cd /news/ai-agents/seven-tricks-for-making-better-actio… · home › topics › ai-agents › article
[ARTICLE · art-173594] src=dev.to ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

Seven Tricks for Making Better Action Games with Claude Opus 5.5

A developer built a set of browser-based action games using Claude Opus 5.5 and an open-source "Agentic Gamedev Skills" repository, in which an AGENTS.md file directs the agent to draw two random constraint tags from a tags.csv pool and implement a playable prototype under tmp/games/. The approach produced six arcade-style titles, including UNDERTOW (from the tags "sink" and "rescue") and ZIG SABER ("sword" and "laser"), with the developer reporting that injecting external randomness yields more varied game concepts than prompting the model cold.

recommendation

When an agent keeps proposing similar ideas, have an external script pick two random constraint tags and pass them in as fixed requirements, keeping the pair even when it is hard to combine.

more in playbook: Designing agent systems
by read12 min views14 publishedNov 21, 2026

Claude Opus 5.5 handles nearly every part of game development well: coming up with ideas, design, implementation, visuals, music, sound effects, and more. You can make a game just by saying "make a game like this" and then "I don't like this part, fix it." But with a few small additions to that process, you can get noticeably better game feel. This post describes one way to do that.

A collection of games created using Agentic Gamedev Skills.

Click a game image to play it in your browser.

Ask an agent to create a game in this repository. The repository instructions in AGENTS.md tell it to use your brief, or draw two constraints from tags.csv when the concept is open-ended, then use the relevant Agentic Gamedev Skills, implement a browser prototype under tmp/games/, and validate it.

To draw an optional pair manually:

node tools/pick-tags.mjs

The source code for each game in this repository lives under docs/.

The source for these games lives in crisp-game-lib-12-games.

They're all in the repo above and playable in the browser, so give them a try. They're action games that aim to capture the feel of old-school arcade games.

Game Description
WRECKFALL A fixed-screen shooter. Enemies you shoot down turn into wreckage that falls straight down, taking out any enemies in its path before crashing onto the ground where your ship is. Clear the descending enemy formation before it reaches the defense line.
UNDERTOW A side-scrolling submarine rescue game. Enemy ships pick up castaways from the sea and carry them back to their own port. Sink an enemy ship and its passengers are thrown into the water and start sinking; grab them before they hit the seabed and bring them to your own port.
HEAD RUSH A dot-eating race. Drive around concentric square lanes collecting dots, and change lanes to dodge rival cars coming the other way. Grab a power-up and you can ram the rivals off the track. Compete for the highest score in three minutes.
ZIG SABER A one-button side-scrolling terrain shooter. Your ship always moves diagonally; pressing the button flips its vertical direction and swings a sword at the same time. The laser is a power-up that lengthens the sword.
OVERTURN Pinball without flippers. Your only control is rotating the whole table, and the ball always falls toward the bottom of the screen. Every time the ball lands on the table's rim, the rim gets damaged, and if the ball falls where the rim has broken through, you lose it. Knock down a full set of targets and the rim is repaired by as many segments as the lamps you've lit.
PARALLAXIA A vertical-scrolling bullet-hell shooter. Its three modes — CHAIN, CLOSE, and GRAZE — differ in enemy bullet speed and in how you build up the score multiplier.

The following tricks make it much easier to get an agent to come up with varied game ideas and to tune the resulting games to a good balance.

If you ask an agent to come up with a game idea cold, it tends to propose similar things every time. So inject noise from the outside: for example, give an external script a pool of seed tags, have it pick two at random, and hand them to the agent as constraints.

For instance, the two words sink and rescue produced UNDERTOW, a submarine game where you sink enemy ships carrying castaways, throw them into the sea, and pick them up before they reach the seabed to bring them back to port. Even when a pair of tags was hard to combine, I didn't reroll — I had the agent use them anyway, even if that meant bending their meaning. ZIG SABER drew sword and laser. It became a one-button shooter where you swing a sword, with the laser as a separate power-up. After playtesting, that laser changed from a long-range beam into a mechanic that lengthens the sword.

