Skip to content
/

Blueprint (Imbue)

blueprint-imbue · imbue-ai/blueprint · ★ 63 · last commit 2026-05-19

Slows down AI coding just enough to ask codebase-grounded clarifying questions before writing the implementation plan, with user-controlled Q&A termination.

Best whenThe highest-value human contribution to AI coding is decisions, not enumeration — the agent enumerates options and asks; the human decides; the agent never s…
Skip ifAsking questions answerable by reading code, Agent auto-terminating Q&A before user is satisfied
vs seeds
spec-kit(clarification loop → spec output) but uses explicit two-skill phase separation (blueprint Q&A only / blueprint-generate…
Primitive shape 2 total
Skills 2
00

Summary

Blueprint (Imbue) — Summary

Blueprint (imbue-ai/blueprint) is a two-skill planning copilot that slows down just enough to ask the right questions before any code is written, then produces a structured implementation plan any coding agent can execute. The pair of skills — blueprint (multi-round Q&A) and blueprint-generate (plan output) — are deliberately minimal: 63 stars, two markdown SKILL.md files, a JSON templates file, and a reference directory. The skill is installed via npx skills add imbue-ai/blueprint (the agentskills.io skills CLI) or a one-line curl. Its distinctive philosophy is that planning should be iterative and human-directed — the agent explores the codebase, generates 3-5 questions grounded in actual code findings, waits for answers, then asks 3-5 more; the user decides when Q&A is done by invoking blueprint-generate. This separation of "gathering phase" (blueprint) and "writing phase" (blueprint-generate) is enforced by skill design: blueprint is explicitly blocked from generating the plan, and blueprint-generate is blocked from asking questions. Compared to seeds, it most resembles spec-kit's speckit.clarify + speckit.specify split but is narrower in scope (plan only, no task decomposition, no implementation) and richer in the Q&A quality rules (codebase-grounded questions only, no questions answerable by reading code, shorthand answer format).

01

Overview

Blueprint (Imbue) — Overview

Origin

Created by Imbue AI (imbue-ai/blueprint). No license file in root (LICENSE file present but not read). 63 stars. Pushed 2026-05-19. Also published on VS Code Marketplace as Imbue.imbue-blueprint.

Philosophy

From README:

"Most coding agents rush to code or guess at the plan. Blueprint slows down just enough to ask the right questions." "You stay in control without doing the tedious work. Blueprint splits planning between you and the agent. The initial idea comes from you. Enumerating all the considerations and choices comes from the agent. Decisions come from you." "Questions that make you think. Blueprint asks questions that surface real design choices — the kind that engage you and make you think about what you actually want."

Blueprint is built on the premise that the highest-value contribution a human can make to AI coding is the decisions (not the enumeration of options). The agent handles codebase exploration and question generation; the human handles decisions; neither does the other's job.

Key design decisions

  1. Codebase-grounded questions only: Blueprint reads actual source files before asking. It must not ask questions whose answers are already obvious from the code.
  2. No "should I continue?" gates: The agent always asks another round of questions. The user terminates Q&A by invoking blueprint-generate.
  3. Template-driven plan structure: Two built-in templates (Default, Concise) plus custom templates via templates.json. The template determines what sections the plan contains.
  4. Progress indicator: Every blueprint message after template selection shows a progress line at the end.
  5. Implementation BLOCKED during blueprint phase: Blueprint explicitly cannot generate the plan; blueprint-generate cannot ask questions.

Target audience

Developers who want structured planning before large features — greenfield projects, large new features, research experiments, new subsystems. Not recommended for visual frontend work, small refactors, or debug/polish tasks.

Distribution note

Listed on VS Code Marketplace as Imbue.imbue-blueprint — likely a wrapper or integration for the same underlying skills, not a separate product.

