How to use the Game combat skill polish agent skill

What is an agent skill?

An agent skill is a folder of instructions, reference files and optional scripts that an AI agent loads when a request matches its description. The skill's SKILL.md file tells the agent when to use it and how to work step by step.

This skill is aimed at video game projects. It is not a general coding helper.

What the skill makes the agent do

The skill covers two bars. The readability bar asks whether a skill reads as itself, and whether the result lands when the blow arrives. The art bar asks about detail, light, trails, impact and art direction, judged against a named reference such as Diablo IV. A score is always quoted with the rubric it came from.

The workflow runs in stages: pin down the job, film the real cast on a stopped, stepped clock, build contact strips and close-ups, have two blind judges score each round, then raise the lowest criterion and repeat. Each skill is shipped as its own version.

• Films casts headlessly at 0.05 s steps of game time
• Builds contact strips (timing) and close-ups (art)
• Anonymises footage and runs two-judge rounds, keeping the lower score
• Measures timing against the hit frame, draw calls, triangles and download size
• Recommends or applies fixes such as landing results on the hit frame

Install the skill

Download the ZIP from the skill's page on this directory and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/game-dev-combat-skill-polish
• Project: .claude/skills/game-dev-combat-skill-polish inside your repository

Prepare your game project

The skill cannot work without a running game development server that exposes review-only hooks on the page. The skill's film script expects hooks that stage a state, issue a command, aim the camera, and freeze and step the game clock. These hooks must be built for each game.

Project specifics such as hook settings, where each skill's code lives and scores so far can go in an optional local references/<project>.md file next to the skill.

• Python 3 with Pillow (PIL) for the image scripts
• Node.js for the Playwright-style film and performance scripts, plus a headless Chrome
• A running game dev server with review-only hooks
• Git, for worktrees and fetching other threads' changes
• Subagent capability to spawn two fresh judge agents per round

How to trigger it with natural requests

Ask in plain language. The skill activates when you ask to score, review, improve, polish or rebuild combat skills, spells or attacks, or when you say a skill doesn't make sense or looks bad. You can also name a target score or ask it to make a skill look like another game.

Name the skill and the bar if you know them. Start a new thread for each skill when several run at once.

• Score every combat skill in the project and list the weakest three.
• The Frost Nova doesn't make sense. The blow lands before the target moves. Fix the timing.
• Polish the Holy Hammer to 7 on the art rubric, matching the Diablo IV reference.
• Next skill: rebuild the Shield Bash so it has its own kneel-then-rise schedule.

What a review run produces

A review film produces a contact strip per cast and a close-up of the key frames. The judges receive these images with neutral ids, return JSON with part scores, a total, the biggest fault and one change to raise it most, and the replies are saved under judges/. The agent keeps the lower of the two scores and re-judges before calling a 7 solid.

Typical fixes in the source include giving a skill its own schedule and keyed move, making results land on the beat, and modelling a hero prop from a painted concept with image-to-3D.

• Results land on the hit frame, not the click
• Painted additive textures instead of procedural rings
• Camera-facing ribbons that grow with the path
• Light placed on the object or beside the body, not inside it