90 lines
4.0 KiB
Markdown
90 lines
4.0 KiB
Markdown
---
|
|
name: blueprint
|
|
description: "Turn a one-line objective into a step-by-step construction plan any coding agent can execute cold. Each step has a self-contained context brief — a fresh agent in a new session can pick up any step without reading prior steps."
|
|
category: planning
|
|
risk: critical
|
|
source: community
|
|
date_added: "2026-03-10"
|
|
---
|
|
|
|
# Blueprint — Construction Plan Generator
|
|
|
|
Turn a one-line objective into a step-by-step plan any coding agent can execute cold.
|
|
|
|
## Overview
|
|
|
|
Blueprint is for multi-session, multi-agent engineering projects where each step must be independently executable by a fresh agent that has never seen the conversation history. Install it once, invoke it with `/blueprint <project> <objective>`.
|
|
|
|
## When to Use This Skill
|
|
|
|
- Use when the task requires multiple PRs or sessions
|
|
- Use when multiple agents or team members need to share execution
|
|
- Use when you want adversarial review of the plan before execution
|
|
- Use when parallel step detection and dependency graphs matter
|
|
|
|
## How It Works
|
|
|
|
1. **Research** — Scans the codebase, reads project memory, runs pre-flight checks
|
|
2. **Design** — Breaks the objective into one-PR-sized steps, identifies parallelism, assigns model tiers
|
|
3. **Draft** — Generates the plan from a structured template with branch workflow rules, CI policy, and rollback strategies inline
|
|
4. **Review** — Delegates adversarial review to a strongest-model sub-agent (falls back to default model if unavailable)
|
|
5. **Register** — Saves the plan and updates project memory
|
|
|
|
## Examples
|
|
|
|
### Example 1: Database migration
|
|
```
|
|
/blueprint myapp "migrate database to PostgreSQL"
|
|
```
|
|
|
|
### Example 2: Plugin extraction
|
|
```
|
|
/blueprint antbot "extract providers into plugins"
|
|
```
|
|
|
|
## Best Practices
|
|
|
|
- ✅ Use for tasks requiring 3+ PRs or multiple sessions
|
|
- ✅ Let Blueprint auto-detect git/gh availability — it degrades gracefully
|
|
- ❌ Don't invoke for tasks completable in a single PR
|
|
- ❌ Don't invoke when the user says "just do it"
|
|
|
|
## Key Differentiators
|
|
|
|
- **Cold-start execution**: Every step has a self-contained context brief
|
|
- **Adversarial review gate**: Strongest-model review before execution
|
|
- **Markdown-first distribution**: The reviewed revision is primarily instructions and templates, but installing or following it can still cause an agent to run commands. Treat the repository as untrusted until inspected.
|
|
- **Plan mutation protocol**: Steps can be split, inserted, skipped with audit trail
|
|
|
|
## Installation
|
|
|
|
Do not clone a moving branch directly into an active skills directory. First ask
|
|
the user to approve network access to the named repository. Then inspect the
|
|
reviewed revision and ask separately before activating it:
|
|
|
|
```bash
|
|
review_dir="$(mktemp -d)"
|
|
git clone --filter=blob:none https://github.com/antbotlab/blueprint.git "$review_dir/blueprint"
|
|
git -C "$review_dir/blueprint" checkout --detach 07c5b305cf2d95d584a0d0398c390e839fec5954
|
|
git -C "$review_dir/blueprint" ls-files
|
|
git -C "$review_dir/blueprint" status --short
|
|
```
|
|
|
|
Read `SKILL.md` and every bundled file at that exact commit. Check for scripts,
|
|
hooks, symlinks, network calls, credential access, and instructions that request
|
|
commands or elevated permissions. Only after explicit user approval, copy the
|
|
reviewed files into the selected host's skills directory. Re-review a newer
|
|
revision instead of silently updating this pin.
|
|
|
|
## Additional Resources
|
|
|
|
- [GitHub Repository](https://github.com/antbotlab/blueprint)
|
|
- [Examples: small plan](https://github.com/antbotlab/blueprint/blob/main/examples/small-plan.md)
|
|
- [Examples: large plan](https://github.com/antbotlab/blueprint/blob/main/examples/large-plan.md)
|
|
|
|
## Limitations
|
|
- Use this skill only when the task clearly matches the scope described above.
|
|
- Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
|
|
- Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.
|
|
- A pinned revision is reproducible, not automatically trustworthy; its contents still require review.
|