02

Architecture

Blueprint (Imbue) — Architecture

Distribution

  • agentskills.io compatible skills (SKILL.md per skill)
  • Install: npx skills add imbue-ai/blueprint
  • Or: curl -fsSL https://raw.githubusercontent.com/imbue-ai/blueprint/main/install-skills.sh | bash
    • Copies to ~/.claude/skills/ (or $CLAUDE_CONFIG_DIR/skills)

Required runtime

  • None beyond the AI host tool

Directory tree

imbue-ai/blueprint/
├── README.md
├── install-skills.sh        # curl install script
├── justfile                 # Just task runner
├── assets/
├── .cursor-plugin/          # Cursor integration
└── skills/
    ├── blueprint/
    │   ├── SKILL.md         # Main Q&A skill
    │   └── references/
    │       ├── templates.json        # Built-in plan templates
    │       ├── questions.md          # Question format spec
    │       └── refine-prompt.md      # Prompt refinement rules
    └── blueprint-generate/
        ├── SKILL.md         # Plan generation skill
        └── references/
            └── templates.json        # Mirrored templates (must stay in sync)

Templates (from references/templates.json)

[
  {
    "name": "Default",
    "description": "Full plan with all sections",
    "prompt": "Overview, Expected behavior, Implementation plan, Implementation phases, Testing strategy, Open questions"
  },
  {
    "name": "Concise",
    "description": "Minimal plan",
    "prompt": "Overview, Expected behavior, Changes"
  }
]

Custom templates added by editing templates.json in both blueprint/references/ and blueprint-generate/references/ (must stay in sync).

Output location

blueprint/<slug>/plan-<slug>.md

Target AI tools

Claude Code, Codex CLI, Gemini CLI, Pi agent, and any agentskills.io-compatible tool. install-skills.sh copies to ~/.claude/skills/ for Claude Code. Cursor plugin in .cursor-plugin/.

03

Components

Blueprint (Imbue) — Components

Skills (2)

Name Purpose
blueprint <description> Start a Q&A session: template selection → codebase exploration → question rounds
blueprint-generate End Q&A: write plan to blueprint/<slug>/plan-<slug>.md

References

File Purpose
skills/blueprint/references/templates.json Built-in plan templates (Default, Concise)
skills/blueprint/references/questions.md Question format rules (multiple choice, format spec)
skills/blueprint/references/refine-prompt.md Rules for the "refined prompt" shown after each Q&A round
skills/blueprint-generate/references/templates.json Mirror of templates.json (must stay in sync manually)

Templates (2 built-in + custom)

Template Sections
Default Overview, Expected behavior, Implementation plan, Implementation phases, Testing strategy, Open questions
Concise Overview, Expected behavior, Changes
Custom User-defined in templates.json with name, description, prompt fields

No other components

No slash commands, hooks, scripts, MCP servers, or subagents. Blueprint is deliberately minimal.

05

Prompts

Blueprint (Imbue) — Prompts

Excerpt 1: blueprint SKILL.md — Step 4 Question Guidelines

Source: skills/blueprint/SKILL.md

### Step 4: Ask clarifying questions

Based on the codebase exploration, ask 3-5 clarifying questions following the format in [references/questions.md](references/questions.md).

Guidelines:
- Keep text between questions brief — a sentence or two of context at most, not a full analysis
- Gather facts before asking — if something can be determined by reading code, searching docs, or looking up external references, find the answer yourself. Do not make subjective decisions on behalf of the user
- Look up external documentation, APIs, or tools when relevant
- Ground questions in what you found in the codebase — do NOT ask questions whose answers are already obvious from the code
- Questions must match the level and perspective of the selected template
- Users generally expect to continue existing patterns — only question existing patterns when the user's change clearly conflicts with them. Focus on what's new or ambiguous
- Do NOT ask questions about the plan template itself

Before the questions, add this hint:

> Answer with shorthand like `1a, 2b, 3e, 4a, 5b` or write freely.

After the questions, add:

> Once you're done answering, I'll follow up with more questions. When you're ready, invoke the blueprint-generate skill to end the Q&A and generate the plan.

Technique: Anti-redundancy constraint. The key rule — "do NOT ask questions whose answers are already obvious from the code" — forces genuine codebase exploration before question generation, preventing the lazy pattern of asking generic questions that any AI could generate without reading a single file.


Excerpt 2: blueprint SKILL.md — Refined Prompt Display

Source: skills/blueprint/SKILL.md

When they answer:
- If they answered a question with a follow-up question of their own, answer it before moving to the next round
- Acknowledge briefly
- Show the updated refined prompt — take the original feature description and add bullet points (using `*` syntax) incorporating all clarifications so far. Follow the rules in [references/refine-prompt.md](references/refine-prompt.md). Display it in a blockquote so the user can see how their answers are shaping the plan.
- ALWAYS ask 3-5 more questions.
- Keep asking rounds of follow-up questions until the user invokes the blueprint-generate skill

IMPORTANT: Do NOT stop asking questions on your own. Only the user decides when Q&A is done by invoking blueprint-generate. Do NOT generate the plan. Do NOT write or modify any code files.

