Blueprint Plan Session Skill
Asks the right clarifying questions before you code, turning your idea into an executable plan for your coding agent.
Evidence shows a read-only planning Q&A flow that explicitly forbids writing code files, requires user decisions, has no dependencies, external calls, or sensitive-data handling; deducted because confirmation/rollback are only workflow conventions rather than enforced guarantees, and publisher identity is unverified.
SKILL.md is self-consistent and all referenced files (templates., questions.md, refine-prompt.md) exist with matching formats; deducted for absence of tests or key-path reproduction evidence — static cap is 10 — and only descriptive guidance for edge-case parsing of shorthand answers.
Fit and non-fit ranges are stated in README (greenfield/large features vs. visual frontends, small refactors); trigger description is precise; deducted for English-only interaction with no Chinese support, reliance on a compatible agent harness, and no mainland-China reachability considerations.
MIT license is clear and docs are well layered (SKILL.md + references) with customizable templates; deducted for missing versioning and changelog, the hidden assumption that templates. must be manually synced across two skill directories, and only indirect maintenance/update-path signals.
README gives a clear value proposition and competitor comparison; output is a directly usable markdown plan file; deducted because static cap is 7, with no executed plan artifacts or third-party evidence of output quality beyond author claims and user quotes.
Auditable primary material exists (full skill source, templates, license, README); deducted for no test suite or CI coverage of the skill's key paths and unverifiable user testimonials — static cap is 5.
- This is a static source review; no skill flow was executed, so actual Q&A quality and plan generation remain independently unverified.
- The skill is English-only with no declared Chinese-language support; mainland-China users may face usability friction.
- templates. must be manually kept in sync between the blueprint and blueprint-generate skill directories; custom templates risk divergence.
- blueprint depends on the companion blueprint-generate skill to produce the final plan; it cannot deliver a plan on its own.
What does this skill do, and when should you use it?
Blueprint, by imbue-ai, is a planning copilot for coding agents; this entry covers only the blueprint skill — the plan-session half of the repository. Given a feature description, it has you pick a plan template, genuinely explores your codebase, then runs multi-round Q&A with 3-5 easy-to-answer questions per round. Your answers progressively refine a prompt that the companion blueprint-generate skill turns into a plan when you invoke it. The skill gathers information only: it never ends the Q&A itself, never writes the plan, and never modifies code.
Parses your feature description; reads references/templates. and asks you to pick a template (with an Other option for custom structures); actually reads project source to understand structure, key modules, existing patterns, and conventions; asks 3-5 clarifying questions per round grounded in what it found, accepting shorthand (e.g. 1a, 2b, 3e) or free-form answers; after each round merges your answers into a refined prompt shown as bullet points; displays an Explore/Plan/Write/Refine progress line on every message after template selection. It does not stop the Q&A on its own, does not generate the plan, and does not touch code files.
- Engineers building a sizable new feature on an existing codebase who need design clarity before coding
- Developers starting greenfield projects who want their idea turned into an executable spec
- Teams making incremental changes large enough to warrant a plan before handing it to an agent
- Researchers running experiments on new models, systems, or subsystems
What are this skill's strengths and limitations?
- Hands the tedious parts of planning to the agent while you keep every decision
- Questions are grounded in real codebase exploration and never ask what the code already answers
- Planning depth is user-controlled — generate after one round or keep going for many
- Plan structure and depth are template-driven, and custom templates are supported inline
- Does not generate the plan itself — it requires the companion blueprint-generate skill
- Question quality depends on choosing a matching template; a poor pick skews the Q&A perspective
- No evidence in the source of an automated test suite or cross-platform compatibility validation
- Poor fit for visual-decision-heavy frontend work and small refactors
How do you install this skill?
Install the skill collection: npx skills add imbue-ai/blueprint. For Claude Code you can alternatively run curl -fsSL https://raw.githubusercontent.com/imbue-ai/blueprint/main/install-skills.sh | bash, which copies the skills into ~/.claude/skills/ (or $CLAUDE_CONFIG_DIR/skills if set).
How do you use this skill?
In a compatible agent, invoke /blueprint <description> (e.g. /blueprint Add a caching layer to reduce API calls). Pick a template; the skill explores your codebase and starts asking questions. Answer with shorthand like 1a, 2b, 3e or in prose. When you've covered enough ground, invoke the companion blueprint-generate skill to end Q&A and produce the plan — this skill alone does not write the plan.