You ask AI to summarize a meeting. Several revisions later, it finally separates proposals from decisions, lists follow-up work, and stops guessing who owns each task. At the next meeting, you find yourself explaining the same rules again.

Once you can recognize a good result, you can preserve the method that produced it. Agent Skills offer a way to package that knowledge for reuse.

The previous article on working with AI in feedback loops covered understanding the problem, doing the work, and learning from evidence. This article asks how those lessons can improve the next task.

What is an Agent Skill?

The Agent Skills documentation describes skills as task-specific instructions and resources. A skill folder contains SKILL.md and may include scripts, references, and templates. Compatible systems can load the details when needed.

Think of a useful handover guide: when to use it, what inputs it needs, what to do, and how to check completion. Writing such a guide supplies working context; it does not train the model or change its weights. Results still need review.

Choose one recurring task you know how to review

Start with something frequent, with a good example and clear acceptance criteria: meeting follow-ups, a team document format, or an article prepared from sourced notes.

Our hypothetical example is turning meeting notes into a draft the team can act on. The input already exists, and the task ends at a reviewed draft. Sending messages or creating tasks elsewhere is outside this example’s scope.

Take one accepted summary and compare it with its source. “It reads well” is a starting point. More useful criteria include keeping proposals distinct from decisions, tracing every follow-up to the notes, and making missing information visible.

A five-step loop: succeed, extract, reuse, verify, improve

1. Get one example right

Work through one meeting with AI. Check names, decisions, owners, and deadlines against the notes. Keep the accepted output and the mistakes you corrected. If “proposed a trial” became “approved the rollout,” the guide needs a rule about decision status.

2. Extract the reusable parts

Record minimum inputs, steps, output structure, and review criteria. Supply project names, participants, and actual dates with each new task.

For this example, require source references for decisions and follow-ups. Missing owners or dates should say “not specified,” with questions collected for confirmation.

3. Try a fresh example

Use notes that were not part of the original example. Include an unassigned task, an unresolved proposal, or a deadline changed during discussion.

Ask a compatible system to use the skill and check that it actually applied the instructions. Discovery and installation depend on the client; placing a file in an arbitrary folder may not make it available.

4. Verify against the source

A complete set of headings checks structure. Reviewing each decision and task against the notes checks meaning. Another AI can flag suspicious claims, but a “passed” verdict needs supporting evidence that a person can inspect.

5. Improve from observed mistakes

If the agent still invents owners, add an example with no owner. If several guides cover the same task, clarify their boundaries or combine them before adding another.

Keep a short change record: what changed, which failure prompted it, and whether the original case now passes. That also makes obsolete rules easier to remove.

An example SKILL.md for meeting follow-ups

This illustrative guide has not been validated in a particular client. The format specification requires YAML metadata followed by Markdown instructions; the folder name matches name.

meeting-follow-up/
└── SKILL.md
---
name: meeting-follow-up
description: Turn meeting notes into decisions, follow-ups, and open questions. Use when notes or a transcript are available and the user needs a draft for team review.
---

# Meeting follow-up draft

## Inputs
- Notes or transcript with an identifiable source
- Meeting date, if available
- Requested output format, if any

## Method
1. Read the full notes. Identify gaps and contradictions.
2. Separate discussion, proposals, decisions, and open questions.
3. Extract follow-ups with explicitly stated owners and deadlines.
4. Reference a timestamp, paragraph, or source heading for each
   decision and follow-up.
5. Compare the draft with the source before returning it.

## Missing information
- Write “not specified” for missing owners or dates.
- Show contradictions and questions; do not resolve them by guessing.
- If source notes are absent, request them before summarizing.

## Output
- Key points
- Decisions confirmed in the notes
- Follow-ups: task / owner / deadline / source
- Open questions and input limitations

## Acceptance criteria
- No proposal has been promoted to a decision.
- No unsupported names, dates, or commitments have been added.
- Every decision and follow-up has a checkable source reference.
- Return a review draft; do not send messages or create external tasks.

Suppose the notes say, “The team proposed trying the new system next month, but has not chosen an owner.” The expected result is a proposal, with questions about ownership and a start date. It should not claim an approved rollout.

That case tests meaning even when the writing looks polished.

Give instructions, tools, and checks distinct jobs

The skill explains how to reason about the task. A script can check fixed conditions, such as whether every follow-up row has a source field. Some clients provide hooks to run checks at particular stages; availability and configuration depend on the tool.

A populated source field still does not establish that the cited passage supports the claim. Choose checks that match the failure you want to catch.

Begin with a short, usable guide. Add a tool when a repeated step becomes clear, and document its inputs, outputs, and limits.

Check whether it helps

Compare several similar tasks using total time to an accepted result: input preparation, reading instructions, tool execution, human review, and revisions. Record significant errors alongside time.

For meeting summaries, track correction time, misinterpreted decisions, and follow-ups that cannot be traced to the source. Keep the model, requested format, and review criteria comparable where possible. Include the cost of building and maintaining the guide.

One or two examples can reveal improvements without establishing a general productivity gain. Look for less repeated explanation, fewer errors, and lower total time to usable work.

Start small and keep the guide current

Choose one recurring task and one guide. Test it before expanding. If you cannot explain the acceptance criteria yet, review more examples first. Update the guide when the team’s process changes.

Read third-party instructions and bundled scripts before adopting them, including what they do and which data they use. Keep passwords and confidential meeting details out of examples you share.

To begin, take one recently accepted piece of work and write down the required inputs, the repeatable steps, and the evidence of completion. Those are the raw materials for a skill grounded in your own work.

Sources and editorial scope

The five-step workflow, meeting example, and evaluation approach are the author’s proposed application. They are not experimental results or performance guarantees from the video.

Related: Use AI Well: Understand the Problem and Work in Feedback Loops and The evidence ladder.

← All notes