G015)
  # Minds-specific (only when type=minds):
  # minds_depth: "advisory|cognitive"
  # minds_sub_type: "self|genius|domain"  # cognitive only
  # genius_profile: "{name}"              # genius only

state:
  phase: "scaffold"                # scaffold | content | validated | packaged | published
  last_validated: null
  last_validation_score: null
  dna_compliance:
    tier1: null                    # 0-100
    tier2: null
    tier3: null
  overall_score: null
  last_tested: null
  test_result: null              # "pass" | "fail" | null
  test_scenarios: null           # "{passed}/{total}" | null

history:
  created_at: "{YYYY-MM-DD}"
  validated_at: []                 # append timestamps
  packaged_at: null
  published_at: null
  version: "1.0.0"

# Intent Declaration — the canonical record of how /create reasoned from creator intent
# to forged form. Written by /create at Step 11 of the 12-step algorithm (express or
# guided mode). Consumed by: /validate Stage 0 (Intent Coherence), /fill (natureza-aware
# section walks), /package (manifest shape), and STATE.yaml decisions_history
# (longitudinal feedback loop per capability-matrix.md §6).
#
# Schema source of truth: references/capability-matrix.md §3 Step 11 (19 fields).
# Enum sources:
#   - references/intent-topology.md §2 (delivery, nature, depth axis enums)
#   - references/intent-topology.md §4 (habitable cells v1)
#   - references/capability-matrix.md §3 Step 1 (verb_family enum)
#   - references/capability-matrix.md §3 Step 9 (unroutable_reason enum)
#   - references/capability-matrix.md §4 (mode enum: express | guided)
#
# Null-safe discipline (Contract C9): every field is populated on every forge. Fields
# not applicable to the current mode/path are set to null (single-value) or [] (list).
# Never omit a field — downstream consumers must not distinguish "absent" from "null".
#
# This block is the canonical home for scout_source (Activation Protocol step 9b).
intent_declaration:
  captured_at: null       # ISO-8601 when /create resolved the triple (Step 11)
  creator_said: null      # verbatim intent text (guided mode), or "(express mode)"
  mode: null              # express | guided (see references/capability-matrix.md §4)
  mode_switches: []       # appended by §5 transition protocol; each entry: {from, to, trigger, at_step, at_time}
  language: null          # mirror of creator.yaml → creator.language at forge time (drives locale-adaptive clause + announcements)
  scout_source: null      # "scout-{slug}.md" if a scout report informed this forge (canonical location for Activation Protocol step 9b)

  # engine_parsed — per-step outputs of the 12-step algorithm. In express mode, these
  # are populated from sub-command flags + type defaults. In guided mode, from the walk.
  engine_parsed:
    verb_family: null         # do_X | advise_on_X | coordinate_X | observe_X | enforce_X | react_to_X  (Step 1)
    continuity_bias: null     # parent | isolated | unclear  (Step 2)
    invocation_mode: null     # remembered | needs_auto  (Step 3)
    pollution_risk: null      # pollutes | safe  (Step 4)
    output_shape: null        # amplified_reasoning | structured_report  (Step 5)
    depth: null               # procedural | advisory | cognitive  (Step 6)
    nature: null              # executor | advisor | orchestrator | observer  (Step 7)
    delivery_mechanism: null  # ambient_constitutional | ambient_path_scoped | invoked_slash_command | invoked_task_spawn | reflex_hook_binding | composed_system  (Step 8 primary)
    host_set: []              # derived from (delivery, nature) per references/runtime-host-dag.md §2 — NEVER declared, always computed

  # Canonical cell lookup result (Step 9)
  matched_cell: null          # habitable cell id from intent-topology.md §4 (e.g. "reasoning_skill_cognitive"), or null if unroutable
  ranked_alternatives: []     # up to 2 alternative cell ids surfaced in the proposal

  # Creator-facing proposal (Step 10)
  proposed_form: null         # canonical_form from matched_cell, or null if unroutable
  creator_choice: null        # accepted | overridden
  override_to: null           # creator's chosen form if overridden
  override_reason: null       # creator's one-line explanation if overridden

  # Unroutable handling (Contract C4 — fall-through is a feature, not a failure)
  unroutable: false
  unroutable_reason: null     # no_habitable_cell | v2_cell_deferred | blocked_by_composition_gap | ambiguous_between_cells
  unroutable_gap_id: null     # gap marker from composition-anatomy.md (e.g. "GAP-COMPOSITION-1") when unroutable_reason == blocked_by_composition_gap

  # Audit markers for post-hoc schema evolution + default-tracking
  discriminators_applied: [] # which of [continuity, invocation, pollution, output, depth] actually fired (not just defaults)
  defaults_applied: []       # intent_profile field names that fell back to declared defaults (schema v3 gap tracking)

# Intelligence Layer fields (populated by /validate Stage 8, /scout, /package)
# Reference: references/intelligence-layer.md + config.yaml intelligence section
intelligence:
  domain: null                     # inferred from scout or content (e.g., "kubernetes-security")
  market_position: null            # blue_ocean | moderate | saturated
  value_score: null                # 0-12 co

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [myclaude-sh](https://github.com/myclaude-sh)
- **Source:** [myclaude-sh/myclaude-creator-engine](https://github.com/myclaude-sh/myclaude-creator-engine)
- **License:** MIT
- **Homepage:** https://myclaude.sh

Install and usage instructions live in the source repository linked above.