structured-output

Project 03 of ai-dojo. Compare three structured-output techniques on a market-research questionnaire extraction task.

Techniques

1. Prompted JSON — plain prompt + schema dump, parse + retry
2. Tool use — Anthropic tools with tool_choice, parse from tool_use block
3. Strict tool use — Anthropic tool with full JSON schema + forced tool_choice, the closest analogue to OpenAI's strict json_schema response_format

Stack

Python 3.14 · uv · Pydantic v2 · instructor · anthropic · pytest · rich

Model: claude-haiku-4-5-20251001 (Anthropic only — OpenAI dropped).

Layout

src/structured_output/   # package
tests/                   # pytest
fixtures/briefs/         # hand-written research briefs (*.md)
runs/                    # sweep artifacts (JSON, gitignored)
notes/                   # technique comparison writeup

Setup

uv sync
cp .env.example .env  # fill ANTHROPIC_API_KEY

Usage

uv run extract --brief fixtures/briefs/01-coffee-subscription.md --technique tool_use
uv run sweep --briefs fixtures/briefs/ --out runs/

Success criteria

See project brief.