Skip to content
Hands-on AI Playbook

Find Your Skill Candidates (Quick Prompt)

Why This Matters

Section titled “Why This Matters”

The best agent skills come from work you already do repeatedly — not ideas you brainstorm on the spot. This prompt puts the AI in the analyst seat: it reviews whatever it can see of your work — your calendar, email, and recent files where it’s connected to them, plus its memory of you and your past conversations — then asks you a few questions to fill the gaps. It returns a ranked table of skill candidates and recommends two to build first: one reference skill and one task skill.

Use this when you want a fast, conversational discovery in one chat — ideal for live sessions or when you want the AI to do the heavy lifting. Prefer a worksheet you can fill out on your own time? → How to Discover Your Best Agent Skills (Worksheet).

How to Use This Prompt

Section titled “How to Use This Prompt”
  1. Copy the prompt from the code block below using the copy button in the top-right of the block
  2. Optional: fill in the “Where I’ll build” line, only if you’ll build your skills in a different tool from the one you run this in. The AI uses it to judge what that tool can do (for example, whether skills there can reach your local files).
  3. Check the “Sources you may use” line. By default the AI may look at your calendar, email, recent files, its memory of you, and your past conversations. Delete any you’d rather it didn’t use; the prompt works with none of them.
  4. Open a new conversation in that tool and paste the prompt
  5. Read the first line it returns. It tells you which of your allowed sources it can actually reach. If it can reach none, it skips straight to the questions. That’s expected, not a failure.
  6. Answer one question at a time in Phase 2. Be specific about real work, not ideal work.
  7. Review the recommendations in Phase 3: a ranked table, then one reference skill and one task skill to build first.

The Prompt

Section titled “The Prompt”
You are an AI workflow analyst helping me find the best work to package as Agent Skills: reusable folders of instructions and reference material that an AI loads when a task needs them, so it handles that task the same way every time.


Where I'll build: [Optional. Leave this as it is if I'll build my skills in the tool I'm using now. If I'll build somewhere else, replace this with that tool, e.g. Claude Code, ChatGPT, Microsoft 365 Copilot.]


Sources you may use: my calendar, my email, my recent files, your memory of me, and our past conversations. [Delete any you don't want the AI to use, then delete this note.]


Two kinds of skill matter here:
- A **reference skill** holds knowledge my work draws on: brand or voice guidelines, a style guide, product facts, a glossary, policies, templates, example outputs.
- A **task skill** holds the steps for one job I repeat, such as drafting a client email, writing a weekly update, or summarizing a meeting. It applies a reference skill when it needs one.


What makes a good candidate (any one is enough):
- Work I repeat, with a recognizable trigger ("every Monday", "after every client call") and output that looks the same each time
- Context I give the AI over and over: guidelines, background, examples
- Instructions I keep re-typing: "always include X", "never do Y", "use this tone"


Leave out: one-off creative work, pure judgment calls with no rubric, and work that needs systems the tool I'll build in can't reach.


Rules:
- Use only the sources I allowed above, and only the ones you can actually reach.
- Never invent or guess details about me. If you have nothing to go on, say so in one line.
- Describe patterns, not private details: don't quote emails, names, figures, meeting titles, or file names, and skip anything that reaches you through a source I removed.
- If you don't know a value, write "?" rather than estimating.


We'll work in three phases.


---


Phase 1 — Review my work


First, in one or two lines, say which of my allowed sources you can actually reach (or say none).


If you can see my calendar, email, or recent files, look at the last two weeks for work I repeat: recurring meetings that lead to the same follow-up, emails I write the same way, documents I produce on a schedule. Then add what you already know about me from memory or past conversations.


List the candidates you can actually point to, one line each: the task, why it stood out, and where you saw it (calendar, email, a file, memory, a past conversation, or "you told me"). If you have nothing, say so and go to Phase 2.


---


Phase 2 — A few questions


Ask me one question per message, and wait for my answer before the next. Ask about whatever Phase 1 didn't cover:
- Work I do weekly or more often, what it produces, and roughly how long it takes
- Guidelines, background, or examples I give the AI every time
- Instructions I keep re-typing
- Tedious work I put off, or work with a checklist I follow


Keep it short: stop when we have at least five candidates, usually after three to five questions. If I say "keep going", ask more.


---


Phase 3 — Recommendations


Give me a ranked table:


| Rank | Task | How often | Time each time | Knowledge it draws on | Where you saw it | Potential |


Potential: 🟢 High (clear trigger, same steps, same output), 🟡 Medium (mostly structured, needs some judgment or missing context), 🔴 Low (too creative or one-off for a skill today). If tables display poorly here, use a numbered list with the same fields.


Judge against what the tool I'll build in can do. If skills there run in the cloud, they work only on what I upload or paste in; a task that needs my local files, inbox, or a live system is at best 🟡, and say why.


Then recommend two to build first:
1. **A reference skill**: knowledge my work draws on again and again. If nothing more specific stands out, suggest my brand or writing guidelines. If that knowledge is a long document, plan for me to upload it whole rather than condense it.
2. **A task skill**: the one task to start with, usually your top-ranked task (if not, say why). It doesn't have to use the reference skill. Prefer something small: a few steps, in order, so I could build both skills in about 25 minutes. Say why it pays off (how often × time saved × consistency).
3. For each of the two, what to give the AI when I build it: the steps, the source documents, and one or two real examples.


Finally, name any candidate better served by something else (an agent, a connector, or a saved prompt), and if the tool I'll build in doesn't support Agent Skills, name its closest equivalent. If you're not sure what that tool supports today, say so instead of guessing.


---


Start now with Phase 1.

What to Expect

Section titled “What to Expect”
  • Sources line — Before anything else, the AI says which of your allowed sources it can reach. With connectors, that includes your calendar, email, and files; in a tool with memory, your history. In a fresh chat or a CLI, it should say it has nothing and move on. If it claims to remember things you never told it, stop and start over — you’re in a tool without memory and it’s confabulating.
  • Phase 1 — Evidenced candidates only, each with where it came from. Skipped entirely when there’s nothing to review.
  • Phase 2 — One question at a time. Expect three to five; say “keep going” for more. Answer with specifics (“Monday status update for my manager”), not categories (“status stuff”).
  • Phase 3 — A ranked table, then two to build first: one reference skill and one task skill, with what to give the AI for each. It also flags candidates better suited to an agent, a connector, or a saved prompt.

Next Steps

Section titled “Next Steps”
  • Take your #1 candidate to the Skills setup guide and install it in your platform — or, if you’re on Claude, Installing Skills on Claude
  • Re-run this prompt in a few weeks as new patterns emerge in your work
  • Browse Claude Skills Resources for examples and deeper reference material — Claude-specific, but the structure and testing advice carries across platforms
Section titled “Related”