Skip to main content

blueprint-setup

Prepares a project for agentic development once its Blueprint documents are written. Use when the plan is closed and the repository has to be made ready, or when the user asks to prepare the environment, the agent entry file or the project journal. Writes the Blueprint block of the entry file, connects memory and observability through a tool or the project journal, adds a playbook per role when a team develops, runs the project check, and says what it could not connect.

Source facts

Repository
sinteselabs/Blueprint
Last source activity
October 11, 2026 at 13:54
Detected SKILL.md language
English
Stars
0
Forks
0

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

File Explorer
8 files

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
name
blueprint-setup
description
Prepares a project for agentic development once its Blueprint documents are written. Use when the plan is closed and the repository has to be made ready, or when the user asks to prepare the environment, the agent entry file or the project journal. Writes the Blueprint block of the entry file, connects memory and observability through a tool or the project journal, adds a playbook per role when a team develops, runs the project check, and says what it could not connect.
# Blueprint: setup The documents are written and the plan is closed. What is missing is the repository itself: an agent that opens this project has to find the rules, the plan and where to record, without anyone explaining it again. This skill writes inside the project folder only, and it never installs a tool on the person's machine. ## How to talk - Answer in the person's language, in plain words. - Say every file you created and every file you changed, one line each. - What did not work is said, never hidden: a tool that does not answer, a file that could not be written. ## 1. Before you write 1. **Find the project.** `.blueprint/state.md`, in this folder or in the folders above. If there is none, say that you found no Blueprint project here and stop: there is nothing to prepare yet. 2. **Read the state whole.** From it: `documents_at`, `tasks_at`, `document_language`, `developers`, `memory`, `observability`, and the plan's situation. 3. **If the plan is not closed,** say so and what is missing. Preparing a project whose plan is not written leaves an entry file pointing at a plan that does not exist. Ask whether to go on all the same. 4. **Look at what the project already has:** `AGENTS.md`, `CLAUDE.md`, a journal folder, a playbooks folder. What exists is kept. ## 2. The entry file `AGENTS.md` at the project root is the person's file. Blueprint keeps one block in it, between two marks, and touches nothing else: ``` <!-- blueprint:start --> ... <!-- blueprint:end --> ``` 1. **If the block is there,** replace what is between the marks, and keep every line outside them exactly as it is. 2. **If the file exists without the block,** add the block at the end, after one empty line. Read the file first: when a rule of the person says the same thing as a line of the block, keep theirs and drop that line, and say which. 3. **If the file does not exist,** create it with the block alone. 4. **Write the block from [assets/agents-block.md](assets/agents-block.md),** in the project's document language, with the project's real paths, taken from the state. Replace every `{{placeholder}}` and leave no section that does not apply. 5. **`CLAUDE.md` and `GEMINI.md`:** some agents read a file of their own and never open `AGENTS.md`. Claude Code reads `CLAUDE.md` first, and the Antigravity CLI reads `GEMINI.md` and does not read `AGENTS.md` at all (read in its version 0.63.0, 11/10/2026). So both files get one line: `See @AGENTS.md for how to work on this project.`, in the document language. One that does not exist is created with that line alone; one that exists and does not name `AGENTS.md` gets the line at the end, and you say so. Never move the person's rules out of either. The block is also what says that this project was prepared. Nothing in the state file records it: the mark in the entry file is the fact. ## 3. Memory and observability The state's `memory` and `observability` say `journal` or name a tool. **With the journal,** which is the way that needs nothing installed: make sure `.blueprint/journal/` exists, with `milestones.md` and the session records. Create what is missing through the `blueprint-record` skill. Say in one line that the project records itself in files, and that there is no automatic measure of cost this way. **With a tool named,** check that it answers before you say it is connected: read something from it, or write a test line and read it back. Then: - it answers: say so, and what it records; - it does not answer, or you have no way to reach it: say that, say what the person has to do once (install it, put a key where that tool reads it), and write to the journal until it is back. Never report a tool as connected because the state names it. - a file the tool needs inside the project: write it from the examples in [assets/project-files/](assets/project-files/), with the values the person gave. A key, a token or a password is never written in a file of the project. If the person says they want a tool that the state does not name, say what it takes, and change the state only through the `blueprint-record` skill. **When the state says `journal` and a tool would be worth it,** offer it once, in one line, with the reason it fits this project, and go on either way. Most people have never heard of the tools in this field and cannot answer a bare "do you want observability?": the offer carries the reason or it is noise. [references/observability.md](references/observability.md) has when it is worth it, when it is not, what to ask first, and what the person does to wire it. Blueprint ships no script for this and installs nothing: a no, or no answer, is a closed subject, and the journal keeps recording. ## 4. A playbook per role Only when the state's `developers` is more than 1. One file per role, in a `playbooks/` folder inside the documents folder, from [assets/playbook.md](assets/playbook.md). The roles come from the plan: who does what. If the plan does not say, ask the person, once, naming the roles you would create. With `developers` at 1, create nothing: a single person has no handover. ## 5. Run the check Go through the project check with the `blueprint-status` skill, and report what it found. A problem is reported with how to fix it; fix only what belongs to this skill: the block of the entry file, the journal, a playbook. ## What you give back In this order, about ten lines: 1. **What the project has now,** in one line: prepared, and where its rules are. 2. **What you created and what you changed,** one line each. 3. **Memory and observability:** one line each, saying what records what, or what is missing and what the person has to do once. 4. **What the check found,** one line, or that it found nothing. 5. **The next thing:** the first task of the plan that is not done, by its number and its spec file, and that the person can say to start it. ## Never - Never touch anything outside the marks of the entry file. - Never delete or rewrite a rule the person wrote. - Never say that a tool is connected without having read or written something through it. - Never write a key, a token or a password in a file of the project. - Never install anything on the person's machine, and never run the project's code. - Never write the first task yourself. Preparing the repository ends where the development begins.
View on GitHub