Best AI Workflow Rules for Knowledge Workers Who Use ChatGPT Daily

Short answer

The best AI workflow rule is: do one human rep before every prompt. Write three bullets, recall the source, state the decision, or name the objection before AI enters. That one delay turns AI from a replacement first draft into a sparring partner.

Here are the practical rules.

1. Write three bad bullets first

Before asking ChatGPT for a plan, summary, email, pitch, lesson, or strategy, write three bullets yourself. They can be bad. The point is not quality. The point is ownership.

Bad bullets create a human anchor. Without that anchor, the model’s first answer can become your frame.

2. Ask for critique before rewrite

Do not start with “rewrite this.” Start with:

Find the weakest claim, the missing evidence, and the sentence that sounds least like me. Do not rewrite yet.

This keeps AI in the critic role. The rewrite role is seductive because it feels useful immediately, but it can also erase voice and reasoning.

3. Use AI after recall, not before recall

If you read a paper, article, or meeting transcript, close it and write what you remember before asking AI to summarize. This protects memory encoding.

The question is not whether the AI summary is good. The question is whether you still own the material after reading it.

4. Separate decision from justification

When you have to make a decision, state your decision first. Then ask AI to challenge it.

Bad prompt:

What should I do?

Better prompt:

My current decision is X. The strongest reasons are A and B. Attack this decision and show what would change your mind.

The second prompt preserves agency.

5. Force source checks on factual claims

For claims that matter, AI should not be the final source. Use it to find candidate sources, then verify against the primary source. The research page is built this way: short claims tied back to the source that carries them.

6. Keep one task human-only each day

Choose one meaningful task per day where AI is not allowed until after the first complete pass. It can be a memo, code review, lesson plan, outline, or decision note.

One human-only task per day keeps the circuit loaded.

7. Use model roles sparingly

“Act as…” prompts can help, but they can also create borrowed authority. The role that matters most is not “CEO,” “professor,” or “therapist.” It is “adversarial reviewer with source discipline.”

Make the model disagree.

8. End with a teach-back

After AI helps, explain the final answer in your own words without looking. If you cannot teach it back, you probably outsourced too much.

The teach-back is the cheapest test of whether the work still lives in you.

The operating system

Use this five-step AI workflow:

1. Human draft.
2. AI critique.
3. Human revision.
4. Source check.
5. Human teach-back.

That is not anti-AI. That is anti-atrophy.