nhouseholder/skill-awareness

Skill Awareness — Use Your Tools

The Problem

Claude has 70+ specialized skills installed but doesn't use them. Real example:

• User asks to redesign a frontend
• frontend-design and ui-ux-pro-max skills exist — designed exactly for this
• Claude ignores both and writes components from scratch
• User has to ASK "did you use the frontend skills I installed?"
• Claude admits it didn't

This happens because Claude treats skills as passive background context instead of active tools to invoke. This skill fixes that.

The Rule (Non-Negotiable)

On EVERY prompt, before doing ANY work, ask yourself:

"Is there a skill, hook, command, or agent designed for this task?"

Then:

1. Scan the skills list in the system prompt (the Available skills section)
2. Identify matching skills by name, description, and trigger conditions
3. Use them — invoke the Skill tool, follow the skill's protocol, or activate the skill's checklist
4. If no skill matches, proceed normally

How to Check

The system prompt contains a list like:

- frontend-design: Create distinctive, production-grade frontend interfaces...
- ui-ux-pro-max: UI/UX design intelligence. 50 styles, 21 palettes...
- fix-loop: Self-healing CI loop...

Read this list. Match against the current task. If a skill's description or triggers match what the user asked for, USE IT.

Matching Rules

| User Request Contains | Check These Skills | |----------------------|-------------------| | frontend, UI, design, component, page, layout, style | frontend-design, ui-ux-pro-max | | fix, bug, broken, error, failing | fix-loop, systematic-debugging, pre-debug-check, website-guardian | | deploy, ship, push, release, go live | deploy | | plan, design, architect, break down | writing-plans, brainstorming, spec-interview | | test, TDD, coverage | test-driven-development, qa-gate | | commit, git, branch, merge, push | git-sorcery, version-bump | | research, learn, understand, how does | deep-research, know-what-you-dont-know | | backtest, model, accuracy, prediction | backtest, profit-driven-development | | website, webapp, site, page broke | website-guardian, site-update-protocol | | review, check, audit, scan | audit, reflexion, receiving-code-review | | parallel, multiple, concurrent | dispatching-parallel-agents, parallel-sweep | | API, Claude API, Anthropic, SDK | claude-api | | article, blog, write content | content-research-writer | | schedule, recurring, every X minutes | loop (skill), scheduled tasks | | PowerPoint, slides, deck, presentation | anthropic-skills:pptx | | PDF, fill form, merge PDF | anthropic-skills:pdf | | Word, docx, document | anthropic-skills:docx | | spreadsheet, xlsx, csv, Excel | anthropic-skills:xlsx |

What "Using a Skill" Means

Different skills are used differently:

1. User-invocable skills (like fix-loop, deploy, backtest): Call with the Skill tool
2. Always-on passive skills (like zero-iteration, anti-slop): Follow their checklist mentally
3. Triggered skills (like website-guardian, error-memory): Activate when their trigger conditions match
4. Heavy skills with agents (like frontend-design, qa-gate): Spawn subagents as the skill directs

The key insight: User-invocable skills MUST be actively called. They don't fire automatically. If frontend-design matches the task, you must either:

• Invoke it via the Skill tool if it's a slash command
• Follow its protocol manually if it's a behavioral skill
• Spawn an agent briefed with its instructions

Failure Modes This Prevents

| Failure | What Happens Without This Skill | |---------|-------------------------------| | Skill blindness | Claude has frontend-design but designs frontends from scratch | | Reinventing the wheel | Claude writes custom deploy scripts when deploy skill exists | | Missing quality gates | Claude skips verification when website-guardian should fire | | Manual research | Claude guesses domain logic when know-what-you-dont-know would force research | | No memory | Claude doesn't log bugs when error-memory should capture them | | Brute force debugging | Claude retries randomly when systematic-debugging has a protocol |

When Multiple Skills Match

If 2+ skills match the task:

1. Use ALL of them — they're designed to complement each other
2. Start with the most specific one (e.g., site-update-protocol before generic website-guardian)
3. Let the skill-manager handle weight limits if too many are heavy

Blocked Excuses (You May NOT Use These to Skip Skills)

Claude has a pattern of acknowledging skills then rationalizing why it didn't use them. These rationalizations are explicitly BLOCKED:

| Excuse | Why It's Wrong | What To Do Instead | |--------|---------------|-------------------| | "I applied its principles without formally invoking it" | Mentally applying ≠ using. The skill has specific checklists, scripts, and protocols that mental application skips. | Read the SKILL.md, follow each step. | | "The fixes were minor CSS/JSX changes" | Minor changes still benefit from frontend-design anti-slop guidelines and ui-ux-pro-max accessibility checks. | Run the skill's checklist even for small changes. | | "No test suite exists in the project" | webapp-testing uses Playwright — it creates its own tests. qa-gate can verify via curl, browser, or preview. | Use the skill's tools, don't require project tests. | | "The Explore agent handled the research" | Explore reads files. Skills provide expert judgment, checklists, and quality gates that Explore doesn't have. | Use Explore for reading, THEN apply skill protocols to the findings. | | "Those skills are more relevant for interactive debugging" | If the hook matched them, they're relevant NOW. | Use them now. | | "I verified manually via curl/build" | Manual verification is not a substitute for skill-driven verification with checklists. | Use the skill's verification protocol, then ALSO do manual checks. | | "It would add overhead for a simple task" | The user installed these skills specifically to avoid the bugs that "simple" tasks keep causing. Overhead is the point. | Use the skill. The overhead prevents regressions. |

The rule: If a skill was matched by the hook, you must either USE it or explain BEFORE starting work why it genuinely doesn't apply (not after you've already done the work without it).

Self-Check (Run Mentally on Every Response)

Before submitting your response:

[ ] Did I check the skills list for this task?
[ ] If a skill matched, did I use it?
[ ] If I wrote code manually that a skill covers, why didn't I use the skill?
[ ] Am I about to claim "done" without running verification skills?
[ ] Am I about to rationalize not using a skill? (Check the blocked excuses table)

If you catch yourself NOT using a matching skill, stop and use it. The user installed these skills for a reason.