What to Do When AI Gives You Junk

1. 01

   Diagnose

   Name the failure: wrong task, missing context, generic, too broad

2. 02

   Add context

   Supply the smallest missing piece that would have prevented it

3. 03

   Revise

   Ask for one focused fix, not 'make it better'

4. 04

   Verify

   Check the fix landed without introducing new problems

Bad output is part of the workflow

Every designer hits this moment:

You write a prompt. The AI responds confidently. The result looks okay at first glance. Then you actually read it.

It is generic. It missed the point. It invented constraints. It used the wrong component. It created a polished version of the wrong thing.

That does not mean AI is useless. It means you need a recovery loop.

The recovery loop

When AI gives you junk, do this in order:

1. Name the failure.
2. Decide whether to refine or restart.
3. Add the missing context.
4. Ask for one specific revision.
5. Verify the fix before moving on.

Do not keep pressing regenerate. Random retries teach you nothing.

Step 1: Name the failure

Most bad AI output falls into one of six buckets.

It misunderstood the task

You asked for a component spec and got implementation code. You asked for critique and got encouragement. You asked for a plan and got a finished answer.

Recovery prompt:

You misunderstood the task. I do not want [wrong output]. I want [specific output]. Keep the answer in this format: [format].

It lacked context

The answer ignores your design system, audience, product constraints, or brand direction.

Recovery prompt:

Revise this using the following context:
- Product:
- Audience:
- Design system:
- Constraints:
- What must not change:

It invented facts

It cited components that do not exist, made up token names, or assumed a workflow your team does not use.

Recovery prompt:

Stop and separate known facts from assumptions. List only what you can verify from the files or context I provided. Put assumptions in a separate section and ask me before using them.

It produced generic design

The output looks like every SaaS template: blue CTA, three cards, vague copy, no point of view.

Recovery prompt:

This is too generic. Rework it with a stronger visual and product point of view. Avoid default SaaS structure, vague benefit language, fake metrics, and decorative sections that do not help the user decide.

It changed too much

You asked for a small edit and it rewrote the entire page, renamed props, or moved content around.

Recovery prompt:

The scope is too broad. Keep the existing structure. Only change [specific part]. Do not rename files, props, sections, or tokens.

It solved the wrong problem

The output is competent but aimed at the wrong user goal.

Recovery prompt:

This solves the wrong problem. The user's actual goal is [goal]. Reframe the solution around that goal and remove anything that does not support it.

Step 2: Refine or restart

Use this rule:

Refine when the structure is right and the problem is surface-level.

Examples:

• The hierarchy needs work.
• The copy tone is off.
• The spacing or naming is inconsistent.
• The output needs to follow a stricter format.

Restart when the foundation is wrong.

Examples:

• The user goal is wrong.
• The selected component is wrong.
• The page structure does not match the task.
• The AI invented constraints or ignored the real system.
• You would need five more prompts just to get back to neutral.

Polishing the wrong structure is slower than starting again with a better prompt.

Step 3: Add the missing context

Bad output usually comes from missing context. Before your next prompt, add the smallest piece that would have prevented the mistake.

Useful context types:

• Goal: what the user is trying to do
• Audience: who this is for
• System: which design system, components, and token rules apply
• Constraints: what must not change
• Examples: what good output looks like
• Anti-examples: what to avoid
• Format: how you want the answer structured

If the task touches a design system, include the system rules before asking for output.

Step 4: Ask for one revision

Do not stack six fixes in one prompt. Pick the highest-impact failure and fix that first.

Weak:

Make this better.

Better:

Revise only the empty state. The current version is too generic and does not explain what the user should do next. Keep the component structure, but rewrite the headline, body, and CTA around this user goal: invite a teammate to review the design system audit.

That prompt names the area, the problem, the constraint, and the desired outcome.

Step 5: Verify before accepting

Before you accept the new output, ask:

• Did it fix the failure I named?
• Did it introduce a new problem?
• Did it preserve what I told it to preserve?
• Does the output cite real context, or did it drift again?

If the answer fixed one thing and broke two others, stop. Ask for a smaller edit or restart.

The troubleshooting prompt

Use this whenever you feel stuck:

This output is not usable yet. Help me diagnose why.

Context:
[paste the original task]

Output:
[paste the bad output]

Evaluate it against these failure types:
1. misunderstood task
2. missing context
3. invented facts
4. generic design
5. changed too much
6. solved the wrong problem

For each failure you find, give:
- evidence from the output
- the missing context or instruction
- a revised prompt I should send next

This turns frustration into a debugging session.

Exercise

Recover one bad AI output

15 min

1. Classify the failure

   Find one recent AI output you disliked. Paste it into a note. Label the top failure using the six buckets in this guide. Write one sentence of evidence. For example: “Missing context: it created new token names even though our design system already has semantic color tokens.”

   • You named one primary failure, not ten vague complaints
   • Your evidence points to a specific sentence, component, token, or section
   • You know whether this is a refine problem or restart problem

2. Send a recovery prompt

   Write a recovery prompt that includes the failure, the missing context, and one focused revision request. Send it. Compare the new output to the original failure you named.

   • The new prompt does not say “make it better”
   • The response fixes the specific failure you identified
   • You can explain what context prevented the mistake this time