From SDD Engineering
Orchestrates the PLAN phase of Spec-Driven Development — the thin main-thread wrapper around the read-only implementation-planner agent, mirroring how sdd-engineering:write-spec wraps spec-creator. Takes an APPROVED spec and plans its IMPLEMENTATION: invokes implementation-planner to produce the file-level HOW, relays the planner's clarifications via AskUserQuestion, fans out researcher agents for any [RESEARCH NEEDED] gap, confirms multi- vs single-agent execution, and — the thing the read-only planner cannot do — PERSISTS the plan to the plan directory (default plans/, files PLAN-*.md). Runs BETWEEN sdd-engineering:write-spec (WHAT) and sdd-engineering:run-plan (build). Orchestrates only; writes no code and authors no spec.
How this skill is triggered — by the user, by Claude, or both
Slash command
/sdd-engineering:plan-implementationWhen to use
Trigger phrases: '/plan-implementation', 'plan the implementation', 'turn the spec into a plan', 'run the planner', 'make the implementation plan'. Runs AFTER an approved spec exists (sdd-engineering:write-spec) and BEFORE sdd-engineering:run-plan. Requires an approved SPEC-*.md in the spec directory — if the spec is still draft/has open NCs, finish sdd-engineering:write-spec first. Explicit, user-invoked.
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
You **orchestrate** the PLAN phase — you do NOT write the plan yourself. The read-only
You orchestrate the PLAN phase — you do NOT write the plan yourself. The read-only
implementation-planner agent produces the file-level HOW; your job is the main-thread work a
read-only subagent cannot do: talk to the user (AskUserQuestion), fan out researchers for
its information gaps, confirm the execution mode, and persist the plan to the plan directory
(default plans/; the planner cannot write). This is the exact mirror of how
sdd-engineering:write-spec wraps spec-creator.
You sit between the two other SDD phase skills: sdd-engineering:write-spec (the WHAT) upstream,
and sdd-engineering:run-plan (build → review → fix → gate) downstream, which consumes the plan
you persist.
approved; the plan (HOW) is produced by the planner; code is sdd-engineering:run-plan's job.PLAN-*.md artifact (in the plan directory, default plans/)
is the whole point: it survives the chat boundary and is what sdd-engineering:run-plan and
plan-verifier consume. HOW lives in the plan directory, never in the spec directory.- [ ] 1. INTAKE — take the approved spec path (a SPEC-*.md in the spec directory, default specs/). If it is
still `draft` or has open [NEEDS CLARIFICATION], STOP and route the user to sdd-engineering:write-spec
to finish it first — do not plan against an unapproved spec.
- [ ] 2. PLAN — invoke the `implementation-planner` agent (Task) with the approved spec path + the instruction
to reuse the spec's AC-N ids verbatim. It returns a file-level plan (task units, parallelization graph,
skills, known pitfalls, test plan) plus its handoff markers.
- [ ] 3. CLARIFY — act on the planner's markers IN THE MAIN THREAD:
• "Open questions / clarifications" → ask the user via AskUserQuestion (batch related, recommend a default).
(The planner may try to ask directly; if it instead returns questions as markers, you relay them — same
graceful degradation as write-spec, since a subagent cannot reliably prompt the user.)
• [RESEARCH NEEDED: …] → fan out one or more `researcher` agents in parallel (main-thread only — the
planner cannot); feed their cited findings back by re-invoking the planner for a follow-up pass.
• Execution mode → confirm multi- vs single-agent (the planner recommends; you confirm via AskUserQuestion).
Loop 2–3 until the plan has no blocking gap.
- [ ] 4. PERSIST — write the plan to `PLAN-<SPEC-ID>-<slug>.md` in the plan directory (default `plans/`; you
do this — the planner is read-only). Reuse the spec's SPEC-NN id + slug so spec ↔ plan ↔ code stay traceable.
- [ ] 5. GATE — plan approval: show the plan summary (task units + tracks + parallel groups, scope, contract/
schema changes, risks) and ask the user Approve / Adjust before any build. Do not hand off an unapproved plan.
- [ ] 6. (OPTIONAL) RETRO — if this phase fanned out several agents (planner re-invocations, researchers), offer
**`sdd-engineering:retro`** while the per-agent telemetry is still fresh in context, before it scrolls out.
- [ ] 7. HANDOFF — point the user to `sdd-engineering:run-plan` on the persisted `PLAN-<SPEC-ID>-<slug>.md`
for build → review → fix → gate.
AC-N (ids
reused verbatim, each with its Verify: hint) — so plan-verifier gets a clean forward pass.[RESEARCH NEEDED] items you resolve (fan out researchers), never guesses
baked into the plan.Orchestration only. No spec authoring (that is sdd-engineering:write-spec), no code (that is
sdd-engineering:run-plan), no plan writing by the planner (it is read-only — you persist). Keep the
plan directory for the HOW, the spec directory for the WHAT.
Converse in the language of the request; keep the plan body's paths, identifiers, commands, skill
names, and AC-N ids verbatim.
npx claudepluginhub rostk/my-monkeys --plugin sdd-engineeringCreates structured, bite-sized implementation plans from specs or requirements before writing code. Useful for breaking down multi-step tasks into testable steps with file structure and task boundaries.