storebound/docs/agents-core.md

storebound · rules for AI agents

These rules apply to every AI tool that uses storebound (skills, MCP server or the CLI in a terminal).

First rule: do it, don't ask

storebound plus you, the AI agent, do the work. The owner is not your pair of hands.

1. You do everything yourself: API calls, uploads, builds, template and CSV downloads and exports, every console navigation (open, Get Started, Edit, Next, Continue, Save draft), console forms, and the verification afterwards. Never tell the owner "run this yourself" or "click X yourself" for any of these.
2. The owner keeps only the final public or irreversible buttons: Submit for review (App Review or Beta App Review), releasing a version or completing / resuming a production rollout, App Privacy Publish, accepting legal agreements; plus deleting store data, changing live prices, sending review replies and inviting or removing testers. (Also the owner's: canceling a submission, starting or stopping a store experiment, a console form's final Submit / Create that a console plan lists in buttons.final, signing in and granting account permissions.) For these you prepare everything, show the dry run, and the owner presses one button or says yes.
3. Ask the owner only for facts that no code, config or doc can prove: compliance answers such as Data safety, App Privacy, age rating, export compliance, partial data deletion. Collect ALL of them in ONE batch at the start (storebound check and storebound validate list them), never one at a time during the run. Reuse answers that are already in the config.
4. Default loop: plan, then you drive (the API first, your browser tool for consoles without an API), then verify (read it back), then continue. No round trip to the owner in between.
5. Catch rejections offline. Before a write, storebound check and --dry must catch what the store would refuse (live-shape fixtures, id and pair checks). A failure belongs in the dry run, not in a 400 at write time; when a store refuses something the dry run passed, fix the check (pitfalls), not just the data.

Every hour of "please click this" is an hour the owner did not want to spend. If you catch yourself writing "you need to ...", do it instead, unless it is one of the owner steps in point 2.

The other rules

Read skills/store-pitfalls/SKILL.md before any store task. It lists the mistakes already made in real releases (store lists that lag after a create, dry runs without a plan, taken App Store names, byte limits, ...) and what to do.

1. The owner presses only the owner steps (point 2 of the first rule). Prepare them completely, say what will happen, and let the owner press the button, or run the command only after a clear yes in chat. storebound's MCP server does not offer these at all. Everything else you run yourself.
2. Dry run first, then run it yourself. Every write command has --dry (MCP: dry_run, default true). Run the dry run, check and show it (the rows that change, not just a count), then run the real command yourself (--yes where it asks for a confirmation) and read the result back. Wait for the owner's yes only before an owner step.
3. Read before you write. storebound scan (offline: rejection risks in the project itself), storebound check (offline) and storebound validate (read-only, both stores) tell you what is missing and in which order to fix it.
4. Credentials only from the environment, .storebound/ in the project (gitignored) or a credential profile (~/.storebound/credentials/<profile>.json, named by --account <profile> or a workspace). Never ask for keys, passwords or tokens in chat, never print them, never write them into the config, docs or commits. The same goes for the App Review demo account (reviewDemoAccount in .storebound/credentials.json): the owner puts it there.
5. Never answer a compliance question you cannot know. Export compliance, Data safety, App Privacy, age rating, target audience: ask the owner when the code does not prove the answer, all such questions in one batch at the start (first rule, point 3), and write the answers into the config. A wrong "no" is worse than a slow "ask".
6. Pick builds by explicit number. Use the build number / versionCode you just uploaded, never "the latest".
7. Long uploads run in the background with a log (storebound play images > .storebound/images.log 2>&1 &) and you report the progress lines, not guesses.
8. No manipulative tactics with reviewers, players or store policies: no fake reviews, no incentives for ratings, no hidden features, no misleading screenshots, no keyword stuffing.
9. Personal data stays local. Tester lists, review reply drafts, health reports and analytics snapshots live in .storebound/ (gitignored). Never put tester emails, reviewer names or crash logs into issues, docs or commits, and never ask for tester emails in chat (storebound masks them in its output).
10. Console forms without an API: drive them yourself, agent-first. storebound console plan <form> --json (MCP console_plan) gives the start URL, steps, the answer sheet, the buttons you must never press and the verified UI facts. Drive the page with your own browser tool, adapt to what it shows, run long flows as an in-page async loop that writes window.__log (one evaluate call times out around 45 s), stop only before the final button (Publish, Submit, Create, a final Save) and tell the owner exactly what to press, then storebound console verify <form>. Never sign in, accept an agreement or grant a permission (a service account's app access) for the owner. storebound console fill is the optional scripted fast path; it never blocks you.