Skill Guide: Turn Repeated Work Into Reusable Skills

Ask AI to find the Skills hidden in your work

One is built; now find the next ones

You have a working . You built it from one repeated task, tested it, and saw the output follow a procedure instead of guessing. Now the question shifts: where else in your work are procedures hiding that deserve the same treatment?

is the practice of looking across your real work and asking one question: which procedures keep coming back?

A good is more specific than a you use sometimes. It has a , a repeatable input shape, a sequence of steps, an expected output, and a few rules the would otherwise miss. A long client email thread that always needs decisions, open questions, promised follow-ups, and a tactful reply is a Skill candidate. A vague wish to write emails better is too broad.

A stylized teaching image showing recent work sources sorted into ranked Skill candidates. — starts with the sources you already create, then filters for repeated procedures worth saving.

Search for procedures before you search for facts

When AI looks through your , the tempting thing is to summarize everything. That is the wrong job. You are trying to find repeated procedures that deserve a reusable home.

The difference is practical. 'Tom met with three clients last week' is a fact. 'Before each client meeting, Tom gathers the last thread, open questions, promised follow-ups, likely objections, and one suggested agenda' is a procedure. The second one can become a because it has a , inputs, steps, and output.

Useful sources reveal different kinds of repeated work

Source	What to look for	Possible Skill
Email	Similar replies, follow-ups, scheduling loops, scope clarifications, decision summaries	Turn a long thread into a polite next-action reply
Calendar	Recurring meetings, prep rituals, review blocks, handoff routines	Prepare a one-page meeting brief from related notes and messages
Past AI chats	Tasks you keep asking for, repeated corrections, stable preferences, recurring failure modes	Mine a finished conversation into a draft
Docs and Drive	Templates, reports, briefs, standards, examples of finished artifacts	Turn rough notes into the same report format every week
GitHub or issue trackers	Review criteria, release rituals, test expectations, repeated implementation mistakes	Review a pull request against a project-specific risk rubric
Slack or Teams	Handoffs, approvals, status requests, cross-team norms, unresolved decisions	Convert a long team thread into decisions, owners, and open questions

The same source can hold many facts and only a few procedures. The procedures are the valuable part for creation.

Try a small scan across meetings, emails, and drafts

Start with three recent work traces you are comfortable reviewing: one meeting transcript or notes file, one email thread, and one draft you asked AI to improve. You are looking for repeated work shape, not private facts.

A short discovery pass keeps the scan practical

Trace	Question to ask	Possible Skill candidate
Meeting notes	Do I keep extracting the same decisions, owners, deadlines, and risks?	Meeting recap with verification checklist
Email thread	Do I keep turning long back-and-forth into the same next-action reply?	Thread-to-follow-up draft
Revised draft	Do I keep correcting the same tone, evidence, or approval issue?	Draft review before sending

Private sources need a narrow first pass

is powerful because it can look through your real inbox, calendar, documents, messages, and code history. That same power is why the first scan should be narrow.

Start with metadata, summaries, titles, dates, sender names, meeting titles, thread snippets, file names, and issue titles. Ask AI to propose a source plan before opening full messages, attachments, private documents, client details, health information, financial details, HR material, legal material, credentials, or personal relationship content.

A stylized teaching image showing metadata and summaries scanned before an ask-first gate protects full content. — Discovery should start with low-risk surfaces and pause before opening sensitive source bodies.

Your turn

Discover reusable Skills from your real work

You'll mine your work patterns across all available sources and surface Skills worth building, refining, or formalizing.

Why this exercise matters

A comprehensive discovery scan across real work produces grounded candidates with evidence. Scanning metadata first and asking before opening sensitive content keeps the search productive without turning your inbox into a data dump.

You’ll leave with

A source inventory showing what the assistant can access and what is unavailable
A PARA-organized list of Skill candidates with evidence, reuse value, and interview questions for gaps
A ranked shortlist of the strongest candidates with your best next build target identified
Privacy notes covering what was skipped and what needs explicit approval

Use the prompt in order

1
Let the assistant inventory your sources
The asks for a source inventory first so you can see what is reachable before any searching begins.
2
Review findings as they arrive
The tells the assistant to present findings progressively rather than waiting until the end, so you can redirect the search in real time.
3
Answer interview questions per category
Where the assistant cannot infer a pattern, it will ask focused questions per PARA category instead of guessing.

Starter prompt text

Open the full text if you want to check what will be copied.

