Fit

Fit is a 5-minute adaptive quiz that answers:

How do you like to use AI — and which setup actually fits you?

It is for everyday people: homework, home life, a shop or studio, a side hustle, or a small team. Not a corporate form.

You leave with:

1. Your AI style — a named interaction profile, not a personality type.
2. Why — the choices that produced that profile.
3. How you like to use AI — checking facts, tweaking, going deeper, handing work off, wanting sources, comparing options.
4. Tools that fit you — what each product should handle.
5. What to use each model for.
6. How your AI should talk to you.
7. Setup files — ChatGPT, Claude / CLAUDE.md, Gemini, Cursor rules, AGENTS.md.

This is a free individual quiz. It is not a $10/month “which AI should I use?” subscription.

Source: github.com/aking-beep/aifit-engine.

The public name is Fit. The repo stays aifit-engine so it is not confused with the unrelated business-matching site at aifitengine.com. Point a custom domain at the Vercel project and set NEXT_PUBLIC_SITE_URL to that origin for share cards.

What it measures

Fit scores what you ask, how you steer, and how much evidence you demand. Two people with the same job, budget, and skill level can need different AI setups.

The quiz usually stops after four scenes (about twelve interactions). It continues only when a core habit still is not clear, and never past eight.

An LLM may label optional free text. It never picks the winner. Ranking is dated registries + weights.

Limitations

• The catalog in data/registry/ was last reviewed on 2026-09-12. Re-run docs/REGISTRY_SEED_REVIEW.md when a vendor changes.
• Your quiz and results stay in the browser (localStorage). Server storage is filesystem locally, or Upstash / Vercel KV when UPSTASH_REDIS_REST_URL + UPSTASH_REDIS_REST_TOKEN (or KV_REST_API_URL + KV_REST_API_TOKEN) are set. Share links also embed a compressed copy of the result.
• There are no accounts.
• Keyword classification is first-class. The optional LLM classifier is off unless AIFIT_LLM_CLASSIFIER=1 and an endpoint is set.
• Fit scores are normalized similarity, not scientifically validated probabilities.
• Not a personality test, clinical instrument, or hiring screen.

Run locally

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest
aifit score examples/sample_session.json
aifit freshness
PYTHONPATH=packages/core/src python3 -m evals.harness validate-seed
PYTHONPATH=packages/core/src python3 -m evals.harness run-cases
PYTHONPATH=packages/core/src uvicorn services.api.main:app --host 127.0.0.1 --port 8472

In another terminal:

cd apps/web
npm install
npm run dev
# or a production-style preview: npm run build && npm run start

Open http://127.0.0.1:43123.

Frontend happy path:

cd apps/web
npx playwright install chromium
npm run test:e2e

Deploy on Vercel

This repo is a Vercel Services project: Next.js in apps/web and FastAPI at /v1 and /health on the same domain. Scoring stays Python.

npx vercel login
npx vercel
npx vercel --prod

Import the Git repo in the Vercel dashboard if you prefer. Leave the root directory at the repository root so vercel.json can see both services. Do not set the root to apps/web.

On Vercel, assessment events are buffered in the browser and scored in one request. Add Upstash Redis or Vercel KV so share IDs survive across instances. Without that, share links still work via a compressed snapshot in the URL hash.

Architecture

Short everyday scenes
        │
        ▼
How you actually use AI
        │
        ▼
Your AI style
        │
        ▼
What you need from tools
        │
        ├───────────────┐
        ▼               ▼
Product list        Model list
        │               │
        └───────┬───────┘
                ▼
         Tools that fit
                │
                ▼
      How AI should talk + setup files
                │
                ▼
     A setup you can paste today

The registry, methodology, and evidence live under How it works. They are credibility infrastructure, not the homepage.

Should I use AI for this? (/check)

A two-minute task check that compares five options — leave it alone, ordinary software, rules-based automation, an AI tool, or building something — with fixed, published rules (GET /v1/decision/rules). Free text is shown back but never scored. Results and the history list (/history) stay in the browser.

AI Fit Teams

apps/teams is a separate product, AI Fit Teams, with its own Vercel project and its own TypeScript packages (packages/teams-core, packages/ui, packages/calibration-contract, packages/platform, packages/deliverables, packages/control-export-contract). It does not link to Fit. See apps/teams/README.md.

Shared decision intelligence

packages/decision-core/rules/decision-rules.v1.json holds every weight and threshold of the five-outcome decision engine. It is implemented twice — TypeScript (@aifit/decision-core, used by Teams) and Python (aifit.decision, used by Fit's API) — and both test suites must reproduce packages/decision-core/fixtures/golden.json (15 workloads, every outcome) field for field, input hash included. After an intended rules change: npm run golden --workspace @aifit/decision-core.

See docs/ai-fit-2/ for the audit, plan and release report.

Agents (optional, off)

packages/agents (TypeScript) and aifit.agents (Python) add suggest-only model agents. They suggest intake answers, explain a decision in plain language, and check catalog prices and registry products every week. Each reply has to pass a schema and a deterministic check written in code before anyone sees it, and nothing is applied without the person. They run only when AIFIT_AGENTS=1 and ANTHROPIC_API_KEY are set. See docs/ai-fit-2/AGENTS.md.

Pack docs

1. docs/PRODUCT_SPEC.md
2. docs/ASSESSMENT_DESIGN.md
3. docs/SCORING_SPEC.md
4. docs/REGISTRY_SPEC.md
5. docs/PERSONA_SPEC.md
6. docs/ETHICS_PRIVACY.md
7. docs/ROADMAP_4_WEEKS.md
8. docs/DEFINITION_OF_DONE.md