Technique: Progressive refinement display. After each Q&A round, the agent shows an updated "refined prompt" blockquote — the original description plus accumulated clarifications as bullet points. This gives users a running summary of how their answers are shaping the eventual plan without requiring them to re-read the entire conversation.


Excerpt 3: blueprint SKILL.md — Template Customization

Source: README.md + SKILL.md

{
  "name": "feature",
  "description": "Plan for a user-facing feature",
  "prompt": [
    "The plan should contain the following sections in order.",
    "",
    "- Overview: motivation for the feature and what users will be able to do",
    "- User experience: walkthrough of the primary flow and key edge cases",
    "- Implementation: files, modules, and data types to add or change, and what each does",
    "- Testing: how to verify the feature, including unit tests, integration tests, and edge cases",
    "- Open questions: unresolved design decisions or trade-offs"
  ]
}

Technique: Template as prompt fragment. The prompt field (string or array) is injected directly into the agent's context to define the plan's structure and perspective. The Q&A questions and the final plan both adapt to this template, creating a coherent pipeline from question selection to plan output.

09

Uniqueness

Blueprint (Imbue) — Uniqueness

Differs from Seeds

Closest seed is spec-kit (clarification loop → spec output). Blueprint differs from spec-kit in three ways: (1) explicit two-skill separation — blueprint (gathering only, generation BLOCKED) and blueprint-generate (writing only, Q&A BLOCKED) vs spec-kit's single unified workflow; (2) user-controlled Q&A termination — the agent never stops asking questions; only the user terminates by invoking a second skill; (3) codebase-grounded question rules — Blueprint requires the agent to read actual source files before generating questions and explicitly prohibits questions answerable by code inspection. The "open questions" section in the Default plan template also distinguishes it — unresolved decisions are first-class plan output, not a byproduct. Compared to BMAD-Method, Blueprint is much narrower (planning only, no implementation workflow) but deeper in Q&A quality control.

Two-Blueprint Disambiguation (see BATCH-NOTES)

imbue-ai/blueprint (this entry) is a 2-skill planning copilot focused on multi-round Q&A before implementation. JuliusBrussee/blueprint (slug: blueprint-juliusbrussee) is a completely unrelated repo called "cavekit" — a compressed spec-driven development framework with caveman token encoding. Same GitHub repository name, completely different products. See batch notes for comparison.

Positioning

Blueprint occupies the "serious planning for serious features" niche — it is explicitly not recommended for small tasks and does not pretend to cover the full SDLC. Its value proposition is concentrated entirely on the Q&A-to-plan conversion, which it does with more quality controls (codebase grounding, anti-redundancy, user-controlled termination) than any other planning skill in this batch.

Observable Failure Modes

  1. Manual templates.json sync: Two copies of templates.json (blueprint/ and blueprint-generate/) must be kept in sync manually. Divergence causes inconsistent question focus and plan structure.
  2. No implementation: Blueprint produces a plan only. Users must separately invoke their implementation agent.
  3. VS Code extension vs skills: The VS Code marketplace extension (Imbue.imbue-blueprint) may have different behavior/features than the GitHub skills — the relationship is not documented in the README.
  4. Q&A can run indefinitely: The user decides when to stop; there is no guidance on "enough planning" for different task sizes.
04

Workflow

Blueprint (Imbue) — Workflow

Two-Skill Pipeline

blueprint <description>
    ↓
Step 1: Parse args + select template (one-time template choice)
    ↓
Step 2: Read template's prompt field
    ↓
Step 3: Explore codebase (read actual source files; spend real effort)
    ↓
Step 4: Ask 3-5 grounded clarifying questions
         + shorthand hint: "Answer with shorthand like `1a, 2b, 3e`..."
         + reminder: "When ready, invoke blueprint-generate to end Q&A"
         + progress indicator at end of message
    ↓ user answers (shorthand or prose)
Step 5: Acknowledge + show refined prompt (blockquote showing how answers shape plan)
         ALWAYS ask 3-5 more questions (no auto-termination)
         Repeat Step 5 indefinitely until user invokes blueprint-generate
    ↓
blueprint-generate
    ↓
Write plan to blueprint/<slug>/plan-<slug>.md

