Your AI agent finishes a change and reports back: "Done. The form is accessible, responsive and works on mobile."

Did anyone check? Often the agent wrote the code, looked at it, and described what it hoped was true as if it had been tested.

Design Decision Method's last step is there to stop that. Step 6, deliver and validate, asks the agent to hand over the work, the decisions behind it, and a short validation status that says what each result rests on.

What step 6 asks for

The method asks the agent to deliver three things:

  • the requested artifact (spec, mockup, prototype, or code)
  • the consequential decisions and assumptions
  • a short validation status

Before validating, the agent defines a few observable acceptance criteria matched to the purpose of the experience. Then it labels each result by its evidence level and says what it did not test.

Five evidence labels

Each label says what a result can support and what it can't:

Level Can support Cannot support
Proposed A target to aim for Any claim about the result
Heuristic review Likely problems, effort comparisons Task-completion rates
Automated check Specific rule failures Overall accessibility or usability
Manual test Behaviour of a flow, keyboard and focus Population-level outcomes
User research Real task outcomes for the sampled users Claims beyond the sample

Two rules come with it. Never fabricate a measurement. If something wasn't tested, the record says "not tested". And a mockup can show routes and states, but it cannot establish task-completion rates.

A worked example: a handoff note

Say you ask your agent to rebuild the sign-up form on example.com. It's one page with name, email and password, and a clear way to fix errors.

Without the method, the handoff is the sentence at the top of this post. With it, the agent's note might read:

Delivered: Sign-up form (code). One page, three fields.

Decisions: Errors show next to the field and in a summary at the top. Entered values are kept after an error. The placeholder for email is name@example.com.

Assumptions: No social sign-in was requested, so none was added.

Acceptance criteria and status:

  • Valid entries survive a validation error. Manual test: submitted with a bad password, and name and email were still filled in.
  • Each error is announced next to its field. Heuristic review of the markup only. Not tested with a screen reader.
  • Text and background contrast. Automated check: the contrast check found no failures on the colours used. This does not show the whole form is accessible.
  • People complete sign-up without help. Proposed. Not tested. No user research was done.

Not tested: screen reader pass, mobile keyboard behaviour, the server's error responses.

Every line says what it rests on. You can see that mobile keyboard behaviour is the next check, and that nobody has tried the form with real users.

Why this matters when you build with agents

An agent that sounds confident is easy to trust. The evidence labels give you a quick way to read the handoff. "Proposed" means it hasn't been checked. "Automated check" means one rule ran. "Manual test" means somebody actually walked through the flow.

The list of untested items also gives you the next prompt: ask for exactly those checks.

The method doesn't claim this saves time. That hasn't been measured. What it does give you is a handoff note that says where its checks stop.

Keep it proportional

Pick a few criteria that fit the design. The method warns against a fixed score or a universal time or click limit. A copy fix might need one line of status. A new checkout flow needs more. The design record template has a short "Validation status" section for this: the evidence levels reached, what was tested and what wasn't, and open questions.

Where this fits

Deliver and validate is step 6 of six: establish context, choose the visual system, map objects before layout, write a decision record, structure each view, then deliver and validate.

Try it

Design Decision Method is a lightweight design decision workflow for developers building interfaces with AI agents. It follows the open Agent Skills format and is MIT licensed. Install paths for Claude Code, Cursor, Gemini CLI, and Codex are in the README.