
Agent use case packs: real jobs, step by step, page 2
A use case pack walks one specific job from blank page to result with an AI agent: the plan, the prompts for each stage, what usually goes wrong and how to check the output.

Never let a booking agent type the service name itself
A team building booking agents found the model sometimes booked services that did not exist. Their fix: the model never names the service. Code maps the request to a real catalogue id, and free text reaching the booking API is blocked.

Point an agent at the order exceptions, not the routine work
Instead of automating the 90% of orders that already run fine, one operator aimed an agent at the messy exceptions. The agent drafts a fix for each one and a person approves it with one click, which cut the daily exception queue from most of a day to about an hour.

Port a Three.js game to Unity and keep the hand-tuned physics
Moving a browser racing game to Unity for phones, the developer asked Claude Code which Asset Store engine to use. It read the code and advised using neither, then ported the 170 lines of drift physics line by line so the feel stayed the same. The port and a new UI took about two hours.

Turn a folder of raw footage into a finished edit with an agent skill
video-use is an open-source skill from the Browser Use team. You drop raw takes in a folder and ask Claude Code or Codex for an edit. It cuts filler words, grades color, adds subtitles and animated overlays, then checks its own render before showing you.

Render repeatable videos from HTML instead of a video model
A hands-on test of HyperFrames, an open-source tool that turns timed HTML pages into MP4 with headless Chrome and ffmpeg. Same input gives the same video, so changes are code edits instead of another roll of the dice. The tester also wrote down the gotchas.

Make a 3D product ad with Claude Code and headless Blender on an old GPU
A 15-second, four-shot product spot rendered on a 2016 graphics card. Claude Code wrote the whole Blender scene in Python, Blender rendered it with no window open, and ffmpeg put the shots, music and overlay together. The builder posted the render times and what broke.

Reverse engineer a console kernel with Claude Code so new controllers work on old hardware
Claude Code read the Xbox 360 kernel's USB stack through a Ghidra MCP server, ported an open Linux driver to C and built a boot plugin, so up to four Xbox Series controllers pair wirelessly with a modded 360.

Share one small kanban board with your coding agents through MCP
Kanzo is a tiny self-hosted kanban board with a built-in MCP server. Claude Code or Codex can take the next task, comment on it and move it, and its tool list costs roughly 800 tokens of context.

Keep an app's rules in one data file so the agent can change behavior without touching code
A builder made a 90s-style scrolling cable TV guide for Plex, Jellyfin and Emby with Claude. A written spec came first, all channel and schedule rules live in one JSON file, and later ports used the first app as the reference.

Judge agent-written libraries with test suites your side did not write
A team built open-source C# libraries with a factory of agents (planner, project manager, orchestrator, engineers, reviewers) and scored each one against public conformance suites such as test262 and html5lib, publishing the failures too.

Give agents reverse engineering output they can check, not guess
borer is an early open-source reverse engineering engine built for agents. It is one Rust binary that runs as an MCP server, keeps lists paged and capped, and checks decompiled C against the machine code in an emulator so the agent knows what is proven.

Put Grok Bot in front of a Hermes agent so unattended jobs keep running
A builder wired Grok Bot to a local Hermes Agent API: Grok Bot starts runs, streams events, polls for results and relays approvals, while Hermes handles scheduled jobs that run without a person clicking approve.

Have the agent build a tool that runs without it
After 30 days with an always-on agent, one user's main lesson was that repeated agent steps are fragile and costly, and the better result is a script or small app the agent writes once that then runs on its own.

Tag every decision in your game's log with who made it
Thirteen weeks into a fishing and exploration game, a solo developer keeps a Decisions.md with more than 350 entries, each tagged [USER] or [PROPOSED], so the agent knows which choices are settled and which it may revisit.

Build your own render tool for a brand look with Claude Code
A designer used Claude Code to build a browser tool for liquid 3D shapes, with a WebGL2 path tracer and denoising, that exports 4K loops. The point: designers can now make the tool, not only the image.