What areSkills?
Teach the agent your way of working.
You keep pasting
the same instructions.
“Start with the recommendation. Three reasons. Keep it short. Use our format. Don’t make up numbers.”
What if you could write that once, and have the agent find it when it needs it?
The agent needs
a way of working.
Model — reasons and writes.
Memory — keeps useful context.
Tools — let it act.
Harness — runs the loop.
A skill supplies the method: a format, a rubric, a checklist. The harness makes that know-how available when a task needs it.
A skill is just a folder.
exec-summary/
└── SKILL.md
star-feedback/
├── SKILL.md
└── references/
└── rubric.mdSKILL.md names the skill, describes when to use it, and gives the instructions.
Extra files can hold rubrics, scripts or templates.
Two parts. Different jobs.
--- name: exec-summary description: Turns messy notes, call notes, transcripts or long text into a one-screen executive summary in answer-first style. Use when the user asks to summarise, brief, condense or write an exec summary of meeting notes, articles, case readings or reports. ---
Recommendation first.
Exactly three reasons, each with a source number.
120 words maximum.
Risks, then who does what by when.
It loads in three steps.
The name and description are available up front. A small amount of context for each skill.
When the task matches, the agent reads the full SKILL.md.
It follows the steps and opens extra files only as needed.
Give it a clear reason
to pick your skill.
“Helps with business tasks.”
“Turns messy notes into an executive summary. Use when the user asks to summarise, brief or condense meeting notes or reports.”
Four things. Four jobs.
| Thing | What it supplies | Example |
|---|---|---|
| Prompt | The request you give now | “Summarise these notes.” |
| Memory | Saved facts or preferences | “I’m an IESE MBA student.” |
| Tool | An action the agent can take | Open a file. Search the web. |
| Skill | A reusable way to do a task | Our executive-summary format. |
Which file wakes up when?
| File · job | Who creates it / when | When the agent reads it |
|---|---|---|
| CLAUDE.md · project rules | You, or Claude Code’s /init at setup | Session start; subfolder rules when needed |
| AGENTS.md · project rules | You, your team or a setup tool | Codex: session start, for the working folder’s scope |
| MEMORY.md · learned context | Claude Code, as it learns with memory enabled | Index at startup; topic files on demand |
| SKILL.md · a reusable method | You, a skill creator or an installer | Name + description first; full instructions on use |
| README.md / notes.md · reference | You or the project’s author | When opened or explicitly imported |
Other formats: settings.json / config.toml configure the app; .py / .js scripts run when called; templates, images and CSVs are used when needed.
Open Claude. Get the kit.
Settings → Capabilities → Code execution and file creation.
Then Customize → Skills. Leave exec-summary off for now.
Open the participant kit ↗
Step-by-step: Claude or Codex ↗
- exec-summary.zip
- star-feedback.zip
- test-inputs.md
- scorecard.md
Same model. Same prompt.
Add one skill.
Keep exec-summary off. New chat. Paste the Pedala notes after:
“Summarise these notes for my manager:”
Score the result.
Customize → Skills → + → Create skill → Upload a skill.
Upload exec-summary.zip and enable it. New chat, same model, exact same prompt. Score again.
Don’t just prefer it. Score it.
| One point for each yes | Without | With |
|---|---|---|
| Starts with a recommendation | / 1 | / 1 |
| Exactly 3 reasons, each with a source number | / 1 | / 1 |
| 120 words or fewer | / 1 | / 1 |
| No invented numbers | / 1 | / 1 |
| Ends with who does what by when | / 1 | / 1 |
| Total | / 5 | / 5 |
What changed? What still needs a human check?
Change one rule.
Watch the output change.
Replace “120 words maximum” with “60 words maximum”.
Replace both bold-label requirements with “Use the same four labels in plain text. No bold anywhere.”
Replace the Next step rule with “End with one yes/no decision question.”
Unzip → edit SKILL.md → zip the folder → delete old skill → upload → new chat.
Predict first. Then test.
| New chat for each prompt | Expected | Observed |
|---|---|---|
| “What’s the capital of Peru?” | No | ? |
| “Condense this article so my boss can read it in 30 seconds.” + article | Yes | ? |
| “Write a poem about Barcelona.” | No | ? |
Look for the skill-file read in the work trace when available. Output style alone is not proof.
First the method.
Then the rubric.
star-feedback/
├── SKILL.md
│ “Read references/rubric.md
│ before scoring anything.”
└── references/
└── rubric.mdFour STAR scores, each from 1 to 4.
A weak sentence rewritten.
An improved answer with placeholders for missing facts.
Your weekly task.
Your new skill.
Ask only what you need. Write a clear description of when it should trigger. Create SKILL.md, then package its folder as a ZIP I can download.
Read the description aloud. Would it trigger for the right requests?
Write a skill when
the method repeats.
Instructions you paste again and again.
A team’s shared format or standard.
A rubric or checklist needed for one task.
Name one skill you would write this week.
A tool lets the agent do something. A skill teaches it how.
If it gets stuck,
check one thing at a time.
Check code execution and account or organization access. Pair up if access is blocked.
ZIP the skill’s folder, with SKILL.md inside it. Save edits as plain text.
Check the toggle. Start a new chat. Try explicitly asking to use exec-summary, then review its description.
Keep the files.
Check the product steps.
Participant kit ↗
Executive-summary source ↗
STAR-feedback source ↗
Claude access and upload steps checked 15 October 2026. Recheck before the session.