Show starter prompt

You are an AI Skills archaeologist. Your job is to mine my real work
patterns and surface reusable AI Skills I should build, refine, or
formalize.

**Discovery scope: be comprehensive.** Search across all available
sources: chat history and past conversations with me, active files and
work surfaces on my computer, connected tools (email, calendar, drive,
meeting transcripts, project boards), and anything else you can reach.
Look for:

- **Repeated workflows**: things I do more than once that follow a
  recognizable pattern
- **Standards and conventions**: naming schemes, formatting rules,
  architectural patterns, style guides, or house rules I enforce
  consistently
- **Universals**: patterns that apply across multiple projects or
  contexts (e.g., how I always structure a deliverable, how I kick off
  a new feature, how I handle client communication)
- **Project-scoped skill candidates**: workflows unique to a specific
  project but structured enough to templatize
- **Gaps**: places where I'm doing something manually or inconsistently
  that a skill could standardize

**Organize everything you find into PARA:**

- **Projects**: active work with a clear goal or deadline. What skills
  would accelerate each one?
- **Areas**: ongoing responsibilities I maintain continuously (business
  ops, client relationships, content pipelines, platform maintenance).
  What skills would reduce friction here?
- **Resources**: reference material, research, reusable templates, or
  knowledge bases I return to. What skills would make these more
  accessible or useful?
- **Archives**: inactive or completed work worth preserving. Are there
  latent patterns in archived work that deserve to become active skills?

For each PARA category, either **propose specific skills** (with a name,
trigger description, and what the skill would do) or **interview me**
with focused questions to fill in gaps you can't infer from the data
alone. Limit interview questions to 2-3 per category max.

**How to search:**

1. Start by scanning metadata: titles, file names, dates,
   senders/recipients, thread subjects, calendar event names, chat
   summaries, and snippets. This is your primary discovery layer.
2. When you need to go deeper into full message bodies, attachments, or
   document contents to understand a pattern, briefly explain what
   you're looking for and ask before opening anything that looks
   sensitive (financial details, health info, legal or HR material,
   credentials, confidential client specifics, or personal relationship
   content).
3. Don't let caution slow you down on routine work artifacts. Most of
   my files, chats, and emails are fair game at the metadata and
   summary level. Use judgment.

**Output structure:**

1. **Source inventory**: what you can access right now and what's
   unavailable.
2. **Search plan**: what you'll scan first, the queries or filters
   you'll use, and approximate date ranges.
3. **PARA-organized findings**: discovered patterns, proposed skills,
   and interview questions, sorted into Projects / Areas / Resources /
   Archives.
4. **Privacy notes**: a short paragraph on what you chose not to open
   and anything you'd need approval to inspect further.

Start by running the source inventory, then execute the search plan.
Present findings as you go rather than waiting until the end.

What the answer should give you

You should receive a source inventory, a search plan with queries and date ranges, PARA-organized findings with proposed Skills and interview questions per category, and privacy notes about what was skipped or needs approval.

What the discovery scan produces

The comprehensive above handles source planning, pattern searching, PARA organization, and ranking in one session. Each candidate should point back to the kind of source that produced it, the recurring situation, the usual inputs, the output, and the corrections you seem to repeat.

Pay attention to confidence. A pattern that appears in ten threads is different from a pattern the AI inferred from one ambiguous note. The asks the assistant to interview you when evidence is thin. A low-confidence pattern can still be useful, but it should not become a until you confirm it.

Rank candidates before writing files

A discovery run can produce too many candidates. Do not build all of them. Rank them first.

The best first is usually frequent, valuable, repeatable, easy to test, and low risk. That often means something ordinary: meeting prep, client follow-up, weekly review, draft review, long-thread summary, PR review, source-backed research brief, or a recurring document format.

Use scoring to separate good Skill candidates from clutter

Score	What it asks	High score means
Frequency	How often does this repeat?	Weekly or more
Value	What improves if this becomes consistent?	It saves time, prevents mistakes, or improves public work
Repeatability	Are the inputs, steps, and outputs stable?	The same pattern appears across examples
fit	Would a formal outperform a one-off ?	The task has procedures, examples, gotchas, or tests
Risk	Could a bad output expose data or cause harm?	Lower risk is better for a first
Implementation ease	Can this start as clear instructions?	No script or integration is needed at first

A stylized teaching image showing Skill candidates ranked across frequency, value, repeatability, Skill fit, low risk, and easy start criteria. — Ranking keeps discovery from turning every possible idea into a file.

