#34
Monthly Rank
#106
Lifetime Rank
Share
18 June 2026
Harness

Harness

Where lawyers learn to trust AI

Andrew BirdGitHubWebsiteVideo
Open Source

Gallery

About the Project

Your best lawyers quit AI first — not from fear, but because their expertise makes its failures loud: the invented citation, the confident wrong date, the case never decided. The people best able to catch those errors leave; the ones who can't, stay and ship the hallucinations (1,500+ court decisions and counting).

Harness reverses that — it shows one legal question three ways (raw Claude → Claude with skills → Claude under audit), so you watch the failure that makes experts quit, then watch it caught: a skill that won't invent authority, a citation check against the actual documents, and a human who signs the record. Then you take the harness home as a file.

Practice Areas

Key Features

Key features

  • Three ways, one question — raw → skills → audit, side by side
  • Plain-English skill strips slop and jargon
  • Citation/source-anchor check refuses invented authority
  • Audit view: what the AI saw vs asserted, each unsupported claim refused
  • Human sign-off → hash-stamped record
  • Live "try your own" under a guided CLAUDE.md (server-side key)
  • Take-home CLAUDE.md; steps 1–3 fully offline & deterministic

About the Creator

AB
Andrew Bird
Independent — legal AI at Building Legalise

I build with frontier AI. Currently focused on legal — ripe for disruption.

Jamming on a few projects: github.com/b1rdmania

Open to collaborators, particularly anyone interested in doing a UK AI law firm with regulatory approval paired with a tech firm.

LinkedIn