If you already know exactly what you want to make, just give the agent that instead of tags. Some of the games above don't depend on tags at all.

Don't start building from the first idea. Have the agent come up with about ten, compare them, and drop any that are:

Before any code is written, have the agent put these four things into words:

The sentence in item 1 becomes the yardstick when you later ask a human player "Did it feel like this?" If the agent can't write item 2, the game will likely turn out monotonous.

The first version exists so a human can check whether the core play feels right. Build it at a small scale — one enemy type, one screen — and leave title screens and stage structure for later. Problems with game feel and readability only show up when a person plays, so getting it played once before adding content means less rework. Still, verify up front that simple play loses quickly and that danger arrives soon after the start so the player is never idle.

If you build bots (programmatic players) so the agent can play the game itself, testing and tuning game balance gets much faster. And if you can build a bot that matches the skill of the player — which, for a solo developer, is the developer themselves — you get close to the right difficulty, which makes playtesting less stressful. However, a player's habits can deviate from what the agent assumed, in either direction: better or worse.

Game Deviation
ZIG SABER The player struggled to line up their ship's height with enemies and couldn't take out formations the bot could.
UNDERTOW The player was better than the strongest bot, so difficulty tuned to the bots was too easy.
PARALLAXIA A bot with delayed reactions made 3–7 times as many mistakes than the human. The player wasn't dodging bullets after seeing them — they were anticipating and moving before the bullets were even fired.
HEAD RUSH The player made more than 20 mistakes in three minutes; the bot made fewer.

So capture the player's play logs and build a bot that plays the way that person does. Using the logs, match the bot to the player’s reaction speed, the situations where they tend to miss things, their aiming error, and the way they lose, then tune until the bot and the player get similar results on the same stage. Simply slowing down a strong bot's reactions won't get you close to the player. As with PARALLAXIA, imitate perceptual habits instead: the delay between seeing the screen and moving, or overlooking bullets when there are many on screen.

With this bot in hand, you can re-measure difficulty for that player every time you change a rule or layout, without asking them to play again. Enemy speed and spawn counts, lives, and extend values can all be decided while watching the bot's results.

Players improve as they play, so a bot isn't a build-once thing. In OVERTURN, the same player's results caught up to the next bot tier by their third session. Every time new logs arrive, check which bots the player falls between, and re-fit if they've drifted.

Beyond the player-like bot, preparing a few other bots lets you check balance from other angles:

Bot What it checks
Does nothing Does it lose quickly?
Simple input (hold, mash, fixed-interval actions) Does it clearly do worse than skilled play?
Strong (looks ahead and picks the best move) Can it still fail even when playing well?
Exploit seeker (stalling, running away) Is there any way to play that beats the intended approach?

Before using bot results, verify the bots themselves. Check that the strong bot is clearly stronger than the do-nothing bot, and that the strong bot's losses really are unavoidable. Sometimes the bot is just weak, and you mistakenly conclude the game is hard.

Also verify that your measurements are correct. In PARALLAXIA, the measurements said about half of the small enemies left without firing — but when the player loses a life, all enemy bullets on screen are cleared, and those cases were mixed into the count. Recounting only enemies that didn't overlap with a player death showed that the share leaving without firing was 4–47%, depending on the flight pattern.

If you can't get the game to a state where even the strong bot can fail, it won't be tense for skilled players. When you find it's too easy, change how score and danger are determined before raising enemy speed or numbers. UNDERTOW was too easy at first — there were no moments where enemies attacked you. So the agent gave some enemy ships destroyer escorts that drop depth charges set to the submarine's depth. Sink an escorted ship and the destroyer chases the submarine for a while. Sinking ships to rescue people now directly invites danger. As a result, players have to choose between attacking from in front of the enemy ship or from a depth the depth charges can't reach. Raising speed or numbers wouldn't have changed the fact that repeating the same action was the best strategy.

