Detect kitchen-sink skill
What you'll learn
A kitchen-sink Skill bundles multiple unrelated goals. Symptom: vague description, partial adherence. Fix: split into one-purpose Skills.
- One Skill, one purpose. 'And' or 'plus' in the description is a smell.
- Symptoms: instruction drift, vague descriptions, partial adherence.
- Detection: description fan-out, unrelated body sections, partial-instruction reports.
A 'kitchen-sink' Skill is one that bundles multiple loosely-related goals into a single Skill — setup + linting + deploy in one Skill, or 'analyze + summarize + recommend' in one Skill. Symptoms: instruction drift (the model partially follows some goals while ignoring others), description-discoverability problems (the description has to mention all the goals, becoming vague), and reuse failures (users can't reach for a sub-piece without inheriting the whole bundle).
The principle: **one Skill, one purpose**. If the Skill's description has to use 'and' or 'plus' to enumerate three things it does, it's probably two or three Skills. Splitting yields focused descriptions, cleaner instruction adherence, and composable building blocks the user can reach for individually.
Detection signature: (1) description starts 'Helper for X, Y, and Z' or 'Various tasks related to…'; (2) the Skill's body has multiple disconnected sections targeting different audiences or use cases; (3) users report 'it sort of works for X but I have to nudge it for Y' — the model is getting partial instruction adherence on the multi-purpose body.
The fix is mechanical split: identify the distinct goals, factor each into its own Skill with its own focused description, and let the user invoke the right one. The kitchen-sink Skill becomes 2–3 Skills. Discoverability improves (each new description names one specific scenario), adherence improves (each invocation has a focused brief), and reuse improves (users can build by composing).
Key points
- One Skill, one purpose. 'And' or 'plus' in the description is a smell.
- Symptoms: instruction drift, vague descriptions, partial adherence.
- Detection: description fan-out, unrelated body sections, partial-instruction reports.
- Fix: split into focused Skills.
Examples
Description: 'Setup, lint, deploy, and explain code in this project.' Body: four unrelated sections. Users report mixed results across the four goals.
Each focused; each with a description naming one scenario. Users pick the right one; instruction adherence is clean per invocation.
Pitfalls
- Bundling 'because they all touch the project'. Bundle by *purpose*, not by *target*.
- Resisting the split because 'it'll be more files'. Files are cheap; cognitive load on the model and user is expensive.
- Ignoring the description-fan-out signal. If it takes 'and' to describe the Skill, you have multiple Skills.
Source notes: 00-academy-basics/notes/07-agent-skills.md
Ask Claude
Build a prompt with this lesson + your question, copy it, and open the Claude Project in a new tab.