Start with the job
Describe the result you need. Skills support the work; selecting a persona is optional.
| Job | Reviewable result |
|---|---|
| Investigate a claim | A finding with evidence, inspected scope and uncertainty |
| Turn requests into scope | Current behavior, proposed change and the decision needed |
| Reconcile artifacts | A maintained source for each fact and preserved reviewer decisions |
| Explain a measurement | A reproducible definition and a qualified decision implication |
| Continue substantial work | Recovered decisions, current sources and the next authorized action |
Give sources their actual authority
Requests, accepted intent, inspected implementation, rollout and measured impact answer different questions. Code is optional when implementation is not the question.
Relevant context
Keep stable instructions short. Read the selected route and task handoff instead of loading all history.
Human decisions
A request is not a commitment. A linked repository is not authorization to change it.
Verified artifacts
Check the result the reader will use. A successful tool response is not proof of correct content or layout.
Preview before applying
From a framework source checkout, create a small local workspace with the starter selection.
node plugin/bin/workspace.js init ../my-product --runtime codex
node plugin/bin/workspace.js init ../my-product --runtime codex --apply
node plugin/bin/workspace.js doctor ../my-product
node plugin/bin/workspace.js start ../my-product --focus "Review a product request"Initialization and diagnostics do not call a model, start services, configure connectors or edit global settings. Add --apply to a launch preview only when you want to start the independently installed agent.
Choose the starter, product or evidence selection, or named skills. Required references and scripts travel with the reviewed bundle. Private sources and accounts stay in the target environment.
Review changes without losing local policy
node plugin/bin/workspace.js compare /path/to/workspaceThe read-only report separates local edits, upstream changes and conflicts, with installed/local/proposed text and review patches. It never applies updates. Reports can contain private local text and should remain private.
Diagnostics report candidate instructions and source configuration; they do not certify the runtime's actual discovery, authentication or current business facts.
Skill ownership and reviewed exports · Portable skill collection and behavioral cases
Follow a complete task
Independently fictional walkthroughs show an initial interpretation, correction, reviewable result and next-session handoff. They are examples, not measured agent results.