Loop Setup

Session 0 — set up the loop

You are setting up a working method in someone else's project. Touch nothing until they have seen what you propose to write. Never invent a fact about their stack: check it.

1 · Read the project

Detect, don't ask, what you can find yourself: ls, git remote -v, git branch --show-current, cat .gitignore, and whichever of these exist — package.json scripts, Makefile, pyproject.toml, Cargo.toml, go.mod, docker-compose.yml, CI config, test config, migrations directory. Report in three lines what this project is and how it is built.

2 · Check the connections — live, not by asking

The method depends on being able to observe things. Find out what is actually available in this session, right now, and print a table of PASS / FAIL / ABSENT:

| Need | How to check | |---|---| | Version control | git status works, and whether a remote exists | | Build / tests | run the project's build or test command once | | Browser | is a browser-driving tool available (in-app browser, Chrome extension, Playwright/Puppeteer in devDependencies)? If yes, open one page and confirm it renders | | Database | is there a DB MCP server, a CLI (psql, mysql, sqlite3), or an ORM with a dev database? If yes, run one harmless read query | | Payments sandbox | if the project charges money: is a test-mode key present in the environment (never print its value) | | Deploy | how does this project publish, and is that command something the owner runs, not the agent |

For anything ABSENT, say in one line what will not be verifiable without it and what the cheapest substitute is. Do not silently accept a gap — an absent connection becomes a declared GAP in .loop/VERIFY.md, not a shrug.

3 · Decide which models run which role

Ask, do not guess: which is the strongest model you have access to, and do you have one you would rather keep for the expensive steps? Plans differ, and a managed install can restrict models outright.

Explain the shape in one breath, then write the answer into .loop/roles/*.json:

• plan and close want the strongest reasoning at the highest effort. These are

the two steps where being wrong is most expensive: a bad plan wastes every session after it, and a soft audit lets a broken slice through.

• build and verify can be a mid-tier model. The hard calls are already made.
• Use aliases (opus, sonnet, haiku, or whatever the strongest tier is called

on their plan), never full model ids — aliases survive model releases.

If they only have one model, say so plainly and write that one everywhere: the method still works, it just loses the quality gradient. Never leave a profile pointing at a model they do not have — an unavailable model can fall back silently, and a silent fallback is the exact failure this method exists to prevent. Then tell them that no session can report which model it is on, so the profile is the only record.

Ask how they work, because it changes what the profiles are worth:

• In a terminal — the loop wrapper launches each role with its model already set.

The profiles do the work; they never choose a model by hand again.

• In the desktop app or an IDE — sessions are launched by the app, so the wrapper is

not the path. They invoke the skills directly and pick the model in the app's model selector. Say this plainly instead of letting them think the profile applied. It costs little: the plan, the council and the audit dispatch subagents whose model is fixed in their own definitions, so the expensive thinking is strong either way.

4 · Ask only what you cannot find out (these are the owner's calls)

1. Are the loop files versioned or local-only? .loop/ holds the plan, the

decisions and the hard-won techniques. Committing them means the method survives a dead laptop and travels to teammates; keeping them local means nothing new enters the history. Recommend committing, and either way write .loop/ACCESS.local.md for anything sensitive and always gitignore that file.

1. Which branch is protected (never pushed to by an agent) and **who authorizes

publishing**.

1. Which commands are forbidden to the agent in this project — the destructive

or expensive ones (a deploy, a production migration, a data wipe).

1. Is there a shared resource two parallel sessions could collide on (one

database, a schema version counter, a staging environment), and the exact command that reads its current true state.

1. Solo or team? Solo: an item closes with a commit on a branch. Team: it closes

with a pull request, green CI, and a reviewer.

5 · Write the files

Copy from this skill's templates/ directory into the project, filling in what you learned. Show the list first and get a yes.

• .loop/STATE.md — the single source of truth for the plan
• .loop/VERIFY.md — the verification contract, filled in with their real commands
• .loop/DECISIONS.md — append-only ledger
• .loop/GOTCHAS.md — techniques learned the hard way, starts empty
• .loop/ACCESS.local.md — how to reach test accounts and environments. **Copy it

from templates/ACCESS.md** (the template is not named .local so that a *.local.md ignore rule cannot swallow it) and always gitignore the installed copy

• .loop/roles/{plan,build,verify,close}.json — one launch profile per session role
• loop at the repo root — the wrapper, chmod +x
• .claude/agents/loop-{planner,builder,verifier,auditor}.md — copy the role agents

here so the project controls them (agents shipped inside a plugin cannot carry permission rules)

• Append .gitignore entries according to their answer in step 4.1 — always

.loop/ACCESS.local.md.

Then append six lines to their CLAUDE.md (create it if absent), so that any future session lands in the loop without anyone saying the word "loop":

## How we work
Before building anything, read `.loop/STATE.md` and follow the `loop` skill.
`.loop/STATE.md` is the only source of truth for what is done and what is left —
nothing lives only in someone's head. Nothing is "done" until it has been observed
working, per `.loop/VERIFY.md`. Never push to the protected branch.

6 · Close session 0

Print: the connection table, the files written, the decisions recorded, and the exact next command for the way they actually work — ./loop plan "" in a terminal, or "switch to your strongest model, then run the loop-plan skill with your goal" in the app. Then stop. Session 0 does not build anything.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: cbdreamer11
• Source: cbdreamer11/CB-loop-kit-claude-plugin
• License: MIT

Install and usage instructions live in the source repository linked above.