Journey Plan Skill for Claude

Journey: Plan

Stage 2 of the guided journey. Turn the idea brief into an actionable Butterbase plan.

When to use

• Dispatched by journey when current_stage: plan.
• Directly via /butterbase-skills:plan.

Inputs

• docs/butterbase/01-idea.md (must exist — if absent, bounce back to journey-idea).
• docs/butterbase/00-state.md (for hackathon_mode, hackathon_deadline).

External services — check built-ins first

When the plan needs email, messaging, calendar, CRM, docs, or project-management integration:

• Invoke butterbase-skills:integrations and call manage_integrations action: "list_available" to see what Composio covers for this app.
• If a toolkit fits, the plan should record "via manage_integrations (toolkit: <name>)" instead of naming an external SDK.

When the plan needs payments:

• Invoke butterbase-skills:payments.
• Default to Stripe Connect via manage_billing unless the user's region forces a fallback (see the payments skill).
• Record the choice ("Stripe Connect" or "<regional gateway> via function proxy") in the plan's Payments section.

Procedure

Work through these sections in order. After each section, write the result to 02-plan.md before moving on. One question at a time per the spec's questioning discipline.

• Tables. Read the capability map. Propose a starter table list with columns and types — recommend, don't ask blank. Example: "Tables I'm seeing: users, orders, items. Missing any?" Then for each table: "<table>.<column>: should this be a uuid / text / int / timestamp / enum?". Confirm primary keys, foreign keys, indexes that are obvious (foreign-key columns).

• RLS model. For each table: "Can user A see user B's <table> rows? ① no, strict isolation ② yes, public-read ③ only shared via explicit grant." Decide policy shape. In hackathon mode, prefer option ① and recommend manage_rls action: create_user_isolation.

• Auth. "OAuth providers: ① Google only ② Google + GitHub ③ email/password too ④ none (anonymous app)." Also ask: "Need a demo / judge account seeded? (hackathon mode only)".

• Functions. For each function from the capability map: "<name>: trigger = HTTP / cron / WebSocket? If cron: schedule? If HTTP: idempotency needed?".

• Storage. If used: "Which objects (avatars, attachments, …)? Public-read or per-user?".

• AI / RAG / realtime / durable. Only if used. Capture model choice (AI), collections (RAG), tables to subscribe to (realtime), object kinds (durable).

6b. Agents. Only if create_agent is tagged in the idea. For each agent, capture:

• name (slug), one-line purpose.
• Tool surface: which builtins (query_table, insert_row, …), which functions (must exist in the Functions section), which MCP servers (URL + transport).
• Visibility: private | authenticated | public. If public AND any write tool is reachable, mark safety_acknowledged_needed: true and set per-IP / per-user / per-app rate limits + daily_budget_usd.
• Default model.
• Note that the agent record will not be carried by clone replay — the spec JSON (under agents/<name>.json) rides along in the repo snapshot, so this matters for the publish stage too.

• Frontend stack. "Frontend: ① Vite + React ② Next.js ③ static HTML ④ none (API-only)." Write to 00-state.md frontend_stack.

7b. Publish-as-template. Read publish_as_template from 00-state.md front-matter (set by journey-idea). If yes or unlisted, plan for it now:

• README outline (one-liner, env-var-per-function list, OAuth setup, agent re-import, MCP server registration, seed data, first-run smoke).
• Which env vars use the auto-mint convention (butterbase_api_key) vs. require manual user input on clone.
• Whether agents/*.json files need to be exported and committed (yes if Agents section is non-empty).
• Note that publishing requires butterbase repo push to upload the source tree as a snapshot — without it, cloners get an empty file tree.

Toolchain question

Ask: "Will your app's code use @butterbase/sdk only in the frontend, only server-side (functions, scripts), or both?" Record under Toolchain → SDK surfaces.

Ask: "Want to install @butterbase/cli for the local dev loop (logs, scaffolding, key rotation)? (yes/no — default yes)" Record under Toolchain → CLI usage.

• Scope cut (hackathon mode). Re-read the must-haves list. For each, ask: "Ship now or post-hackathon?" Write the cut list into 02-plan.md's "Post-hackathon" section.

• Annotate skipped stages. For every build stage NOT used in this plan (check the capability map and feature list), update 00-state.md's checklist to read - [ ] <stage> (n/a) for that row. Also do this for rls if hackathon_mode: true (mark as (folded into schema)).

10. Final approval. Show the user the assembled plan and ask: "Plan looks good? (yes / revise <section>)". Loop until yes.

02-plan.md format

# Plan

## Tables
- `users` (id uuid pk, email text unique, created_at timestamp)
- `orders` (id uuid pk, user_id uuid fk→users.id, status enum[pending,paid,shipped], total int, created_at timestamp; index on user_id)
- ...

## RLS
- `orders`: user-isolation (create_user_isolation, owner column = user_id)
- ...

## Auth
- Providers: Google
- Demo user: yes (email [email protected], password set via seed)

## Functions
- `stripe-webhook` — HTTP, idempotency table `_processed_events`
- `daily-digest` — cron 0 9 * * *  UTC

## Storage
- bucket: `avatars` (per-user, private; download via presigned URL)

## AI / RAG / realtime / durable
- (none)

## Agents
- `order-summariser` — purpose: summarise a user's recent orders on demand.
  - Tools: builtin `query_table`, function `format-currency`.
  - Visibility: authenticated. Rate: 60/hr per user. Daily budget: $5.
  - Default model: claude-haiku-4-5-20251001.
  - Spec file: `agents/order-summariser.json` (committed to repo).
- (omit section entirely if no agents)

## Publish-as-template
- Intent: yes / unlisted / no
- README outline: <bullet list of sections>
- Env vars cloners must supply: <list per function>
- Auto-mint eligible keys: `butterbase_api_key` (etc.)
- Agent specs to bundle: `agents/*.json`
- Snapshot push: `butterbase repo push` at end of journey-templates.
- (omit section entirely if publish_as_template = no)

## Frontend
- Vite + React

## Toolchain

- **SDK surfaces**: <client-side only | server-side only | both>
  - Client-side: install `@butterbase/sdk` in the frontend; use `auth`, `db`, `storage`, `realtime`.
  - Server-side: install `@butterbase/sdk` in functions / scripts; use the service-key flow for elevated access.
- **CLI usage**: <yes / no>
  - Yes (default): use `butterbase` CLI for local scaffolding, log tailing (`butterbase logs <fn>`), function invocation, and key rotation.
- **Why both**: MCP tools provision and orchestrate; SDK + CLI are the runtime + dev loop.

## Build order
1. schema
2. rls (folded into schema in hackathon mode)
3. auth
4. storage
5. functions
6. ai          (if used)
7. rag         (if used)
8. realtime    (if used)
9. durable     (if used)
10. agents     (if used — must come after functions, ai, and any MCP-server setup)
11. frontend
12. deploy
13. templates (optional — only if publish_as_template != no)

## Post-hackathon
- email notifications (deferred)
- admin dashboard (deferred)

Outputs

• Writes docs/butterbase/02-plan.md.
• Updates 00-state.md: tick - [x] plan, set frontend_stack, set current_stage: preflight, annotate skipped build stages with (n/a).

Anti-patterns

• ❌ Asking the user to design every table column without recommending defaults first.
• ❌ Skipping the scope-cut section in hackathon mode.
• ❌ Writing the plan only at the end — write as you go so progress survives a crash.