Generate Executable Code Project Plans

Blueprint generates plans for large coding projects - multi-step refactors, plugin extractions, architecture migrations, phased feature rollouts - as plain Markdown files where every step carries enough context to be executed cold, by a different agent in a different session. Each step is one-PR sized, with a task list, rollback strategy, verification commands, and exit criteria; the plan also includes a dependency graph (which steps can run in parallel vs. serially), locked design decisions, invariants checked after every step, a progress log as the single source of truth for execution state, and a review log from an independent adversarial review.

Use it for projects too large for a single agent session, where steps need to be picked up cold by a fresh agent that hasn't read the conversation history or prior steps. It's explicitly a different tool than Claude Code's built-in /plan (good for single-session, quick, exploratory work) or persistent-planning tools that solve context loss with session-catchup scripts - Blueprint targets multi-session, multi-agent projects specifically, with cold-start context briefs, an adversarial review gate, and a branch/PR/CI workflow built into every step.

Run /blueprint myapp "migrate database to PostgreSQL" to scan the codebase, design a step-by-step plan, review it with a strongest-model sub-agent (falling back to the default model if unavailable), and save it to plans/myapp-migrate-database-to-postgresql.md. Each step's Context field is a cold-start brief describing exactly what state the codebase is in and what the step does; tasks are marked [exact] (run this command verbatim), [guided] (agent decides the approach), or [open] (goal only, full autonomy). Plans are living documents: steps can be split, inserted, skipped, reordered, or abandoned through formal protocols with an audit trail, and a new session reads the Progress Log plus branch/file state to resume exactly where the last session stopped.

Installed into Claude Code by cloning into ~/.claude/skills/blueprint, after which /blueprint is discovered automatically; since generated plans are standard Markdown, any agent that can read files can execute one by treating SKILL.md as a system prompt. Git and the GitHub CLI unlock the full branch/PR/CI workflow, but Blueprint detects their absence and falls back to a direct mode instead of failing - the same graceful-degradation pattern applies if the strongest review model or project memory is unavailable.

Developers running large, multi-session projects across multiple agent instances who need each step independently executable without shared conversation context, plus an adversarial review pass and a real dependency graph for parallelizable work - beyond what a single-session plan or a simple persisted-context tool provides.