start

name	start
description	Use when starting curdx-flow, creating a spec, resuming work, or routing intent.
argument-hint	[name] [goal] [--fresh] [--quick] [--mode auto|fast|deep] [--task-granularity auto|coarse|standard|fine] [--review minimal|standard|strict] [--commit-spec] [--no-commit-spec] [--specs-dir <path>]
allowed-tools	Read Write Edit Bash Agent Skill AskUserQuestion
disable-model-invocation	true

Smart Start

Act as the curdx-flow router. Decide the next action from facts first; ask the user only when a fact is missing or the action is destructive.

Dev Context

Before routing, assume the current project may be frontend-only, backend-only, full-stack monorepo, split frontend/backend repos, a CLI/library, or a Claude Code plugin. The deterministic router reads project CLAUDE.md Dev sections, .claude/curdx-flow.local.md, and cheap manifests to infer code roots.

Users do not need to know --add-dir. If routing returns a missing code root, show the exact nextAction from the router and stop.

Route First

1. Parse $ARGUMENTS into optional spec name, goal text, and flags.
2. Run the deterministic router:

   curdx-flow route \
     --name "$name" \
     --goal "$goal" \
     --flags "$ARGUMENTS"
   

   For routes that are not direct-change or blocked-ask-user, compile the execution contract before planning:

   curdx-flow route --compile --record \
     --name "$name" \
     --goal "$goal" \
     --flags "$ARGUMENTS"
   

3. Run the last-mile autopilot:

   curdx-flow last-mile \
     --name "$name" \
     --goal "$goal"
   

   Treat phase, problemTypes, capabilityPlan, evidenceRequired, and coordinatorInstruction as the automatic execution policy. Do not ask the user which skill to run when the autopilot already identifies the needed capability.
4. Treat the returned route and intent as the source of truth. Do not invent a different workflow unless the router says blocked-ask-user and the user's answer changes the facts.
5. Always read these adaptive facts before acting:
   • topology.workspaceState: empty, scaffolded, existing, or split-repo
   • intent.intentKind: scaffold, product, prototype, import-spec, feature, fix, refactor, release, or unknown
   • intent.clarity, intent.stackSpecified, intent.artifactProvided, intent.deliveryExpectation, and intent.missingFacts
6. If the router returns recommendedCapabilities, treat them as phase-specific hints, not mandatory steps. availabilityState is the operational signal: available means visible now, expected means the environment should provide it, and missing means skip or fix setup before relying on it. availability: plugin-dependency means a Claude Code plugin dependency; availability: external-expected means an MCP provided by the user's external setup such as context7 or sequential-thinking. curdx-flow recommends existing wheels (context7, claude-mem, ui-ux-pro-max, chrome-devtools-mcp, sequential-thinking, pua) and must not reimplement or bundle duplicate MCP config for them. Workflow/policy hints such as docs-query, tdd-cycle, security-review, stack-specific-verification, and context-budget need no installation.
7. For last-mile phase/capability/evidence interpretation, use ${CLAUDE_PLUGIN_ROOT}/references/last-mile-autopilot.md.
8. For stack profile, quality gates, suggested verifier, and context-budget interpretation, use ${CLAUDE_PLUGIN_ROOT}/references/intelligent-routing.md only when the compact router output is insufficient.
9. For execution brief fields, completion contract, and .curdx/brain.jsonl interpretation, use ${CLAUDE_PLUGIN_ROOT}/references/execution-brief.md only when the compact brief output is insufficient.

Route Actions

Route	Action
resume-current	Resume the active spec at nextAction. Do not create a new spec.
direct-change	Handle the change directly in the current turn. Do not create a spec or tasks.md.
lite-spec	Create a lightweight spec, then generate 1-3 value-slice tasks.
full-spec	Run the normal research -> requirements -> design -> tasks -> implement workflow.
epic-split	Invoke /curdx-flow:triage with the same goal. Do not force the work into one spec.
scaffold	Select the best scaffold source, create only the explicitly requested skeleton, write assumptions, then run baseline verification.
product-inception	Create product context before application code: mission, constraints, roadmap, tech-stack assumptions, and constitution.
greenfield-spec	Create a greenfield spec with product context, technical plan, walking skeleton, and vertical-slice tasks.
prototype	Create a bounded prototype spec with an explicit success criterion and minimal verification loop.
import-spec	Import the provided PRD/spec/design/API artifact into curdx-flow phase artifacts, then plan implementation.
blocked-ask-user	Ask one focused question, then rerun the router with the new fact.

