Which skills do people actually keep installed for their AI coding agent?
Skill lists for coding agents keep getting longer, but I suspect most people only use a handful day to day. Which skills have actually earned a permanent place in your setup for Claude Code or similar agents, and which ones did you try and drop?
The skills that stick are the ones that change how the agent works rather than what it knows. Coding agents already know most libraries; what they lack is discipline. Skills that make the agent plan before editing, find the root cause before fixing, and write a failing test before the code tend to stay installed because they help on every task, not one kind of task.
Skills that get dropped are usually broad reference dumps or role prompts such as 'senior engineer' with no steps. They cost context on every run and rarely change the result. If you cannot name the behaviour a skill is supposed to change, it probably is not doing anything.
A practical starting set is three: one for planning, one for debugging and one for testing. Use them for a week, then read a few sessions back. Keep the ones the agent actually triggered and that changed what it did; remove the rest.
Check each skill's source before installing it. A skill is instructions the agent will follow with your permissions, so read the file, look for scripts it asks to run, and prefer ones from repos with a clear licence and recent maintenance.
Listings mentioned
- Writing plans · skill by obraTurns a spec into small tasks with the files, code and test for each step, and does not allow TBD placeholders.
- Systematic debugging · skill by obraMakes the agent find the root cause before changing code, and stop after three failed fixes.
- Test-Driven Development (TDD) · skill by obraThe agent writes a failing test first, watches it fail, then writes the least code needed to pass it.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.