Preview first and follow the app's AGENTS.md

[raghubetina]

Why

On 2026-10-12 the owner changed what the user's agent does after Compile and asked for the Skill's restatements to point to the generated app's AGENTS.md instead of copying it. The service side, which rewrites that AGENTS.md, is firstdraft/firstdraft#1149.

After Compile. The test run and continuation now follow the app's AGENTS.md: write any Behavior to write item that makes Home raise, preview without asking, and debug other failures without weakening a spec. The old text held the preview until every item was written and handed other failures to the user. A missing prerequisite still asks once, with the exact install command.

Steps that changed no outcome, removed:

• the opening "level" question; the agent makes implementation choices, names them in the read-back, and asks only on request;
• the read-back stop after "make the rest of the decisions for me", now a progress update followed by Compile;
• gaps-readback.md, the read-back's technical block, and the SHA rechecks (the CLI's local_plan_changed covers byte changes);
• the local JSON Schema validator procedure (plan push validates), the exact CLI version probe, and the readable-file check (local_input_unreadable covers it);
• the destroy-without-show read-back question, now a default (add show) named among delegated choices.

No email verification by default (owner, 2026-10-12): the Accounts checklist no longer asks about email confirmation; verification stays out unless the user asks, so sign-up signs people in at once and they can share the app with friends right away. Recovery and lockout stay authored, and the After-Compile email item says their emails are not delivered until a mail provider is added. Plan semantics are unchanged. The Drawing Board and kit notes had no email-confirmation wording to align.

The hand-off report shrinks to where the app is, what works, what the agent writes next, and how to preview. Two stale lines go: the multi-target comment gap example and the "Compiler without stubs" hedge. "An already approved Plan with no data is not permission to silently add records" goes too; after Compile, seeds are ordinary source.

Evals and tests. Level cases expect no level question; delegation compiles after a progress update; the failing-spec case expects debugging (its "preview anyway" follow-up is gone); the two local-validator cases and their fixture are removed. A generated AGENTS.md fixture is staged where cases read it, and the README fixture and interview protocol follow. The release-compatibility test no longer looks for the version probe, and RELEASING.md stops naming it.

Not changed on purpose: the ambiguous push/direct-start recovery (the CLI's errors.md says not to rerun; the CLI owns that and changes first), the public-exposure confirmation in examples (a data-exposure boundary), and running login for the user.

Release. The version moves to 0.8.16 (sync-plugin-version.mjs + npm version), because 0.8.15 is published and CI's release-order rehearsal requires its compatibility bytes to match the tag. The CHANGELOG lead names 0.8.16. plugin_source.tarball_sha256 is unchanged; the release records it. Locally on the rebased head: sh script/check --cli-root with pinned CLI 92f27714, check-cli-contract, check-plugin-release-order --prospective, and npm audit pass. Behavioral evals were not run.

Docs: updated skills/create-full-stack-app/SKILL.md, its interview, modeling-guide, diagnostics-and-recovery, and foundation-plan-023 references, CHANGELOG.md, RELEASING.md, evals/README.md, and the eval cases, fixtures, and interview protocol.
Review: Codex, approve (3 rounds, 1 finding fixed)