Discovery Copilot

Turn a raw requirements meeting into structured, ready-to-groom user stories.

Discovery Copilot is a small enterprise-AI agent that reads the transcript of a business meeting and returns clean user stories, each with acceptance criteria and a suggested target module. It is built for the messy first mile of any ERP or software project: the moment when stakeholders talk for an hour and someone has to turn that noise into a backlog.

Why this exists

Requirements discovery is where most enterprise projects quietly leak value. The knowledge is in the room, but it dies in an audio file or a wall of notes. This agent compresses the gap between "what the client said" and "what the team can build".

It is intentionally narrow: no UI, no meeting integration, no database. One transcript in, structured user stories out. That constraint is the point.

What it does

• Input: a single meeting transcript (plain text or markdown).
• Process: an LLM agent extracts discrete user stories, writes acceptance criteria, and proposes the functional module each story belongs to.
• Output: a clean markdown backlog of user stories.

Quick start

pip install -r requirements.txt
export ANTHROPIC_API_KEY=your_key_here
python src/extract.py samples/transcript_sample_01.md > outputs/my_stories.md

See outputs/stories_sample_01.md for an example of the expected output, generated from the fictional meeting in samples/transcript_sample_01.md.

Design notes

• The extraction logic lives in a single, auditable prompt (prompts/extract_user_stories.md), so the behavior is transparent and easy to govern rather than buried in code.
• The sample transcript is fictional. No client data is included.

Roadmap (not in this slice)

• Batch mode for multiple transcripts.
• Deduplication and clustering of overlapping stories.
• Export to project tools (Jira, ClickUp, Odoo Project).