What You Need to Know
An adjustment round means changing a request, running it again and judging the output. Objective 1.3 taught what good feedback sounds like inside a conversation. This objective covers the discipline around that feedback. A systematic round gives you more than a different result. It gives you knowledge that you can reuse tomorrow and explain to a colleague. A systematic round needs three parts. You change one thing at a time. You hold the test input fixed. You judge the result against a criterion that you state before you look.
This lesson uses one made-up example throughout: the finance office of a university. An officer there has Claude draft the fee letter that the office sends to each new student. Students continue to phone the office to ask for the total.
Change One Thing at a Time
One change at a time shows what fixed the output. Suppose the officer changes only the audience line and the letter improves. Then the audience line was probably the problem. Suppose she changes the wording, the model and the attached example at once. The letter improves, and nobody knows which of the three worked. Tomorrow, when she does the same kind of task, she knows no more than before.
One change at a time does not give you a better draft, or a faster one. A bundled change, which changes several things at once, can also give a good draft. What one change gives you is knowledge of which change worked, so you can use it again. Anthropic's Help Center says that an effective prompt contains as much context as possible. About adjustment, the Help Center advises giving Claude feedback, clarifications and rewrite requests. Anthropic's Academy teaches one change at a time for improving a Skill (objective 7.3). A Skill is a set of instructions that tells Claude how to do one kind of task. After each change, run the same prompts again to see whether the change held.
A change that worked applies only to other tasks that fail for the same reason. Naming the reader fixes a letter that was written for the wrong reader. It does not fix a letter that is missing the fees, or a request with too many requirements.
Hold the Test Input Fixed
A new prompt that you try on a new document tells you nothing, because two things changed. A better result may only mean an easier input. When you hold the test input fixed (the same document, report or source file), the comparison between rounds is meaningful. For the officer, the test input is one student's fee record. She uses this record in every round. Choose a typical input, not the hardest or the smallest. What you learn from an unusual case may not apply to your normal work. So the officer uses the record of a student on a standard one-year course, not the record with the most unusual fees.
To run a new round on the same input, edit your first message and send it again. When you are comparing two wordings, do not ask Claude to rewrite its last draft. A rewrite starts from the old draft, so the two rounds no longer start from the same point. When you edit an earlier message, Claude creates a different version of the chat, so you can compare the two answers.
One good result is not proof. Claude can give a different answer each time you send the same request. Before you use a new wording for every future job, run the old wording and the new wording on the same input. Use the new wording only if it is still better.
State the Criterion Before You Look
"Better" is not a criterion. "A student can find the total fee in five seconds" is a criterion. Stating the criterion in advance stops your judgement from drifting towards whatever the new draft happens to do well. That drift is how you accept a round that fixed nothing. A stated criterion also makes other people's reactions usable. A returned draft and a criterion together identify which requirement failed. "The manager didn't like it" identifies nothing.
Feedback Arrives from Readers, Not Just the Text
The results that matter often come from outside the conversation. A colleague says the tone is wrong. A manager returns a draft. A reader asks the same clarifying question three times. These reactions usually show what the output lacks more precisely than you do when you re-read the text yourself. The students who phone about the total are such a result. When you treat these results as inputs to the next round, not as noise, you can iterate on whether the output works. When you ignore them, you iterate on how the output looks.
Know When to Stop Adjusting
Stop adjusting when the draft passes the criterion that you wrote before the round. More rounds without a new criterion let your judgement drift, and they can break parts of the draft that already work.
Also stop when the rounds keep failing. Adjustment assumes that the approach is right and that the execution is wrong. When several well-aimed rounds have not changed the output, that assumption has failed. The task may need different source material, a feature such as web search or Research, or a different setting. The task may be work that you should not give to Claude at all. The third failed adjustment is a signal to return to the cause list in 7.1 and diagnose the cause again. A fourth round of slightly different words tests the same failed assumption. If students still phone after the officer's third round, she stops rewording.
In the exam
The exam has questions on each of the three parts. The exam presents the bundled change as decisiveness.
Where this method comes from
The Help Center's advice ends at giving Claude feedback and asking for rewrites. Anthropic's Academy teaches one change at a time and the same test prompts each round, for improving a Skill. Branching a chat means editing an earlier message so that Claude creates a different version of the chat. It only gives you versions to compare. The criterion that you state before you look is based on the wording of the exam guide's objective and on ordinary professional practice.
Exam Traps
Changing several things at once and calling it decisive
Changing the wording, the model and the example in one round leaves nobody knowing which change worked, so the improvement cannot be attributed or repeated. Change one thing at a time and keep the rest fixed.
A fourth round of rewording after three aimed rounds changed nothing
Several well-aimed rounds that changed nothing mean the approach, not the execution, is wrong. Go back to 7.1's cause list and re-diagnose: different source material, a different feature or setting, or work not to be handed over.
Testing the adjusted prompt on a different document and crediting the prompt
When the input changed too, the comparison is meaningless; the improvement may only mean the new document was easier. Round-to-round learning requires the same test input every time.
Practice Scenario
A training coordinator at a consulting firm drafts the welcome email she sends attendees before each client workshop, and attendees keep saying it is too long and too formal. She has time for one adjustment round this week. Which round is set up to teach her something?
Build Exercise
Run One Honest Adjustment Round
What you'll learn
- How to choose one instance of a task and keep it as a fixed test case
- How to write a pass-or-fail criterion before adjusting
- How to change exactly one variable per round
- How to recognise the third failed round as a signal to re-diagnose
- Take a real recurring task whose last output disappointed you, and choose one instance of it and keep it as your test case
Why: A fixed test input is what makes round-to-round comparison meaningful; without it, a better result may only mean an easier input.
You should see: One named test case: the same document, request and materials you will use for every round.
- Write down the criterion before you start: one sentence describing what the improved output must let its reader do
Why: A criterion stated in advance stops your judgement drifting to whatever the new draft happens to do well.
You should see: A single written sentence with a pass-or-fail reading, not a wish for something better.
- Change exactly one thing about your request: the audience line, the attached material, or the order of requirements, and note which you chose
Why: One changed variable is what lets an improvement be attributed, repeated and explained; a bundle of changes teaches nothing.
You should see: A one-line record of the single change this round makes.
- Re-run the request on the same test case and judge the result only against the written criterion
Why: Judge the round by the criterion; general impressions let a round that fixed nothing get accepted.
You should see: A pass or fail verdict against the criterion, recorded next to the change that produced it.
- If the round fails, run at most two more aimed rounds; if the third fails too, stop adjusting and re-diagnose against 7.1’s cause list
Why: Repeated failed adjustment is evidence that the approach is wrong, and the professional skill is noticing the signal rather than rewording again.
You should see: Either a passing round, or a written note that adjustment has stopped and diagnosis has started again.
Sources
- My prompt isn't giving me a helpful answer — Anthropic
- A statistical approach to model evaluations — Anthropic
- Define success criteria and build evaluations — Anthropic
- What are artifacts and how do I use them? — Anthropic
- Introduction to Claude Cowork: Validating skills for plugins — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic. The Academy course above teaches one change at a time with fixed test prompts. The stated criterion draws on the exam guide's objective and standard practice; the Help Center documents the feedback direction.