Prompt Clarifier / Prompt Coach MVP

FastAPI + chat-style web UI implementing:

• Clarifier stage (/v1/clarify): Prompt A (clarified intent) and Prompt B (Vision B).
• Execution stage (/v1/execute): user selects A, B, or both and runs selected prompt(s).
• Direct stage (/v1/direct): bypass A/B and ask GPT directly.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
cp .env.example .env

Open .env and paste your API key. Then switch backend to openai.

PROMPT_COACH_BACKEND=openai
PROMPT_COACH_MODEL=gpt-4.1-nano
PROMPT_COACH_ANSWER_MODEL=gpt-4.1-mini
PROMPT_COACH_ENABLE_WEB_SEARCH=true
OPENAI_API_KEY=PASTE_YOUR_OPENAI_API_KEY_HERE

Run

uvicorn app.main:app --reload

Open: http://127.0.0.1:8000

API

POST /v1/clarify

Request:

{
  "message": "where should I go for sushi tonight?",
  "locale": "en-US",
  "context": "optional"
}

Returns latent goal, interpretations, assumptions, ambiguities, Prompt A, Prompt B, rationale, follow-ups, and confidence.

POST /v1/execute

Request:

{
  "original_message": "where should I go for sushi tonight?",
  "selected": ["A", "B"]
}

Returns one or two runs depending on selection.

POST /v1/direct

Request:

{
  "message": "where should I go for sushi tonight?"
}

Returns one direct answer without generating A/B prompts first.

Notes

• If PROMPT_COACH_BACKEND=openai and the key is missing/invalid, the API returns a clear 400 error.
• If backend is mock, no external model is called.
• .env is local-only and ignored by git. Commit only .env.example.
• Live web retrieval is enabled in both A/B answer mode and Direct mode when PROMPT_COACH_ENABLE_WEB_SEARCH=true.

Test

pytest