If blocked-ask-user includes topology.missingRoots, do not ask an open-ended question. Print the router's Next line exactly, such as /add-dir ../frontend, and tell the user to rerun /curdx-flow:start after adding the directory.

Pre-Question Discovery (Greenfield / Inception / Blocked)

**For `product-inception`, `greenfield-spec`, `prototype`, `import-spec`, or `blocked-ask-user` routes with non-empty `intent.missingFacts`, you MUST fan out a Discovery subagent batch BEFORE asking the user any clarifying question and BEFORE writing any product-context artifact.**

This is the Discovery domain of ${CLAUDE_PLUGIN_ROOT}/references/bounded-parallel-dispatch.md. The independence rules and anti-patterns in that doc apply here.

Why

Asking the user before fanning out is an anti-pattern (#14 in bounded-parallel-dispatch). It serializes work that should be parallel, surfaces low-leverage questions whose answers are already in the codebase or memory, and produces interview options without [Recommended] grounded in facts. Anthropic's canonical guidance frames subagents as the primary tool for investigation precisely because main-context exploration "fills with file contents you won't reference again."

When this section fires

All of the following must hold:

1. Route ∈ {product-inception, greenfield-spec, prototype, import-spec, blocked-ask-user}
2. intent.missingFacts is non-empty (parse from curdx-flow route output)
3. At least one missingFact is answerable from: prior memory, this repo's files, or current external docs

If route is blocked-ask-user AND topology.missingRoots is the only missing fact, skip this section and follow the /add-dir instruction above. Missing-roots is a deterministic user action, not a discoverable fact.

Fan-out composition (minimum 3 agents in ONE message)

For empty workspaces and product-inception, dispatch ALL of these in a single message:

1. general-purpose agent invoking mcp__plugin_claude-mem_mcp-search__smart_search
   → Search user's claude-mem for prior decisions, failed attempts, similar projects
   → Output cap: 250 words, concrete corpus/observation references

2. Explore agent rooted at .planning/, docs/, CLAUDE.md, .gitignore, package.json
   → Surface existing conventions, prior planning notes, declared stack
   → Output cap: 200 words, file:line references

3. N research-analyst agents — ONE per candidate stack OR prior-art topic in missingFacts
   → Each agent owns ONE topic; never combine
   → Output cap: 300 words per agent, cite sources
   → If a candidate stack is named (e.g., "node-pty", "xterm.js"), prefer Context7 MCP via the agent's tool surface before generic web search

For non-empty workspaces or blocked-ask-user with codebase-resolvable facts, drop the research-analyst agents if missingFacts are all repo-internal.

Synthesis contract

After all agents return, the coordinator MUST:

1. Cross-validate memory hits against current repo state — claude-mem snapshots can be stale. A function foo() referenced in memory may have been renamed; treat memory as "what was true when written", not "what is true now". Verify with grep/Read before recommending.
2. Map every missingFact to evidence: which agent answered it, with what citation. Unanswered facts become the user-decision interview questions.
3. Re-derive interview options with [Recommended] grounded in the synthesis. No fabricated options; every option must trace to a finding or an explicit assumption marker.
4. Persist findings to .planning/discovery/<route>-<timestamp>.md if .planning/ exists, so re-runs and other sessions can reuse them.

What this section is NOT

• Not a license to write product-context artifacts directly. Artifact writing still goes through the normal phase agents (research-analyst / product-manager / etc.) once a spec is created.
• Not a replacement for the interview. It produces grounded interview options; the user still confirms the irreducible decisions.
• Not for direct-change, scaffold (with clear stack), lite-spec, full-spec, epic-split, or resume-current routes. Those have their own dispatch patterns elsewhere.

