Copied to clipboard!
Module 3 Level 3: AI Design Engineering & Productization advanced 30 min

Building Design Skill Agents

Package the repeatable. Stay the art director.

Scaffold a Design Skill Agent (SKILL.md)

You do the same design task over and over: the same kind of brief, the same brand checks, the same export format. Instead of re-explaining it every time, you package that repeatable behavior as a SKILL.md, a small instruction file Claude can load as a skill and run the same way every time. AI helps you write and pressure-test the skill. You stay the art director: the skill only ever produces a draft you review and own.

Recommended tool

claude

You upload a SKILL.md to the Claude app and switch it on as a skill, so you build and test the skill in the same place it will run. Strong at decomposing a fuzzy "thing I always do" into clear, repeatable steps and at writing the instruction file. Keep redacted examples only; the skill is a reusable behavior, not a place to store client data.

Bring to the exercise

  • One repeatable design task you do often (e.g. social set from a campaign, brand-checked first drafts, asset-spec audits)
  • Your brand and quality standards for that task, in your own words
  • Two or three real examples of the task done right, redacted
  • Your Product Brief (the Context Document from this level)
Step 1

Pick the Task and Name What "Good" Means

Choose one repeatable task and pin down the standard before writing any instructions.

What to substitute before pasting

  • [PASTE YOUR PRODUCT BRIEF] The Product Brief Context Document you build in this level.
  • [DESCRIBE THE REPEATABLE TASK] The recurring design task you want to automate the boring parts of, in plain words.
Here is my Product Brief:
[PASTE YOUR PRODUCT BRIEF]

I want to package one repeatable design task as a skill agent. The task is:
[DESCRIBE THE REPEATABLE TASK]

Before we write anything, interview me:
- What does a great result on this task look like, specifically?
- What are the brand and quality checks that must pass every time?
- Where does my judgment have to stay in the loop (what should the agent never decide alone)?

Play my answers back as a short "definition of done" for this task. Do not write the skill yet.
What good output looks like
  • Interviews you instead of guessing the standard.
  • Plays back a clear, specific definition of done.
  • Names where your judgment must stay in the loop.

Verify before using AI's output

Step 2

Teach It From Real, Redacted Examples

Extract the repeatable pattern from examples, not from a generic template.

Confidentiality Guardrail: Read Before Pasting
  • Use role labels, not real names (for example: the client, a partner, the team).
  • Strip identifying details, signed dollar values, account numbers, and anything that re-identifies a person or organization.
  • Confirm the AI tool you are pasting into is approved for this kind of information.
  • You, not the AI, own the final product. AI cannot invent facts.

Use the PII Redactor tool to scrub before pasting.

What to substitute before pasting

  • [PASTE 2-3 REDACTED EXAMPLES] Real outputs of this task done well, with client names and confidential details removed.
Here are two or three real examples of this task done right, with client names and confidential details redacted:
[PASTE 2-3 REDACTED EXAMPLES]

Find the repeatable pattern across them: the structure, the moves, the checks, the voice. Do not invent steps that are not in the examples.

Tell me, in a few lines, the pattern you see. We will turn it into the skill's instructions next.
What good output looks like
  • Names the real pattern across your examples, not a generic template.
  • Picks up your structure, checks, and voice.
  • Invents no steps that are not actually in your examples.

Verify before using AI's output

Step 3

Draft the SKILL.md

Write the instruction file with a description, steps, guardrails, and a human checkpoint.

Draft a SKILL.md for this design task that I can upload to Claude as a skill. Include:
- A frontmatter description: what the skill does and when to use it
- Numbered steps that produce the task, following the pattern we found
- The brand and quality checks it must run every time
- A guardrail section: what it must never do, and what data must never go in
- A required HUMAN REVIEW step at the end: the agent produces a draft I approve; it does not ship anything itself

Keep the body tight and plain. This is an instruction file, not an essay.
What good output looks like
  • A real SKILL.md with frontmatter, numbered steps, and guardrails.
  • Bakes in your brand and quality checks.
  • Ends with a required human-review step.

Verify before using AI's output

Step 4

Dry-Run the Agent on a Fresh Case

Run the skill once on a new, redacted case and watch where it drifts.

What to substitute before pasting

  • [DESCRIBE A NEW CASE, REDACTED] A new instance of the task to test the skill on, with confidential details removed.
Run this skill on a fresh, redacted case:
[DESCRIBE A NEW CASE, REDACTED]

Follow the SKILL.md exactly. Then tell me:
- Where did the instructions leave you guessing?
- Where did you almost skip a brand or quality check?
- What one instruction, if tightened, would make the next run cleaner?

Do not "fix" the skill silently. Show me the drift so I can decide what to change.
What good output looks like
  • Follows the SKILL.md and surfaces where it had to guess.
  • Flags any check it nearly skipped.
  • Proposes one concrete tightening instead of silently rewriting.

Verify before using AI's output

Step 5

Finalize and File the Skill (manual)

Walk the manual gate before you trust the agent on real client work.

Walk this checklist by hand before you put this skill into rotation:

[ ] The definition of done matches what I would actually accept.
[ ] The brand and quality checks run on every pass.
[ ] The guardrail names what must never go in (no PII, no unreleased work).
[ ] There is a required human-review step and the agent ships nothing on its own.
[ ] I dry-ran it on a fresh case and tightened the weak instructions.
[ ] I know where this SKILL.md lives and how to update it.

If any box is unchecked, do not trust it on real client work yet.
What good output looks like
  • You can check every box honestly.
  • The human-review step is real and the agent ships nothing alone.

Verify before using AI's output