# Make your agent play the game before it says done

Free sample from Game Dev with AI Agents by PaperChaseWebb. Written for Claude Code, Codex and Claude.

## The problem

An agent can report a game as finished because a test script moved the player through the level. Real players press keys and tap screens. In one public comparison ([source](https://x.com/theSethian/status/2110440063821545479), credit @theSethian), the agent that tested with real controls found the steering was reversed. The faster agent that tested with a script never noticed.

## The rule to add to your agent

```text
When the game runs, play it yourself using the same inputs a player would use (key presses or taps sent to the game window), not by setting positions in code. Complete one full round. Log what you pressed and what happened in PLAYLOG.md. Fix every bug you hit, replay, and only then say done.
```

## The playtester skill

Save as `.claude/skills/playtester/SKILL.md` (Claude Code) or add to `AGENTS.md` (Codex).

```markdown
---
name: playtester
description: Use before reporting any gameplay milestone as done. Plays the game with real player inputs and logs bugs.
---
1. Launch the game the way a player would.
2. Send real inputs to the window. Never set positions or state directly.
3. Complete one full round: start, play to win or lose, restart.
4. Write PLAYLOG.md: input, expected, observed, bug yes or no.
5. Fix each bug, then replay from step 1 until a clean round.
6. Report rounds played, bugs found and fixed, and anything you could not test.
Never upload or publish. Ask first.
```

## PLAYLOG.md template

| Time | Input | Expected | Observed | Bug |
|---|---|---|---|---|

## Want the rest?

The full pack adds the seven-stage method, a staged brief template, 16 stage prompts, four more skills (frame critic, perf runner, asset smith, release gate), a benchmark harness, asset-as-code recipes, a 48-hour jam checklist, engine notes for Godot, Unity, Three.js and Unreal, and 30 game concepts.

Contact: mgmt@paperchasewebb.com