Hard Rules

• Use behavior route names exactly as returned: direct-change, lite-spec, full-spec, epic-split, scaffold, product-inception, greenfield-spec, prototype, import-spec, resume-current, blocked-ask-user.
• Top-level tasks are value slices. Never split a slice into separate "write test", "write implementation", "run test", or "commit" tasks.
• If the route says direct-change, skip branch prompts, spec creation, phase documents, task planning, and subagents unless the user explicitly asks for them.
• If the route says scaffold, do only what the user explicitly asked to scaffold. Prefer official or ecosystem-maintained generators for the named stack when current docs show one exists; self-author the skeleton only when that is safer, smaller, or no trustworthy generator exists. Record assumptions in CLAUDE.md or .curdx/assumptions.md when no project convention exists, then run the best detected baseline command through curdx-flow dev verify or the project equivalent.
• If the route says product-inception, do not write application source yet. Produce compact product context artifacts and ask at most the missing high-leverage questions from intent.missingFacts.
• If the route says greenfield-spec, bootstrap curdx-flow artifacts before app code: product context, constitution, requirements, design, tasks, then implementation.
• If the route says prototype, constrain scope to a success criterion; do not silently expand it into a production product.
• If the route says import-spec, preserve source artifact traceability in research/requirements/design/tasks rather than re-interviewing from scratch.
• If recommendedCapabilities includes tool, workflow, or policy hints, use the recommendation only at its listed phase and only when it materially reduces uncertainty, improves context efficiency, or verifies real behavior. If a future recommendation is marked check-if-installed, skip silently when that capability is absent.
• If the route says epic-split, stop single-spec creation and run triage.
• If a spec state is created, store autoPolicy for compatibility and set maxGlobalIterations from policy, defaulting to 30 if the helper fails.

Empty Workspace Rules

An empty folder is not a small feature request. When topology.workspaceState == "empty", route by intent:

• scaffold: the user named a stack or starter and expects files now.
• product-inception: the user has a product idea but missing domain, user, MVP, or acceptance facts.
• greenfield-spec: the product and stack are clear enough to plan a usable app.
• prototype: the user wants a demo/POC/spike.
• import-spec: the user supplied or referenced a PRD, design, OpenAPI, or other source artifact.

For greenfield work, create or preserve these context artifacts before coding:

docs/mission.md          # product/user/problem
docs/roadmap.md          # first usable slice and later slices
docs/tech-stack.md       # chosen stack and assumptions
docs/constitution.md     # quality, testing, security, UX, deployment rules

If the user already has equivalent files, reuse them. Do not create a separate methodology folder when the repository has a better convention.

The first implementation task in a greenfield app must be a walking skeleton: selected runtime roots, contract, dev environment, and verification command run together. For example, a full-stack app should prove a real UI route can call a real service health/status endpoint before business features begin.

Use ${CLAUDE_PLUGIN_ROOT}/references/greenfield-delivery.md as the detailed contract for these artifacts and tasks.

Scaffold Source Selection

For scaffold, choose the source before writing files:

1. Identify the requested target shape: frontend app, backend service, full-stack workspace, CLI, library, plugin, data tool, or another explicit shape from the user.
2. When the user named a stack or framework, check the latest official docs or trusted ecosystem docs for the current scaffold command/API before running it. Treat package-manager create flows, documented remote initializer APIs, framework CLIs, templates, and project-local generators as candidates. Do not make any single stack special in the route logic.
3. Prefer non-global, repeatable commands (npm create, pnpm create, npx, package-manager dlx, curl to a documented initializer API, or a project-local generator). Avoid installing global CLIs.
4. If the generator is interactive and missing choices materially affect the result, ask one focused question. Otherwise choose conservative defaults from the user's request and record them.
5. Self-author the scaffold only when no trustworthy generator exists, the generator cannot satisfy the requested constraints, the requested skeleton is intentionally smaller than the generator output, or a custom skeleton is clearly easier to verify. Record the reason.
6. After generation, normalize scripts and run baseline verification. Do not claim the scaffold works without evidence.

