Skip to content
← Selected projects

Skill Eval Pack

2026 · open source · claude code + codex

A useful-sounding skill is easy to write and hard to prove. I built this pack to compare the same agent on the same task with and without the skill, then keep it only when fresh heldout cases clear the gate. Code on GitHub.

Figure 01 · AI toolchain · three open-source projects

Write the instruction, prove it helps, check the facts

Skillpack supplies the instruction. Skill Eval Pack compares it with the same agent and task without the skill. Groundplane then checks declared output fields against recorded tool facts at runtime.

  • Inline
  • Queued
ONE JOB AT EACH STAGESkillpackload SKILL.mdSkill Eval Packblind heldoutsKeep or retiremachine gateGroundplanefacts + checksAgent outputpass or raiseSkillpackload SKILL.mdSkill Eval Packblind heldoutsKeep or retiremachine gateGroundplanefacts + checksAgent outputpass or raise
fig. 1 — the three projects cover different failure points. Skillpack supplies the instruction. Skill Eval Pack asks whether it improved the work. Groundplane checks declared facts when the accepted instruction runs.

The control is the same agent without the skill

The baseline and treatment run in separate fresh contexts with the same model, tools, task and rubric. Development cases are for fixing the candidate. Fresh heldouts make the decision. Once a heldout result influences a revision, that case is retired.

The judge does not get the answer key

eval_gate.py prepare counterbalances anonymous A/B packets. Independent judges see complete transcripts and outcomes, while exact properties run through deterministic graders first. The decision unions critical failures instead of averaging them away. A treatment can score well overall and still fail for one core regression.

A failed revision does not replace the proven one

New candidates get up to three serious attempts by default. An existing skill keeps its last proven version until a revision passes. The lifecycle helper checks the attempt history, retired heldouts and final action, while the payload helper confirms that Claude Code and Codex loaded the exact frozen files.

unit tests
54
gate versions
3 · older bundles still accepted
decision scripts
3 · evaluation, lifecycle, payload
clients
Claude Code + Codex
model calls
0 · supplied by the host
runtime
Python standard library
status
MIT · github.com/ong6/skill-eval-pack

// AI TOOLCHAIN

Instructions, evidence, runtime facts

Ong Jun Xiong

ENGINEER · BUILDER · SINGAPORE

ContactHobbiesArchiveNotesUI PackGitHubLinkedInSource

© 2026 Ong Jun Xiong