Skill Guide: Turn Repeated Work Into Reusable Skills

Save repeated procedures as Skills

Procedures need a different home than preferences

A memory is best for a stable fact or preference: 'I prefer concise meeting recaps with clear owners.' A is best for a repeatable procedure: 'When I give you a transcript, extract decisions, owners, deadlines, risks, and a follow-up draft, then ask me to verify anything ambiguous.'

Use the simpler container first. If the instruction is a fact, preference, or standing rule, save it as memory or project . If the instruction has steps, inputs, examples, failure modes, or a review checklist, make it a .

A stylized teaching image showing repeated work passing through one-time, preference, project rule, procedure, and approval gates. — Repeated instructions should go to the smallest container that can hold the work safely.

Memory remembers a preference. A repeats a procedure.

The meeting-recap workflow from the previous chapter is a good example. 'I like short bullet-point recaps' is a memory. 'Extract every decision, assign an owner, flag missing deadlines, format as a table, and never invent action items' is a procedure with steps, sequencing, output requirements, and a boundary. The preference goes in memory. The procedure goes in a .

Most one-off prompts do not need to become Skills

Most AI work is genuinely one-off. You ask a question, get an answer, and move on. A handles that fine. The question is when a prompt crosses the line into work you should save.

When a prompt is enough and when you need a Skill

Check	A prompt handles this	This should become a Skill
Frequency	You will do this once or twice	You do this every week or more
Output shape	The output is unique each time	The output follows the same structure every time
Corrections	You have never corrected the AI on this task	You keep correcting the same things
Audience	Only you will see the output	The output goes to clients, colleagues, or the public
Complexity	One step: ask and receive	Multiple steps with sequencing, conditions, or boundaries
Risk	A weak answer is merely annoying	A weak answer creates legal, financial, reputational, or operational risk

If you keep re-explaining the same instruction, something should be saved somewhere. The question is where: memory, project , or a . The comparison table above helps you decide.

A stylized teaching image showing a short preference going into memory and a multi-step procedure going into a Skill. — A preference is one sentence that always applies. A procedure has steps, inputs, and boundaries.

Better memory still needs explicit procedures

Memory and personalization are becoming a bigger part of everyday AI tools. That is where stable facts and preferences belong: 'I like short bullet-point recaps.' 'I work in healthcare.' 'My team uses Slack, not Teams.' The better that layer gets, the less you need to restate basic background.

Procedures are different. 'Extract every decision from the transcript, assign an owner to each one, flag anything without a clear deadline, format the output as a table with columns for decision, owner, deadline, and status, and never invent action items that were not explicitly stated' is too structured to store as a simple preference. It has steps, sequencing, output requirements, and a boundary. A saves that procedure so the AI can load and follow it whenever the task comes up.

The repeated-correction test

Mini-project: audit your last week

Your turn

Decide where repeated work should live

You'll classify each repeated instruction into the container that matches its frequency, complexity, and stakes.

Why this exercise matters

This choice prevents your library from filling with instructions that would be simpler as memory, project , or a one-off .

You’ll leave with

A recommended container for each instruction with the reasoning
An explanation of what makes each instruction too structured or too simple for the adjacent containers
The smallest next action for each, including a reviewable Skill plan when a Skill is the right home

Use the prompt in order

1
Paste the repeated instructions
Replace the bracketed lines with your three real instructions so the assistant judges the actual work instead of inventing examples.
2
Check the recommended home
Review the reason for each recommendation because a global , project rule, memory, and one-off solve different problems.
3
Approve the next action
Accept only the smallest next action that fits so you do not create a before the instruction has enough procedure, examples, and review value.

Prompt details

Each detail will be inferred, or you will be asked to clarify in the chat.

3/3 details

3 details will be inferred, or you will be asked to clarify in the chat

Starter prompt text

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

Show starter prompt

You are a workflow placement advisor. Your job is to look at repeated
AI instructions I have collected and decide which container each one
belongs in, so the right work gets the right level of structure.

