<< All versions

Skill v1.0.0

currentAutomated scan100/100
tiger-ai-lab/rewardharness/evaluating-multiple-constraints
──Details
PublishedOctober 16, 2026 at 09:30 AM
Content Hashsha256:96110ab8fffa7b3a...
Git SHA0c1e7051ab4f
──Files
Files (1 file, 1.2 KB)
SKILL.md1.2 KBactive
SKILL.md · 18 lines · 1.2 KB

version: "1.0.0" schema_version: 2 name: evaluating-multiple-constraints kind: skill description: How to systematically evaluate prompts with more than one instruction.


Evaluating Multiple Constraints

When a prompt contains multiple distinct instructions (e.g., "Remove X AND brighten Y", "Change A to B WITH a C"):

  1. Deconstruct the Prompt: Break the prompt down into a checklist of atomic constraints.
  • Example: "Change the Wii-remote to a clear plastic bottle with a nipple." -> Constraint 1: Replace Wii-remote with clear plastic bottle. Constraint 2: The bottle must have a nipple.
  1. Evaluate Independently: Check each edited image against every constraint on the checklist. Use visual analysis tools to verify specific details if unsure.
  2. Strict Prioritization: An image that successfully fulfills ALL constraints is strictly better than an image that fulfills only some, even if the partially successful image has slightly better visual quality.
  3. Identify Partial Failures: Explicitly note which constraints were missed (e.g., "Image A added the bottle but forgot the nipple").

<!-- Release review: RewardHarness v0.4.1 (2026-08-26). -->

All versions