AI Rules & Prompts

Fakebase is built to be driven by AI coding agents. The @byronwade/ai package and the fakebase ai CLI generate machine-readable rules, schema summaries, and task prompts so that agents understand the project's schema and the hard constraints of a local/dev-only backend.

Generate the files

fakebase ai init

This scaffolds:

File	Purpose
`fakebase.rules.md`	Primary rules file: project overview, schema tables, critical warnings, migration workflow, and the list of unsupported capabilities
`schema.summary.md`	Human-readable description of tables, columns, relationships, enums, and functions
`policies.summary.md`	RLS policies, the role access matrix, and limitations vs real Postgres RLS
`compatibility.summary.md`	Capability table (`SUPPORTED` / `PARTIAL` / `STUB` / `UNSUPPORTED`)
`migration.checklist.md`	Checklist to complete before switching to real Supabase
`.cursor/rules/fakebase.mdc`	Cursor-native rules
`AGENTS.md`	Generic agent guidance

All generated docs are regenerated from the current schema IR — never hand-edit them; re-run fakebase ai init instead.

Task prompts

fakebase ai prompt --target cursor          # cursor | claude | copilot | generic
fakebase ai prompt --target generic --output prompt.md

ai prompt emits a prompt tailored to the target tool, pre-loaded with the schema summary and the Fakebase guardrails so an agent starts with correct context.

Guardrails baked into the rules

The generated rules encode hard constraints so agents don't drift into unsafe assumptions:

Local/dev-only. Fakebase is never production auth or RLS; sessions, tokens, and users are ephemeral.
RLS is approximate. Policies are evaluated in JavaScript, not SQL. Always verify against real Supabase before deploying.
Never invent unsupported APIs. If a capability returns CapabilityError.notImplemented(), surface the error and suggest a compatible alternative — don't work around it.
Always export SQL (fakebase migrate export) before claiming a backend feature is complete.
Always run fakebase verify supabase before production handoff.
Use cookie storage + PKCE-shaped flows for Next.js SSR.

Why this exists

Supabase itself ships AI prompts, agent skills, and MCP tooling because workflow guidance measurably improves agent output. Fakebase follows the same lesson in a smaller, sharper form: the generated context keeps agents productive while preventing them from treating local approximations as production guarantees.