Here are the repeated instructions I noticed:
1. Infer from what you know about me or ask if it is unclear
2. Infer from what you know about me or ask if it is unclear
3. Infer from what you know about me or ask if it is unclear

**Containers to choose from** (simplest to most structured):

- **One-off prompt**: keep as-is, no persistence needed
- **Memory**: a stable personal fact or preference the AI should
  recall across conversations
- **Custom instructions**: standing preferences your AI app applies
  to every conversation (tone, format, boundaries)
- **Project instructions**: a project rule file such as CLAUDE.md,
  AGENTS.md, or equivalent, scoped to one workspace
- **Skill**: a reusable instruction package with a trigger, procedure
  steps, expected output, gotchas, and test scenarios

**How to decide:**

For each instruction, evaluate along these dimensions:
- **Frequency**: how often do I repeat this? (daily, weekly, monthly,
  rarely)
- **Complexity**: is it one sentence or does it have multiple steps,
  inputs, and an output shape?
- **Stability**: does this change often or stay consistent?
- **Stakes**: what happens if the AI gets this wrong? (inconvenience
  vs. real cost)
- **Scope**: does this apply globally, to one project, or to one tool?

If any instruction is too vague to classify, interview me with one or
two questions about how I use it in practice before deciding.

**Output for each instruction:**

1. Recommended container and why
2. What makes it too structured or too simple for the next container
   up or down
3. The smallest next action I should take
4. If it should become a Skill, draft a reviewable Skill plan (trigger,
   procedure, output shape, gotchas) before writing any files

What the answer should give you

You should receive three recommendations such as "keep as memory," "put in this project instruction file," or "draft a Skill plan," with one reason and one next action for each.

I reviewed my recent AI sessions for repeated instructions.
I identified at least three repeated patterns.
I wrote each one as a single sentence.
At least one involves a procedure with multiple steps, the kind of thing that would be too long for a single Memory entry.

Six situations where a is the wrong container

Knowing when to hold off is just as important as knowing when to build. Before you create a , check whether the repeated work has enough stable structure to justify one.

A stylized teaching image showing true Skill fit at the center while rare, simple, unstable, sensitive, unverified, and project-only work stays outside. — A earns the center only when the work is stable, repeatable, reviewable, and worth maintaining.

When to hold off on creating a Skill

Situation	Why a Skill is wrong	Where it belongs instead
Rare tasks you do less than monthly	The maintenance cost will exceed the time saved. A you update once and forget is a dead file.	A one-off , or a saved prompt you paste when needed
Short preferences without procedure	A single sentence about how you like things done has no steps, triggers, or tests.	Memory or
Unstable work with no repeatable pattern	If the inputs, steps, or output shape change every time, the will fight the task.	A you adapt in real time, possibly with foundation files for
Sensitive decisions that need real-time evaluation	A can prepare the evidence, but approval decisions that affect health, legal claims, finances, or reputation need a person in the loop.	A that surfaces the review surface, with human approval before any action
Tasks where you cannot cheaply verify the output	If the only way to check the result is to redo the work yourself, automating it creates false confidence.	A plain with careful review, until the verification step itself can be Skilled
Project-specific rules masquerading as global procedures	One client's compliance requirements or one repo's naming conventions will contaminate unrelated work if installed globally.	or a project-scoped

References

2 sources

1
Equipping agents for the real world with Agent Skills
Anthropic · 2025 · Anthropic Engineering Blog
Anthropic's Skill architecture preloads the Skill name and description, then loads the full SKILL.md only when the agent judges it relevant to the current task.
Retrieved May 2026. Skills use progressive disclosure: the AI reads metadata first, then loads full instructions only when relevant. This means the description you write is the single most important factor in whether your Skill gets used.
View source
2
Agent Skills
OpenAI · 2026 · Codex Developer Documentation
Codex reads each Skill's name, description, and path first, then reads the full SKILL.md only when it selects the Skill for the current task.
Retrieved May 2026. Codex follows the same progressive-disclosure pattern: the description decides whether the full instructions ever load. Writing a clear description matters more than perfecting the procedure.
View source