built by aktagon.com
aigrd scores each file against a rubric you wrote from your own best work — a pass or fail per criterion, with a quoted span as evidence and a note on how to fix it.
aigrd is a quality gate for agent output, judged by a second model. Below the bar, the Claude Code Stop hook blocks the turn and the agent redoes the work. aigrd judges once per call. The harness owns the loop.
$ aigrd init Created aigrd.toml, docs/rubrics/ $ aigrd rubric derive --genre POST content/posts/a.md content/posts/b.md Wrote docs/rubrics/001-post.md — 12 criteria $ aigrd judge fail[C-001]: opens with a preamble, not the claim --> content/posts/a.md:1 | 1 | In this post, we will explore a few ideas about | ^^^^^^^^^^^^^^^^^^ | fix: lead with the claim the post argues 1 must failed · should ratio 0.83 (threshold 0.75)
Judge the work, not the diff. · One call. The harness owns the loop.
01 For work only a reader can grade
The judge sees the file and the criteria and nothing else, so a verdict cites the span it judged.
--- id: RUBRIC-001 title: Blog post genre: POST --- ## Criteria - **C-001** (must): Opens with the claim, not a preamble. - hint: the first paragraph - **C-002** (should): Every number names its source. # The judge sees the file and these criteria. # Nothing else.
02 Measure the judge, not just the work
Every judged file appends one line to .aigrd/runs.jsonl — model, thinking cap, cost, duration, verdict. Label the verdicts you disagreed with and aigrd agreement scores the judge against you, overall and per criterion.
$ aigrd label content/posts/a.md C-001 pass recorded — 10 labels on file $ aigrd agreement overall 0.90 (9/10 agree) C-001 0.60 judge stricter than you C-002 1.00 # "the judge seems harsh" is now 0.60
03 Every capability
Write aigrd.toml, docs/rubrics/ and the .gitignore line.
Draft a rubric from your best pieces, then edit it by hand.
Warn on any criterion a regex could have decided instead.
Score every file that matches a genre. Verdicts cached on content.
Print the Stop-hook entry to paste. It never writes settings.
Compare your labels with the judge's, per criterion and per model.
MIT licensed. The judge runs on your existing Claude login. Not yet publicly released.