The interface is localized. Skill descriptions and documents are machine-translated by AI into your language; code, commands, paths and technical terms stay in the original, and the author's document governs.
Explore skills
Productivity · SKILL.md

spec-implementation-planner

spec-implementation-planner

Turns approved specs into waves, dependency indexes, AFK tickets, and status tracking. Use when specs need parallel-agent planning, ticket readiness, blockers/unblocks, or plan gap checks.

Independent GitHub snapshot · Automated classification · Original documentation governs

Ssebastianwessel / skillsApache-2.0

Skill documentation

GitHub

name: spec-implementation-planner description: Turns approved specs into waves, dependency indexes, AFK tickets, and status tracking. Use when specs need parallel-agent planning, ticket readiness, blockers/unblocks, or plan gap checks.

Spec Implementation Planner

Turn approved specs into vertical-slice waves and implementation-ready AFK tickets agents can execute in isolation without guessing.

Hard Gate

Read specs/.readiness-report.yaml. Stop unless:

  • status: approved
  • human_approval.status: approved
  • approval_evidence.manifest_digest equals spec-manifest.yaml.content_digest
  • readiness/gate simulation passed or can be rerun successfully
  • checklist_walk_gate.status: passed with zero blocking findings
  • end_to_end_definition_gate.status: passed
  • current_dependency_research_gate.status: passed
  • representation_reuse_gate.status: passed when data shapes are in scope
  • language: en and semantic_judge_gate.status: passed when deterministic English smoke checks are used

If blocked, write a spec gap/readiness note. Do not create executable tickets.

Plan only from approved specs, ticket rules, contracts, and verification. Otherwise create a spec/plan gap.

Workflow

  1. Verify all ticket fields and action-plan steps are fillable from approved specs.
  2. Create plans/implementation-plan.md, plan-manifest.yaml, and structurally reconciled _registry, _status, _dependencies, and _scope. Use only the strict YAML subset in references/planning-gates.md; never rely on prose or a global N/A to satisfy a ticket gate.
  3. Prefer vertical slices. Each wave ends with a working, testable end-to-end increment. If too large for one agent, split isolated tickets that converge into the same wave result.
  4. Use horizontal waves only for approved foundation/interface work or reliability refactors; document rationale, unblocks, tests, and next vertical slice.
  5. When specs select a contract-first clean rebuild, plan the generated foundation next to the old boundary and block handwritten service work until mapping metadata, generators, generated tests, compile checks, and drift checks pass.
  6. Plan contract/codegen foundations before parallel backend/client/adapter work; keep parallel writes isolated and sidecar agents read-only unless their write scopes are disjoint.
  7. Plan test-driven order: derive unit, contract, integration, and E2E tests from specs/contracts/acceptance/unhappy paths before business logic tickets implement the behavior those tests prove.
  8. Map the approved 00-traceability.yaml requirement, capability, path, and acceptance IDs into tickets and local testable acceptance rows; collectively cover every non-N/A source ID. Map capability inventory rows and end-to-end definition chains into tickets: actor/consumer, entrypoint/reachability, contracts, data lifecycle, states, side effects, permissions, recovery, observability, acceptance, verification, NFRs, operations, supply chain, frontend/client UX, design/component reuse, tests, and coverage.
  9. Split tickets until each active ticket has one bounded deliverable, exact write scope, exact prerequisite artifacts, and a numbered action plan with files/commands/proof for every step. Use phase-gated exceptions only when the approved work is atomic and each phase blocks on explicit evidence.
  10. Require ticket sections for spec-drift controls, generator-first artifacts, representation catalog/reuse, TDD including unhappy paths, strict typing, modular domain structure, reuse/no-duplication, and review against ticket plus specs. Domain structure and dependency choices must follow the approved readiness evidence.
  11. Track the lifecycle planned -> ready -> in_progress -> implemented -> review_pending -> accepted; use partial, blocked, and skipped only as explicit side states. Maintain exact depends_on, blocked_by, and unblocks in the ticket and every plan index.
  12. Self-audit vertical-slice completeness, generation-map coverage, strong boundary types, path/NFR/frontend/release ownership, parallel risk, fake-work risk, and blockers.
  13. Run references/planning-gates.md, then node <spec-implementation-planner-skill-root>/references/generate_plan_manifest.mjs <repo-root> [plans-root] [specs-root], then node <spec-implementation-planner-skill-root>/references/check_plan.mjs <repo-root> [plans-root] [specs-root].

Ticket Contract And Handoff

references/planning-gates.md owns the exact structural ticket contract, autonomy, command, traceability, generated-contract, reuse, action-step, acceptance, lifecycle, and anti-pattern rules. Tickets pin both content digests and may authorize only D0/D1 work. Missing semantic evidence is a spec gap; never emit a ticket that asks an implementer to decide, infer, or invent.

After accepted, keep work historical. New gates create remediation/migration tickets; use blocked, partial, skipped, superseded_by, and resume notes explicitly. Ready tickets go to implementation; waves go to independent review.

Saved skills