Small Fish

Find local businesses by what they actually do, prove every match, and charge only for matches.

A user types a place, a type of business and what they are looking for, in plain English — "med spas in Dallas that offer Botox and have no online booking." Small Fish reads each candidate business's own website and returns only the ones that match, each with the sentence and page that prove it. Businesses it cannot judge are shown as "couldn't tell", never guessed.

Where things are

Path	What it is
PROJECT_PLAN.md	Start here. Stages, gates, every task, live measurements.
CLAUDE.md	Working rules, and links to the founding documents.
docs/critique.md	Review of the founding documents; the reasoning behind the plan's departures from them.
docs/stage0-coverage-report.md	Measured results from Stage 0.
docs/icp-discovery.md	Design for the ICP discovery flow. Not built.
stage0/	Stage 0 "prove it" work: coverage measurement, the engine, the accuracy benchmark.

The three founding documents — product, PMM foundations and GTM — are live Claude Docs, not files in this repo. Their links are in CLAUDE.md. Strategy lives in the docs; numbers and decisions live in PROJECT_PLAN.md, and where they disagree the plan wins and records why.

Current stage

Stage 0 — Prove it. Nothing is built for users yet. Stage 0 exists to answer four questions with measurements rather than estimates:

1. Does open places data see enough of each niche, or does Google gap-fill break the unit economics?
2. Can the engine judge criteria at ≥ 90% precision — including on absence criteria like "no online booking", which is the hardest case and the one the marketing leads with?
3. How many true matches does it actually show (recall)?
4. What does a match really cost?

The gate to building the product is in PROJECT_PLAN.md. It is a hard gate.

Stage 0 quickstart

Requires Python 3.11+.

pip install duckdb

# S0-02 — pull Overture Places candidates for the three benchmark markets.
# Reads the public parquet release over HTTPS; no download, no AWS credentials.
python3 stage0/src/coverage/overture_extract.py

Benchmark markets and criteria are defined in stage0/fixtures/benchmarks.json. Outputs land in stage0/data/.

Principles

Carried from the product document, and binding on the code:

1. Relevance over volume. A short list the user trusts beats a long list they must re-check.
2. Show the proof. Every match carries the sentence and page that justify it. If we can't show it, we don't claim it.
3. "Couldn't tell" is an honest answer. A forced guess looks identical to a real match and destroys trust.
4. Never invent. No guessed emails, no made-up facts, no contact the business did not publish.
5. Charge for value, not effort. Users pay for matches; scanning cost is ours to engineer down.
6. Plain English in, useful list out.
7. Respect the businesses we read. Public business information only, polite crawling, and a way for any business to opt out.