inference academy ← inference.kg
Course page
Day 04Week 2 · Control week

Day 4 — Skills: teach it once

By now you've typed similar instructions several times. A skill is that instruction saved as a file with a name — the agent applies it on demand, the same way every time. Today you use skills, then write your own.

Today's goal

Understand what a skill is, use one, then write your own skill for a task you personally repeat — and prove it works twice in a row.

You need: your laptop with Codex · one task you do repeatedly (weekly report, meeting notes, standard reply — anything) · attention decay article read tonight, before Day 5.

Block A Using skills What a skill file looks like, how agents summon it, and what separates a good skill from a note nobody follows. We use one on real input before the break.

You've already felt the problem this solves. On Day 2 you wrote “show me your plan and wait for my OK” for the third time. Imagine instead typing /plan-first — and the agent applies the full instruction. That's a skill: a Markdown file with a name, a trigger, and steps. Codex keeps personal ones in a prompts folder and summons them as slash commands; the same idea exists in every serious framework (Claude Code's skills ecosystem made the pattern famous).

A skill file has four parts — click each:

The anatomy of a skill. Click each card to see why the part matters.

1 · Name — what you type to summon it
2 · When to use — one sentence
3 · Steps — what to do, in order
4 · Output example — what “done” looks like
1

Watch a skill work on real input

Here is a complete skill — small enough to read in 20 seconds, useful every week:

Create it and feed it something real:

PROMPT — copy & paste into Codex 
Save the following as a reusable prompt named meeting-notes in my Codex prompts folder (~/.codex/prompts/meeting-notes.md), exactly as written:

[paste the skill from above]

Then confirm where you saved it and how I invoke it.

Now type / in Codex — your skill appears in the menu. Invoke it, paste any messy notes (a Telegram thread works), and compare the output to the skill's example.

If /meeting-notes doesn't appear in the slash menu
  1. Ask Codex: “List the files in my prompts folder and check meeting-notes.md is valid.”
  2. Restart the Codex thread — the menu reads the folder on start.
  3. Still missing? Invoke it manually: “Apply the instructions from ~/.codex/prompts/meeting-notes.md to this: …” — same effect, no menu.

Step done when: the skill produced structured notes matching its own output example.

✦ You explained the task once, and now it's a word. That's the difference between using an agent and equipping one.

Block B Writing your own You pick a task you repeat, draft the skill with the agent's help, then run the honesty test: does it produce the right result twice, on two different inputs, without extra explanations?
2

Pick the task and draft the skill

Good first skills come from weekly annoyances: formatting a report, drafting a standard reply, turning a list into a table, translating your posts with your tone preserved. Let the agent interview you — like the spec on Day 3:

PROMPT — copy & paste into Codex 
I want to create a skill for a task I repeat: [name the task in one sentence].

Interview me — 4 questions, one at a time: what I give it, what I expect back, what mistakes it must never make, and one example of a perfect result.

Then write the skill as ~/.codex/prompts/[short-name].md with: name, "When to use", numbered steps, and an output example. Show me the file.
The same request, written like a pro — open once, the structure is the whole trick
PRO VERSION — Role · Goal · Constraints · Output · Done when 
ROLE: you are my process engineer.
GOAL: turn my repeated task — [task] — into a reusable skill file.
CONSTRAINTS: interview me first (4 questions, one at a time); the finished skill must run with zero mid-run questions.
OUTPUT: ~/.codex/prompts/[short-name].md with: name, "When to use" (one line), numbered steps each ending in a check, one output example.
DONE WHEN: two different real inputs produce correct results without any extra explanation from me.

Same request, five labeled lines. Simple prompts are fine; when the result matters, this structure is why professionals get consistent answers.

Step done when: the skill file exists and reads like instructions you'd hand a new assistant.

3

The honesty test: twice, on different input

Run your skill on a real input. Check the result. Then run it on a different real input — no extra explanations allowed. If run two needed you to clarify something, that clarification belongs inside the skill: tell Codex to add it and test again.

✦ Each correction you fold into the file makes every future run better. You're not prompting anymore — you're maintaining a tool.

Step done when: two different inputs produced correct results with zero mid-run explanations.

What makes a skill reliable — four rules
  1. End every step with a check. “…then confirm every row got an owner” beats “process the rows”. Steps without a finish line make agents stop early.
  2. One rule lives in one place. Write the same rule twice and one copy will rot — then the skill behaves differently on different days.
  3. Delete lines that change nothing. “Be careful” changes nothing. Test each line: would the agent behave differently without it? No → cut it.
  4. Use one strong anchor word, not three fuzzy ones. “Return a verdict: SHIP or FIX” gives the agent something solid to hold onto every run.
Three skills worth stealing today (from our lab's daily toolkit)
  • grilling — “Interview me relentlessly about this plan, one question at a time, with your recommended answer for each, until nothing is vague.” Run it on any plan before building. You'll aim it at your own capstone spec tonight.
  • bug fix — “Act as an elite senior engineer chasing a bug in production: find the root cause; give the cause, the fix plan, and the implementation.”
  • performance fix — “Optimize as an elite performance engineer: measure first, find the bottlenecks, return the numbers before and after.”

Save all three as prompts the same way you saved meeting-notes. Week 3 uses the last two on your capstone.

Homework ≈ 30 minutes + tonight's reading
Tonight

Harden your skill on real work, and read the article that prepares Day 5 — it explains why agents forget, and tomorrow's tools exist because of it.

Time estimate: 30m
  4. Open the article →
Done when:
  • Your skill handled real work without hand-holding.
  • Its text is in Telegram.
  • You can name two symptoms of a decayed session from the article.
Short on time? Post the skill as-is and read the article's “Six Habits” section only — those six habits are tomorrow's vocabulary.
Day complete?

One skill used · one skill written & tested twice · reading done

Tick when the homework above is finished — it syncs to your account.

Self-checkAnswer first, then reveal

Three quick checks

What are the four parts of a skill, and which one replaces the most explanation?

Name · when to use · steps · output example. The output example carries the most weight — one concrete “done looks like this” prevents more misunderstandings than paragraphs of description.

Your skill worked on input A but needed clarification on input B. What's the correct move?

The clarification goes into the skill file, then test again. A skill that needs a human whisper on every second run isn't a skill yet — it's a draft. Corrections compound only if they're written down.

How is a skill different from just keeping a good prompt in your notes?

A skill is installed, named, and versioned where the agent looks — summoned by a word, applied identically every time, improved in one place. Notes rot in a notebook; skills live in the workshop.

What you can do now
  • Read and judge a skill file in seconds.
  • Install and summon skills in Codex.
  • Turn a repeated task into a tested, reusable skill.
  • Improve a skill by folding corrections into the file.
Next session: You install your second agent — Hermes, the one that rewrites its own memory files as it learns you. Tonight's article explains exactly why that ability matters.