PrepFor.Me

A war room for high-intent job applications. PrepFor.Me tailors your resume truthfully to each role, builds company-specific interview prep that compounds with every recap you log, and keeps you in charge — nothing is ever auto-submitted.

This repo is meant to be open source. A hosted instance (when it exists) can charge for the cloud model, storage, and ops. Self-hosting stays yours: run the same app locally against a local LLM (Ollama) or against your own API keys (Anthropic + OpenAI). Nothing in the product requires the hosted billing path.

Live: https://prep-for-me.vercel.app/ — pre-v1, a working CRUD skeleton.

Documentation

Doc	What's in it
PROJECT.md	The source of truth. What the product is, the moat, scope and non-scope, decisions, current state, open questions, version history.
TECHNICAL.md	Stack, hosting, environment, code layout, data model, migrations, deploy, and the gotchas that have already cost time.
AGENTS.md	Conventions and invariants for anyone — human or agent — changing this repo.

Stack

React 18 + TypeScript, built with Vite. Supabase for Postgres, magic-link and Google auth, and later Edge Functions. TanStack Query for server state, React Router for navigation. No UI framework: inline style strings are parsed into React style objects by src/css.ts, so oklch colors and gradients stay exactly as designed.

Getting started

You need Docker (Docker Desktop, or Colima on a Mac) and the Supabase CLI. npm run dev alone is the UI — Postgres, auth, and AI live in other processes.

npm install
cp .env.example .env.development.local
# After `supabase start`, copy API URL + publishable key from `supabase status`
# into .env.development.local. Set VITE_AI_PROVIDER=edge to talk to functions.

supabase start                 # Docker stack: API :54321, Studio :54323, inbox :54324
npm run dev                    # http://localhost:5173  — use localhost, not 127.0.0.1

The app boots without credentials — the landing page renders and the sign-in screen tells you what's missing — so a blank env file never crashes anything.

supabase start does not serve Edge Functions. Those are a second process; see Running AI locally.

What has to be running

Process	Command	Needed for
Docker VM	colima start / Docker Desktop	local Supabase
Supabase stack	supabase start	auth, DB, storage
Vite	npm run dev	the web app
Functions	supabase functions serve --env-file …	any real AI (analyze, tailor, prep chat)
Model backend	Ollama + bridge, or Anthropic via supabase/.env.local	what the functions call

First-time Supabase setup

1. Create a project at supabase.com.
2. Run every file in supabase/migrations/ in filename order, via the SQL editor or supabase db push. 0003_grants.sql is not optional — without it the tables exist and every query still fails with 42501 (why).
3. Copy the project URL from Settings → API and the sb_publishable_… key from Settings → API Keys into .env.development.local. Both are safe in the browser; row level security is what protects the data.
4. In Authentication → URL Configuration, set the Site URL to http://localhost:5173 and add http://localhost:5173/** to the redirect allow list, or magic links / OAuth will send you somewhere that isn't your dev server.
5. (Recommended) Google sign-in — see Sign-in with Google below.
6. Sign in at /login. The first sign-in creates your account, and a trigger seeds your profiles and user_settings rows.

Sign-in with Google

The login screen offers Continue with Google, plus the email magic link. You create the OAuth apps; Supabase holds the secrets (never VITE_).

1. Google Cloud Console → APIs & Services → Credentials → Create OAuth client ID (Web):

• Authorized JavaScript origins: http://localhost:5173, https://prep-for-me.vercel.app
• Authorized redirect URIs (both):
  • Local: http://127.0.0.1:54321/auth/v1/callback
  • Hosted: https://<PROJECT_REF>.supabase.co/auth/v1/callback

2. Hosted Supabase → Authentication → Providers → enable Google → paste client id + secret for each.

3. Local — add to supabase/.env (or .env.local loaded by the CLI; not the Vite .env.local):

SUPABASE_AUTH_EXTERNAL_GOOGLE_CLIENT_ID=...
SUPABASE_AUTH_EXTERNAL_GOOGLE_SECRET=...

Then supabase stop && supabase start so GoTrue picks up [auth.external.*] from supabase/config.toml.

Redirect allow-list must include ${origin}/app (already required for magic links).

Scripts:

npm run build      # typecheck + production build to dist/
npm run preview    # preview the production build
npm run typecheck  # tsc --noEmit

Running AI locally

The hosted product can bill for cloud inference. Locally you pick the backend. Every generation call (analyze, rewrite, tailor, prep chat, ingest extraction) goes through ${ANTHROPIC_BASE_URL}/v1/messages. Point that at Anthropic or at a local Ollama bridge.

supabase start does not inject model keys. Serve functions yourself, from a file that process reads — not a VITE_ variable (those are inlined into the browser).

In .env.development.local:

VITE_AI_PROVIDER=edge

Restart Vite after changing it. Leave it mock (or unset) for labelled sample output with no model at all.

Only one supabase functions serve at a time — a second one takes over the shared runtime container. Stop the current serve, then start the env you want.

Option A — your own API keys (Anthropic + OpenAI)

Bills your Anthropic/OpenAI accounts. Analysis is roughly $0.09–0.10 on Claude; embeddings for prep chat/ingest use OpenAI.

1. supabase/.env.local (gitignored):

   ANTHROPIC_API_KEY=sk-ant-…
   OPENAI_API_KEY=sk-…
   # optional: ANTHROPIC_MODEL=claude-haiku-4-5-20251001

2. With the stack up:

   supabase functions serve --env-file supabase/.env.local

3. npm run dev in another terminal. Each Analyze / Improve / Tailor / Ask press is an explicit spend.

Option B — local LLM via Ollama (no Anthropic bill)

Generation hits Ollama. Prep-source embeddings still need OpenAI unless you skip ingest/chat retrieval.

1. Install Ollama and pull the strongest model that still feels fast on your machine for the whole product — resume analysis, tailoring, and prep chat, not only chat. What we used: qwen2.5:7b on a MacBook Pro with 64GB (M1 Max), and the same tag on an M4 Max; resume parsing and the other AI surfaces stayed usable.

   ollama pull <your-model-tag>

2. Copy supabase/.env.ollama.example to supabase/.env.ollama. Set ANTHROPIC_MODEL to the Ollama tag. Add OPENAI_API_KEY if you want prep chat / ingest.
3. Bridge (translates Anthropic /v1/messages → Ollama OpenAI /v1/chat/completions). Leave this terminal open:

   deno run --allow-net --allow-env supabase/functions/_stub/ollama-bridge.ts

4. Functions, pointed at the bridge (host.docker.internal:8788):

   supabase functions serve --env-file supabase/.env.ollama

5. npm run dev. The functions send ANTHROPIC_MODEL; the bridge uses that tag. Don't override it in the shell with a tag you have not pulled.

The functions container cannot see localhost on your Mac; the bridge must bind 0.0.0.0:8788 on the host. If analysis dies with ECONNREFUSED on host.docker.internal:8788, the bridge is not running.

More detail: TECHNICAL.md §4 and the shared-container gotcha in §10.

Not built yet

Discover's job-feed queries, Practice, and drag-and-drop on the kanban board. Each of these says so on screen instead of pretending. See PROJECT.md for what's planned and in what order.

Browser extension

A first version lives in extension/ — tailors your resume to the job posting in the current tab and autofills the application, using this app's own Supabase auth and Edge Functions. Real field-mapping so far covers Greenhouse plus a label-matching fallback for other sites; LinkedIn, Workday, and Lever are not mapped yet. See its own README for setup.