| name | blueprint |
| description | Turn a one-line objective into a cold-start-executable, multi-PR construction plan for work that spans several sessions or agents. |
| when_to_use | Use when a task is too big for one PR or one session — a feature that breaks into multiple PRs, a refactor or migration across many files/sessions, or work coordinated across sub-agents where context loss between sessions would cause rework. Triggers: "plan this", "give me a roadmap", "blueprint", "break this into PRs", "how do we sequence this", "multi-step plan". Skip when the task fits one PR / fewer than ~3 tool calls, or the user says "just do it" — use the `planner` agent or `/feature-dev` instead.
|
| allowed-tools | Agent Read Grep Glob Bash |
Blueprint
Source: concept adapted from affaan-m/ecc skills/blueprint (MIT). Wired to our planner/
architect agents and adversarial-review skill.
Produce a construction plan where any fresh agent can execute any step cold — without reading
prior steps or carrying session context. This is the layer above /feature-dev: /feature-dev
builds one feature in one session; blueprint sequences many of them across PRs and sessions.
The output is a single Markdown plan at docs/plans/<slug>.md (newest source of truth for the
effort — keep it current as the plan mutates, per latest-spec-wins.md).
Pipeline
1. Research (pre-flight + context)
git rev-parse --is-inside-work-tree 2>/dev/null && echo git || echo no-git
gh auth status 2>/dev/null && echo gh || echo no-gh
git remote -v; git branch --show-current
- With git + gh: plan a full branch → PR → CI workflow per step.
- Without them: direct mode — edit-in-place steps, no branches.
Read the existing project structure, any current docs/plans/*, docs/tasks.md, and docs/spec/*
so the plan builds on what's there rather than reinventing it.
2. Design (decompose)
Delegate decomposition to the planner agent (opus) for multi-file/architectural objectives:
Agent(subagent_type: "nextc-ecc:planner", model: "opus",
description: "Decompose objective into PR-sized steps",
prompt: "<objective> + repo context. Return PR-sized steps with dependency edges, parallel/serial ordering, model tier per step, and a rollback strategy per step.")
Each step is one-PR-sized (3–12 steps typical). Assign per step: dependency edges, whether it
can run in parallel (no shared files / no output dependency on an unfinished step), the model tier
to execute it (strongest for interface/design steps, default for mechanical ones), and a rollback
strategy.
3. Draft (write self-contained steps)
Write docs/plans/<slug>.md. Every step must be executable cold — it carries its own context
brief, so it declares what it consumes and what it produces:
## Step N — <title> [depends: Step K] [parallel-with: Step M] [model: opus|sonnet|haiku]
**Context brief:** what a fresh agent needs to know to do ONLY this step (the relevant files,
the invariant to preserve, the prior decision it builds on).
**Consumes:** exact inputs/signatures produced by earlier steps (so this step needs no other reading).
**Produces:** exact outputs/signatures later steps will consume.
**Tasks:**
- [ ] …
**Verify:** the commands that prove this step is done (build/test/lint/manual check).
**Exit criteria:** observable conditions that must hold before the step is "done".
**Rollback:** how to undo this step if it goes wrong.
Consumes/Produces are the contract that lets later-step agents work from their brief alone
instead of re-reading the whole plan.
4. Review (adversarial gate)
Do not ship the plan unreviewed. Run the adversarial-review skill on it, or delegate to the
architect agent (opus) against this checklist:
- Completeness — does executing all steps actually achieve the objective?
- Dependency correctness — no cycles; every
depends references a real earlier step; parallel steps
truly share no files/outputs.
- Cold-start integrity — can each step be executed with ONLY its brief +
Consumes?
- Anti-patterns — steps too big for one PR, vague exit criteria, missing rollback, hidden ordering.
Fix every critical finding before finalizing.
5. Register
Save the plan, add/refresh a pointer in docs/tasks.md, and report to the user: step count, the
parallelism summary (which steps can run concurrently), and the first step to execute.
Plan mutation protocol
Plans change as reality intrudes. When that happens, edit docs/plans/<slug>.md in place and record
the change — never let the plan drift from what's actually being built (latest-spec-wins.md):
- Split / insert / reorder / skip / abandon a step explicitly, with a one-line reason appended to
a
## Changelog section at the bottom of the plan file.
- Re-validate the dependency graph after any structural change.
Anti-patterns
- Steps that can't be executed without reading other steps (no real context brief).
- Steps bigger than one PR ("implement the whole backend").
- Vague exit criteria ("works correctly") instead of observable/verifiable conditions.
- Skipping the review gate because the plan "looks fine."
- Leaving the plan file stale after the work diverges from it.
Relationship to other skills
planner agent — does the decomposition inside step 2; blueprint adds cold-start briefs,
dependency graph, PR workflow, and the review gate around it.
/feature-dev — executes a single step/feature. Blueprint sequences many.
adversarial-review — the step-4 gate for the plan itself.
/clarify — run first if the one-line objective is ambiguous.