Claude Fable 5 Games: Prompts and Plugins
Quick answer: The 3 July 2026 recording uses Claude Fable 5 on a small island-game example, then shows effort modes, plan mode, on-screen prompt briefs and the Compound Engineering plugin. The YouTube title says games in seconds. This page only repeats what that session actually ran, not a guaranteed build time.
For which Claude model to pick, use Claude Opus vs Sonnet vs Haiku vs Fable: which to use. First-time Claude Code setup is Claude Code tutorial for beginners: install and use it. The infographic in this video was made in ChatGPT Images: ChatGPT Image Editing: Sketch-to-Scene and Local Edits.
What this recording shows
The host had started testing Fable 5 after a re-release the day before this 3 July 2026 video. He opened with an island game as a simple one-shot example he had found, not as a timed bake-off of a new game from a blank folder. The rest of the recording is prompting practice, credit control and plugins.
A prompt pack sits at every.to/p/claude-fable-5-prompt-library. The page on camera said it was already used by 100,000 founders, operators and investors. That is the claim on that page in the recording, not a figure this guide independently counted. Copy a demo prompt only after you fill in details for your own case.
A Bridgebench graphic compared Fable 5 before a temporary pause with the post-1 July re-release. The host said the model had been nerfed and that he was still impressed, depending on effort mode. Those charts are what appeared that day, not current proof.
1. Effort modes and credits
The recording treats effort as the main dial. High is the default. Extra high is for the toughest work. Medium or low is for routine tasks. Max and ultra code were named as the modes that burn credits fastest. The host mostly ran medium or high. He did not have ultra code access in that session because credits were too low. Usage on camera was already maxed on Fable 5, with a reset in about 15 minutes, and occasional fallback to Opus 4.8.
Switch modes with:
/effortHe used medium for most of his own coding work in that video. Fable 5, he said, ate credits faster than any other model he had used so far. For a later claude.md cleanup he rolled back to Opus 4.8 on purpose. GLM 5.2 on the desktop is a separate walkthrough: ZCode GLM Desktop: Install, Plans and Review.
2. Plan mode and Superpowers
Rather than reading the full Claude platform docs, the host made a ChatGPT Images infographic of the takeaways versus Opus 4.8: longer-horizon autonomy, better first-shot correctness on complex tasks, stronger vision, more reliable docs and slides, and improved code review, debugging and repo search. That list is from the documentation graphic shown on 3 July 2026.
Short, clear instructions were recommended over long rule lists. A detailed prompt can start in Claude Desktop, then get shortened before it hits the coding harness. Combine that with plan mode so the model does not build immediately, and with Superpowers, described as a skill already common in Claude Code, Codex and other harnesses. When the plan is ready, switch back to build mode.
Shift+Tab cycles accept-edits-on, plan mode and auto mode. Other advice on the graphic: lead with the outcome and pause only when the task needs a person, ground progress updates in real tool results, put explicit boundaries in a claude.md file, and keep a memory system for lessons from prior runs. The docs line on screen said Fable performs best with harder work, clear boundaries, the right effort level and scaffolding for long autonomous runs.
3. Prompt briefs from the recording
Two on-screen briefs were read aloud. The auto-generated captions scramble parts of the first one, so this page does not reconstruct that full prompt. As described in the recording, that brief asks Fable 5 to decide which work is worth escalating, not to execute yet, then to turn the best candidates into ready-to-run Fable briefs.
The overnight brief was spoken as this opening line:
I'm handing you this task to run unsupervised overnight.A related slash command sets a goal Claude checks before it stops, which the host uses for repetitive jobs such as optimizing blog articles for AI search visibility:
/goalLoop runs a prompt or slash command on a recurring interval. The example spoken was about every 24 hours:
/loopFill in your own task after those commands. This page does not invent the rest of the prompt pack.
4. Compound Engineering and other plugins
Compound Engineering is the plugin Every said they use daily for a brainstorm, plan, build, review, test and pull-request pipeline, then to compound lessons after a run. On camera the loop was plan, work, review, compound and repeat. The pitch in the recording: 80% of the time in planning and review, 20% in execution, so Fable 5 is not spooling up for every small step.
In Claude Code the install step spoken was:
/plugin marketplace addThe same marketplace add was run in Claude X. After install the plugin showed 28 skills. Setup created a config file, which was allowed and added to gitignore. The recording does not walk every embedded skill.
A claude.md already existed in the project. A claude.md improver skill from skills.sh was installed to audit that file. For that simple edit the host switched to Opus 4.8. The audit score shown afterward was 95 out of 100. Memory Optimize, also from skills.sh, was installed to clean auto-memory files in four interactive steps (faster context loading, no duplicate rules, cleaner instructions, lower API cost, as listed on that page). Caveman was installed globally and turned on to cut verbose replies. The listing on camera showed 300,000 installs. Those install counts and scores are what appeared on 3 July 2026.
A later git commit and push to master, which deployed on Vercel, also ran on Opus 4.8. The host’s rule in this video: fall back to Opus 4.8 when the work is not mission-critical. Other Claude Code projects on this site include Claude Code Second Brain and Claude Code Video Editor Skill: Build and Run.
Watch the walkthrough
FAQ
Did the recording show that Fable 5 always ships a game in seconds?
No. The YouTube title uses that phrase. On camera the host showed a one-shot island game example he had found, then spent the rest of the video on effort modes, plan mode and plugins. This page does not treat seconds as a guaranteed build time.
Which effort modes were named?
High was the default. Medium or low were recommended for routine work. Extra high, max and ultra code were named for harder or costlier runs. The host mostly used medium or high, and did not have ultra code access in that session.
What is Compound Engineering in this recording?
A plugin installed from a GitHub marketplace add in Claude Code and in Claude X. The host described its loop as plan, work, review, compound and repeat, with most of the time in planning and review. The recording showed 28 skills inside it after install.
Should every task use Fable 5?
The recording says Fable 5 ate credits faster than any other model the host had used. For a simple claude.md cleanup he switched to Opus 4.8. Use the Claude models comparison page when you are choosing which model to run.
Sources
- Video: Claude Fable 5: Build Games in Seconds (Full Guide) (3 July 2026)
- Every, Claude Fable 5 prompt library: https://every.to/p/claude-fable-5-prompt-library
- Anthropic, Claude Code product page: https://claude.com/product/claude-code
You may also like
Build the next one with AI Income Lab
Step-by-step courses, reusable templates, and a community that answers when you get stuck.
Members get the Claude Code & Skills and Codex Tutorials & Courses lessons.
- Courses for builds like this one
- Copy-ready templates
- Ask questions while you build
- Weekly coaching on VIP