New Spec Creation

For lite-spec, full-spec, greenfield-spec, prototype, and import-spec:

1. Ensure the current branch is appropriate using ${CLAUDE_PLUGIN_ROOT}/references/branch-management.md.
2. Validate or ask for a kebab-case spec name only if missing.
3. Validate or ask for the goal only if missing.
4. Resolve the spec directory from --specs-dir or the default specs dir.
5. Create the spec directory, update $defaultDir/.current-spec, and ensure .gitignore covers $defaultDir/.current-spec, $defaultDir/.current-epic, and **/.progress.md.
   • If writing the marker under the default specs dir, write only the bare spec name, e.g. greet-helper.
   • If writing a non-default spec root, write the relative path, e.g. packages/api/specs/auth-flow.
   • Do not write specs/<name> for default-root specs; that can be interpreted as a path by newer runtime resolvers.
   • Do not write a project-root .current-spec; runtime state lives under the configured specs root.
   • Preferred write pattern:

     mkdir -p "$defaultDir"
     if [ "$basePath" = "$defaultDir/$name" ]; then
       printf '%s\n' "$name" > "$defaultDir/.current-spec"
     else
       printf '%s\n' "$basePath" > "$defaultDir/.current-spec"
     fi
     

6. Compute compatibility policy:

   ROUTE_JSON=$(curdx-flow route --name "$name" --goal "$goal" --flags )
   POLICY_JSON=$(   | node -e )
   

If policy computation fails, use this fallback cap:

{ "maxGlobalIterations": 30 }

For lite-spec, keep interviews minimal and generate only 1-3 value-slice tasks. For full-spec, continue with research and the normal phase flow. For greenfield-spec, include product context and constitution before design, then require a walking-skeleton task before feature slices. For prototype, keep tasks bounded to the success criterion. For import-spec, keep traceability to the imported source artifact in every phase artifact.

When a spec state is created and router output includes topology, intent, or recommendedCapabilities, store compact copies in .curdx-state.json as projectTopology, intent, and recommendedCapabilities. If last-mile output is available, store a compact lastMile object with phase, problemType, problemTypes, capabilityPlan, evidenceRequired, and lastDecisionAt.

Quick Artifact Contract

When --quick causes this skill to generate phase artifacts inline instead of delegating through each phase command:

The quick path does not lift the runtime-CLI contract for `.curdx-state.json`. Even in `--quick --mode fast`, do not Write/Edit the state file beyond the initial schema in "New Spec Creation" step 7. Verification blocks must be produced by `curdx-flow verify run`, and completion/iteration mutations must use `curdx-flow state merge "$basePath/.curdx-state.json" ''`. Hand-written `verificationBlocks` invariably miss `srcMtime` and break `check-verification-blocks` / `verify-blocks` invariants.

• tasks.md must still use the task-planner format contract.
• Start tasks.md with ## Source Coverage Audit.
• Every executable top-level task must be a checkbox list item, not a heading:

- [ ] 1.1 Implement greet helper
  - **Do**:
    1. Edit `src/greet.js`
    2. Run `npm test`
  - **Files**: `src/greet.js`, `test/greet.test.js`
  - **Done when**: `greet(" Ada ")` and empty-name fallback pass
  - **Verify**: `npm test`
  - **Commit**: `feat(greet): implement greeting helper`
  - _Requirements: FR-1, AC-1_

• Do not create heading-only task sections such as ## T1; runtime task parsing ignores them and /curdx-flow:status will report empty-tasks.
• If implementation is completed in the same quick run, run the final task's exact Verify command through the runtime recorder before marking completion:

  curdx-flow verify run --phase execution --command "<exact Verify command>" --spec "$name"
  

  This command actually executes the verification, records .curdx-state.json::verificationBlocks.execution with the command, exit code, timestamp, and source mtime, and returns the verification exit code. If the verification fails, do not set completed: true; fix the issue and rerun the recorder.
• If the goal, requirements, or acceptance criteria require browser evidence, Playwright, or Chrome DevTools MCP, --check/file-existence commands are not sufficient completion evidence. The quick run must first perform the real browser interaction, then record the shell verifier with --description containing the concrete browser proof: tool used, URL, action sequence, final DOM assertion, and console-error result. Do not mark completion when that proof is absent.
• After recorded verification passes, mark the checkbox [x], set completed: true, and leave taskIndex equal to the number of top-level checkbox tasks.

Skill Discovery

Only scan and invoke additional skills when the route is lite-spec, full-spec, greenfield-spec, prototype, or import-spec. Match skills by semantic relevance to the goal. Skip discovery for direct-change and scaffold unless the user explicitly asks to use a skill.

Output

Always start with a short routing summary:

Route: <route>
Reason: <reason>
Next: <nextAction>

Then perform the next action. If the route is blocked-ask-user, ask exactly one focused question.