Docs — How To Use

How To Use

Agent Manifesto 3.0 ships as a plugin for the Claude Code and Codex marketplaces. Install it, then run it in ordinary language — the ai-landscape skill routes the request to one of five workflows. You can also point your tool at a workflow file directly.

Install From a Local Checkout

Both CLIs accept a clone of the repository as a marketplace, so you can install and inspect the plugin without publishing anything. Run these from the repository root. For Claude Code the trailing slash matters — a bare . is rejected.

Claude Code

claude plugin marketplace add ./claude plugin install agent-manifesto@agent-manifesto --scope local -y

Codex

codex plugin marketplace add .codex plugin add agent-manifesto@agent-manifestocodex plugin list

To install the published plugin instead, swap the first line for claude plugin marketplace add AlexeyPlatkovsky/agent-manifesto or codex plugin marketplace add AlexeyPlatkovsky/agent-manifesto. In Codex you can browse the result with /plugins; start a new session before using it.

The Five Workflows

Every workflow inspects the repository first, proposes a scoped change, and waits for your approval before it edits anything. Moves, renames, and deletions always need explicit consent. Paths below are relative to plugins/agent-manifesto/; from a clone, run @plugins/agent-manifesto/workflows/setup.yml.

setup

landscape-setup

Ask: “set up an AI landscape here”
Run: workflows/setup.yml
Why: Builds the smallest landscape that gives every concern one owner, so an agent starts with your facts, conventions, and authority limits instead of inferring them. It inspects the repository before it asks anything, and asks only questions whose answers change the result.
When: A project has no AI instructions yet, or the ones it has grew by accretion and need structural adjustment rather than another file.

review

landscape-review

Ask: “review this project’s AI instructions”
Run: workflows/review.yml
Why: Judges the active landscape against the smallest viable harness and separates durable context from scaffolding a capable model no longer needs. A clean result is a valid result: the workflow is allowed to recommend no change.
When: Your model or tool changed materially, a large instruction change just landed, or six months have passed since the last review.

extend

landscape-extension

Ask: “add a migration-review skill”
Run: workflows/extend.yml
Why: Adds one bounded capability — a context source, skill, workflow, agent, contract, or enforcement rule — and places it with the concern’s existing owner rather than beside it. It justifies fresh context for any new agent and real repetition for any new workflow.
When: A recurring task has earned a policy of its own, or repeated ordered work has earned a procedure. Not for a one-off plan.

adopt-tool

tool-adoption

Ask: “adopt this plugin into the project”
Run: workflows/adopt-tool.yml
Why: Sorts an external tool’s surface into what the project keeps and what it discards: runtime dependency, native adapter, reusable policy, procedure, contract, or demo noise. Adapters keep their source trigger, authority, tool limits, and output semantics.
When: You are pulling in a plugin, library, or instruction bundle that ships its own opinions about how an agent should work.

retire

landscape-retirement

Ask: “retire this AI landscape”
Run: workflows/retire.yml
Why: Moves surviving facts, policies, and authority boundaries to a destination you choose before anything is deleted, so a retirement that stops partway loses nothing. It also states what stops being enforced and becomes yours again.
When: A model upgrade made most of the harness unnecessary, or a different framework is taking over. Removal is a supported outcome, not a failure.

Validating a Landscape

The validator uses only the standard library. It checks workflow shape, skill and agent frontmatter, cross-layer references, and dependency cycles. Setup installs a copy into the landscape it creates, so the check travels with the harness.

validate.py

python3 plugins/agent-manifesto/scripts/validate.py # the framework itselfpython3 plugins/agent-manifesto/scripts/validate.py path/to/repo # a generated landscapepython3 .claude/scripts/validate.py . --snapshot # fingerprint a dirty tree firstpython3 .claude/scripts/validate.py . --proposal approved-change.json # check the diff against approval

With --proposal, the validator reports out-of-scope changes, drifted baseline files, mismatched authority-sensitive content hashes, and private paths staged for commit. That is what turns an approval into a check rather than a promise.