Approval Gates

  1. Template selection (user choice after blueprint invocation)
  2. Q&A termination (user invokes blueprint-generate — agent never terminates Q&A)

Rules for Questions (from SKILL.md)

  • Never ask questions answerable by reading code
  • Ground questions in actual codebase findings
  • Questions must match template's level and perspective
  • Never ask about the plan template itself
  • Only question existing patterns when the new change clearly conflicts
  • Keep text between questions brief (sentence or two)

Artifacts

Phase Artifact
blueprint Conversation with refined prompt blockquote in each message
blueprint-generate blueprint/<slug>/plan-<slug>.md

Progress Indicator

Every message after template selection ends with a single progress line (format defined in references/refine-prompt.md).

06

Memory Context

Blueprint (Imbue) — Memory & Context

State Storage

File-based. Output written to blueprint/<slug>/plan-<slug>.md.

Persistence

Project-level. Plan files survive session boundaries.

Cross-Session Handoff

Yes — plan files are the handoff artifact. The user can share the plan file with their coding agent in a new session.

Context Compaction

The "refined prompt" blockquote shown after each Q&A round serves as a running summary of the session, reducing the need to re-read the full conversation to understand accumulated decisions.

Memory Type

File-based markdown. No database, no vector search.

None.

Template sync risk

The templates.json file must be kept in sync between blueprint/references/ and blueprint-generate/references/. There is no automated sync — this is a manual maintenance burden noted in the README.

07

Orchestration

Blueprint (Imbue) — Orchestration

Multi-Agent

No. Single agent handles both blueprint and blueprint-generate skills. No subagents.

Orchestration Pattern

Sequential with explicit user-controlled phase boundary. The user invokes blueprint-generate to end the Q&A phase — the agent cannot self-terminate.

Isolation Mechanism

None.

Multi-Model

No. Single model.

Execution Mode

Interactive-loop. Blueprint is a conversation, not a background task.

Crash Recovery

Partial — plan files persist. The refined prompt blockquote provides a session summary. No explicit state machine.

Context Compaction

Implicit — the refined prompt blockquote compresses accumulated Q&A decisions into a single paragraph that can be read independently of the full conversation.

Cross-Session Handoff

Yes — blueprint/<slug>/plan-<slug>.md is the primary handoff artifact.

Prompt Chaining

Yes — blueprint conversation output → blueprint-generate reads conversation → writes plan.

Subagent Definition Format

None.

08

Ui Cli Surface

Blueprint (Imbue) — UI & CLI Surface

CLI Binary

None. Skills are invoked via the AI host tool's skill invocation mechanism.

Local UI

None. No dashboard.

IDE Integration

  • VS Code Marketplace: Imbue.imbue-blueprint (published extension)
  • Cursor: .cursor-plugin/ directory
  • Claude Code: via install-skills.sh → copies to ~/.claude/skills/
  • Any agentskills.io tool: npx skills add imbue-ai/blueprint

Observability

None beyond the plan file output. No audit log.

Install Options

Method Target
npx skills add imbue-ai/blueprint Any agentskills.io tool
curl -fsSL .../install-skills.sh | bash Claude Code specifically
VS Code Marketplace VS Code / Cursor

Related frameworks

same archetype · same primary tool · same memory type

Superpowers ★ 207k

Enforces spec-first, TDD, and subagent-reviewed development as mandatory automatic workflows rather than optional practices.

Anthropic Skills (Official) ★ 141k

Official Anthropic reference gallery for the Agent Skills standard, spanning creative, technical, and enterprise capability…

OpenAI Skills (Official) ★ 20k

Tiered library of integration and workflow skills for Codex, with a self-bootstrapping skill-installer that lets agents discover…

daymade/claude-code-skills ★ 1.1k

52 production-hardened skills with emphasis on skill craftsmanship — a fork of Anthropic's official skill-creator with security…

Superpowers Lab ★ 344

Experimental skill incubator extending Superpowers with interactive CLI automation, on-demand MCP discovery, semantic…

openspec-skills (chyiiiiiiiiiiii) ★ 6

Ports the OpenSpec proposal/apply/archive SDD lifecycle into three plain SKILL.md files requiring zero runtime dependencies.