The comprehensive discovery above handles ranking as part of the PARA-organized output. Each candidate arrives with evidence, so you can apply the scoring rubric above to choose your next build target.

Some findings belong outside Skills

often finds useful instructions that should not become Skills. A stable personal preference belongs in memory. A repository-wide working agreement belongs in AGENTS.md or a similar project instruction file. A recurring scheduled action belongs in an automation. A specialized parallel research role may belong in a subagent. A rare task can stay a one-off .

That classification step prevents bloat. A Skill library should feel like a small set of well-labeled procedures you trust, not a junk drawer of impressive-sounding files.

tests keep discovery honest

Before you turn a candidate into a real , test whether the is specific enough. Write prompts that should activate the Skill and prompts that should stay outside it. Near misses are the most useful tests because they show whether your description draws a clear enough line between tasks that belong to this Skill and tasks that do not.

A stylized teaching image showing a candidate Skill passing should-run, should-skip, and near-miss tests before build. — Discovery produces candidates; tests decide whether a candidate is ready to build.

Near misses make trigger testing useful

Candidate Skill	Should trigger	Should stay quiet
meeting-prep-from-calendar-and-email	Prep me for tomorrow's client call using the last thread and notes.	What meetings do I have tomorrow?
email-thread-to-followup	Turn this long thread into a reply with decisions and open questions.	Summarize this newsletter.
weekly-review-from-calendar-email-and-tasks	Help me review the week and decide what carries forward.	Schedule a meeting next Tuesday.
codex-pr-review-rubric	Review this diff for correctness, risk, missing tests, and production-readiness.	Explain what this function does.

Mini-project: run one bounded discovery scan

I chose one or two sources I am comfortable using.
I asked AI for a source plan before opening full content.
The scan returned repeated procedures, not private facts.
Each candidate has evidence, frequency, inputs, output, and confidence.
I ranked the candidates and chose one low-risk to build next.

References

6 sources

1
Best practices for skill creators
Agent Skills · 2026 · Agent Skills Documentation
The Agent Skills guide warns that Skills built from generic model knowledge tend to become vague. Stronger Skills come from real expertise, hands-on tasks, project artifacts, corrections, preferences, and execution traces.
Retrieved May 2026. Skills grounded in real work evidence produce better results than Skills drafted from abstract ideas.
View source
2
Apps in ChatGPT
OpenAI · 2026 · OpenAI Help Center
OpenAI documents that enabled ChatGPT apps may access relevant conversation context, may use relevant memory context when Memory is enabled, and are governed by each app's terms and privacy policy. Workspace admins can configure app availability and actions.
Retrieved May 2026. Connected apps can see your conversations and memory, which makes source-based discovery useful but also means you need clear boundaries about what the assistant opens.
View source
3
ChatGPT agent
OpenAI · 2026 · OpenAI Help Center
OpenAI's agent guidance says agent mode can access sensitive data when apps or websites are enabled, recommends enabling only the apps needed for the current task, and warns against vague open-ended prompts such as asking the agent to check email and handle everything.
Retrieved May 2026. Vague discovery prompts like 'check everything' create risk. Telling the agent exactly what to search and what to skip keeps discovery productive.
View source
4
Custom instructions with AGENTS.md
OpenAI · 2026 · Codex Developer Documentation
Codex reads AGENTS.md files before doing work, layering global guidance with project-specific overrides. Files closer to the current directory override broader guidance.
Retrieved May 2026. Stable project-wide guidance belongs in instruction files like AGENTS.md, which are always loaded for that project, rather than in a reusable Skill that might fire in the wrong context.
View source
5
Optimizing skill descriptions
Agent Skills · 2026 · Agent Skills Documentation
The Agent Skills description guide recommends realistic trigger evaluation queries, roughly balanced between should-trigger and should-not-trigger cases, with varied phrasing, detail, complexity, and explicitness.
Retrieved May 2026. Testing the description with realistic examples catches vague triggers before you commit to building the full Skill.
View source
6
Agent Skills
OpenAI · 2026 · Codex Developer Documentation
Codex includes an initial list of available Skills in context and recommends concise descriptions with clear scope and boundaries because implicit matching depends on the description.
Retrieved May 2026. Codex picks Skills based on the description alone, so a vague description means a useful Skill may never fire.
View source