<< All versions

Skill v1.0.0

LLM-judged scan90/100
rahulchaube1/-skills-for-agents/evaluation-harness-design
──Details
PublishedOctober 6, 2026 at 02:13 AM
Content Hashsha256:2799ef46b1791684...
Git SHA
──Files
Files (1 file, 1.9 KB)
SKILL.md1.9 KBactive
SKILL.md · 52 lines · 1.9 KB

name: evaluation-harness-design description: Designs repeatable evaluation harnesses for AI and software workflows. Use when you need to measure quality, compare approaches, or prevent regressions from landing.


[!NOTE]
Skills for Agents card
Every skill in this repo now follows a cleaner, card-like reading experience.
Keep the workflow focused: Overview → When to Use → Process → Verification.
[!NOTE]
Evaluation Harness Design
Create stable scoring and comparison loops.
Workflow lens: Overview → When to Use → Process → Verification

Evaluation Harness Design

Overview

Evaluation Harness Design is a focused workflow for handling this class of work with more clarity and less rework.

When to Use

  • The task fits the evaluation harness design workflow.
  • You need a repeatable process rather than an ad hoc answer.
  • You want evidence, not just a plausible result.

When NOT to use: Pure curiosity, unrelated cleanup, or a task that belongs to a different phase.

Process

  1. Clarify the target outcome.
  2. Gather the minimum context needed to act.
  3. Apply the workflow in small, verifiable steps.
  4. Verify with concrete evidence before moving on.

Common Rationalizations

RationalizationReality
"This is too small to need a process."Small changes still create hidden regressions.
"I can verify later."Verification is cheaper before merge than after release.
"The obvious path is probably fine."Explicit assumptions beat silent guesses.

Red Flags

  • The task is moving forward without a clear outcome.
  • The workflow is being compressed into a single unverified step.
  • Output is being judged by appearance instead of evidence.

Verification

  • [ ] The result matches the requested outcome.
  • [ ] Assumptions are documented if anything was ambiguous.
  • [ ] The change can be reviewed or rerun by someone else.
All versionsv1.0.1 →