Even though the agent can handle balance tuning well, human playtesting is still necessary. Some problems can be found by having bots play and measuring; others only show up when a person plays.

Bots only find problems when a rule flaw shows up in the metrics. Issues like "you score without doing anything" or "you win by holding the button" can be found by having bots play the same level many times and comparing results across play styles.

These are the problems that only show up when a person plays:

All of these involve how people feel, see, and hear, and no matter how capable the agent is, they can't be verified without a person playing. The agent asks the player about the playtest and analyzes the play logs to find and address the following issues.

Ask whether the player felt the "what playing it feels like" sentence from section 3. If some part was experienced differently from the intent, change the rule causing that feeling. In ZIG SABER, the laser changed from a long-range beam into a mechanic that lengthens the sword, and every enemy within the sword's reach now gets the close-range slash bonus, so the power-up is always a benefit.

Ask whether the player could read enemies, hazards, scoring objects, and the rules that shape how to play while they were playing. If a player's questions reveal a rule they couldn't read, show the moment the rule takes effect on screen. In OVERTURN, unlit lamps are now drawn as empty sockets and lit lamps as glowing bulbs, and when the rim is repaired, the lit lamps fly to the spots being repaired.

Use the play logs to find which actions trip the player up. Fix the actions people struggle with by slowing enemies, widening hitboxes, and so on. In ZIG SABER, the player had trouble lining up their ship's height with enemies, so slashes now connect even with some height mismatch.

Ask about game tempo and ease of control. I decides the parts that come down to taste, but sometimes play logs can back things up with numbers. In HEAD RUSH, the rival cars were called a bit too fast, so they were slowed down and a brief window of invulnerability was added right after taking one out. In PARALLAXIA, my play logs showed a single CHAIN mode run lasting nearly twice as long as the other modes, so CHAIN mode's enemy bullets were sped up and the extend interval was widened.

You can measure volume numerically, but you can't know how something sounds without asking a person to listen. In WRECKFALL, the sound effects were called almost inaudible, and their volume was raised 8–10×. Judging whether the BGM fits the game or whether the sound effects are appropriate is still hard for today's agents.

Questions from the player after a playtest are clues to ways of playing the agent never tested. When you get one, build a bot that plays that way and check the results. In WRECKFALL, a bot that left the last enemy alive and kept shooting scored 1.2–2.2 times as much per stage. So the game was changed to end immediately when enemies reach the defense line, and the bonus grows the farther the enemy formation is from the line when you clear it, making it more profitable to clear stages quickly.

Once playtesting convinces you the game is fun, move on to the full version.

English | 日本語

This repository collects agent skills extracted from game-development work and related agentic-workflow research. Each skill lives under .agents/skills/, uses SKILL.md as its entry point, and may include references/, assets/, scripts/, tools/, or agents/ directories.

The main use case is building mini-games with one-button controls, strong visual feedback, procedural audio, telemetry-guided tuning, and optional pixel-art assets. The collection also covers pre-implementation concept exploration, adversarial concept review, and mechanically diverse portfolio curation. A few adjacent skills support those workflows: skill extraction, workflow refinement from real execution artifacts, and gating or dispatching expensive agent work.

Games built with these skills live in agentic-gamedev-games.

description field.SKILL.md as the canonical workflow for that capability.AGENTS.md for repository-level maintenance rules. Seven…

The skills are in the repo above — feel free to use them. That said, recent LLMs are so capable that I'm not sure these skills are really necessary. If you want to try any of the above, give this post and the skills repo to your agent of choice and talk it over with them. The agent will probably figure out the best approach.

── more in #ai-agents 4 stories · sorted by recency
── more on @claude opus 5.5 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/seven-tricks-for-mak…] indexed:0 read:12min 2026-11-21 · —