Iterate: bad to better

Mastery: your iteration playbook

You iterate well when you diagnose before editing, change one dial at a time, score versions against the job, verify claims, and stop with a human decision.

~14 minMastery check

Before you start

Why this matters

How to use this check

Take one real draft you recently improved with AI—or invent a concrete task such as a status update, study summary, or customer reply. Answer each section with that task in mind. “I would make the prompt better” is not enough. Name the failure, the single change, the comparison criteria, and the stop rule.

This page consolidates the ladder into a reusable playbook. Keep it short enough to use under time pressure.

1Learn the idea

Read

1. Diagnose before you rewrite

[ ] I can name the intended job in one sentence: audience, outcome, and success look.
[ ] I can classify the failure as vague ask, missing context, wrong format, unsupported claim, tone mismatch, or scope creep.
[ ] I avoid changing five instructions at once after a disappointing first draft.
[ ] I check whether the model lacked evidence versus ignored evidence I already supplied.
[ ] I decide whether the next move is prompt revision, more source material, a different tool, or human authorship.

Mastery prompt: Write a one-line diagnosis for a vague email draft that invents a deadline. State the single change you will make first.

Read

2. Run controlled loops

[ ] Each revision changes one primary dial: job, context, constraints, examples, format, or tone.
[ ] I keep earlier successful constraints instead of replacing the whole prompt every round.
[ ] I can show a short trail: Round 1 problem → change → Round 2 result.
[ ] I notice when more rounds are polishing style while facts remain unchecked.
[ ] I can explain why one-change loops make learning transferable across tasks.

Mastery prompt: Take a weak prompt and produce Round 2 and Round 3 that each change only one dial. Label the dial.

Read

3. Give feedback that guides

[ ] My feedback names what to keep, what to remove, and what evidence to use.
[ ] I avoid vague praise such as “make it better” or “be more professional.”
[ ] I separate style preferences from correctness requirements.
[ ] I ask the model to flag unknowns instead of filling gaps.
[ ] I can turn recurring feedback into a reusable constraint or example.

Mastery prompt: Rewrite “make this more professional” into three concrete instructions tied to audience, claims, and length.

Read

4. Compare, score, and verify

[ ] I use a compact rubric with criteria tied to the job, not only “sounds good.”
[ ] I score at least two versions side by side before choosing.
[ ] I independently verify names, dates, numbers, and commitments against sources.
[ ] I treat fluent structure as a presentation choice, not evidence of truth.
[ ] I know which claims require human approval before sending or publishing.

Mastery prompt: Build a five-criteria rubric for your task and score two drafts. Record one claim that still needs source checking.

Read

5. Stop for the right reason

[ ] I have a written “good enough” bar for this task.
[ ] I stop when criteria are met, when missing facts block progress, or when another tool is required.
[ ] I do not keep iterating to avoid reading the output carefully.
[ ] High-stakes actions keep a human decision point even after a strong draft.
[ ] I can explain the difference between stopping early from impatience and stopping because the job is done.

Mastery prompt: Write stop rules for a low-stakes brainstorm and a customer commitment. Make the difference explicit.

Read

Playbook card

Copy and adapt:

Job:
Audience:
Must include:
Must not invent:
Sources I will verify against:
Failure of the current draft:
One change this round:
Rubric criteria:
Stop when:
Human approval needed for:

Use the card at the start of a chat, after Round 1, and before you accept a final version. If the card is empty, you are guessing.

Checking tutor…

Continue learning · glossary & guides