Becoming an AI Power User

2.3Working With the Model

They follow up until the output matches because the first response is a measurement

Developing

The first response gives evidence of the assistant's interpretation

Picture asking AI to draft a short update to your manager. A clean, confident answer appears, ready to copy. Fluent wording can conceal a misunderstanding of the task. Before using the draft, compare what it says with what you wanted the manager to know or do.

Read the first answer for evidence of the assistant's interpretation. Does it emphasize the right issue? Did it assume a deadline you never supplied? Did it omit the decision you need? A mismatch can help you choose the next instruction, although it does not reveal the model's internal reasoning.

A simple mismatch may take one correction; a difficult task may need a different source, tool, or model. Judge progress by the remaining gap. If the first response already meets the standard, accept it. There is no required number of turns.

Making progress is measured against a direct-status-update purpose and source evidence that a test failed; the correction is to state the failed test and accept when it fits. — Compare the first answer with the task to decide what, if anything, needs changing.

The first response may already meet the requirements

Treat the first response as something to evaluate. For a draft, it may be version zero. For a straightforward calculation or lookup, it may already be sufficient once checked. The habit is to compare the answer with the task before deciding what comes next.

A weak response can help you articulate your standard. If the draft feels wrong but you cannot explain why, ask the assistant to compare it with your goal and identify possible mismatches. Check its diagnosis against the source and your intention. A useful criticism gives the next revision a specific target.

Claude users with older accounts showed more task in Anthropic's March 2026 report. The comparison did not establish why a person continued a conversation, whether the first answer looked satisfactory, or that more turns caused better work. Choose a follow-up because it addresses a need in your task.

Choose a follow-up from what the answer needs

A small vocabulary can help you choose a follow-up. State the change the answer needs. Reframing helps when it answers the wrong question; critique helps identify weaknesses; comparison opens alternatives. The labels below are practical prompts, not a complete taxonomy of how to work with AI.

Not every follow-up is a correction. Some of the best ones are experiments: 'wouldn't it be better if we tried it this way,' or 'show me a version that takes this further than I asked.' Power users iterate to fix a gap, and they also iterate to play, using the model to stretch what they imagined the output could be. Comparing approaches with materially different trade-offs can be playful too: an alternative may suggest a possibility you had not considered. Judge it by what it adds to the task.

The nine follow-ups below are grouped into three families by purpose. Direction covers the question, audience, and choice of approach. Depth covers language and level of detail. Pressure covers checking weaknesses, assumptions, and factual support. These are practical groupings to help you find a suitable request, not stages every answer must pass through.

Direction: change the question, audience, or approach

Use a direction follow-up when the question or audience needs to change. Reframe redirects the emphasis: 'Explain how the delay affects the business decision, using the supplied facts.' Adapt changes the explanation for its reader: 'Rewrite this for the board. Include financial implications supported by the notes, and flag missing figures.' Compare opens alternatives: 'Suggest approaches that differ in cost, effort, or likely outcome, and explain the trade-offs.' Each request gives the assistant a different purpose for the material.

Depth: adjust the language and level of detail

Use a depth follow-up when the language or level of detail does not fit the reader. Simplify asks for familiar language and patient steps. Deepen asks for a worked example or the reasoning behind a claim. Tighten removes repetition. For example: "Show the calculation using the figures in the source. If an input is missing, state what is needed rather than inventing it." Specificity should come from evidence or an explicitly hypothetical example.

Pressure: check weaknesses, assumptions, and factual support

Test a plausible answer against specific criteria and evidence. Critique asks where a draft meets or misses the goal: 'Compare the draft with the stated goal for this audience. Quote any passage that falls short and explain the consequence. Report when a criterion is met.' Challenge assumptions asks: 'Which conclusions depend on facts about my situation that I have not confirmed?' Fact-check asks: 'Which claims do the supplied documents support, contradict, or leave unresolved?' Inspect the support for the critique as carefully as the support for the draft.

Use the table to find a relevant follow-up

Use the table to find a possible follow-up during a task. The families help organize the options. Combine or adapt the prompts when the draft needs something different.

Follow-up	Family	When to use it	Example follow-up
Reframe	Direction	The output answers the wrong question or misses the point	"I want this framed around the business risk, not the technical details"
Adapt	Direction	The output is good but wrong for the audience	"Rewrite this for a board-level audience. Remove unnecessary implementation details, include financial implications supported by the notes, and flag missing figures."
Compare	Direction	You want to see alternatives before committing	"Suggest approaches that differ in cost, effort, or likely outcome, and explain the tradeoffs"
Simplify	Depth	The language or explanation is too difficult for the intended reader	"Explain this for someone with no technical background. Keep the steps needed to understand it."
Deepen	Depth	The output is too surface-level or generic	"Show a worked example using the supplied facts, or clearly label an illustrative example as hypothetical."
Tighten	Depth	The output rambles or repeats itself	"Remove repetition and filler. Preserve the facts, qualifications, examples, and steps the reader needs to understand and use the answer."
Critique	Pressure	The output looks plausible but you want to pressure-test it	"Compare this draft with the goal and source material. Quote any passage that needs improvement and explain why. Report when the draft already meets a criterion."
Challenge assumptions	Pressure	You want to understand what the model is taking for granted	"What are you assuming about my situation that I have not confirmed?"
Fact-check	Pressure	The output makes claims you cannot verify	"Which of these claims can you verify from the documents I provided? Flag anything you are inferring or guessing."

Choose a follow-up that addresses an identified gap. Critique can help when you need to examine a plausible draft; reframe is useful when its emphasis misses the goal; simplify fits an explanation whose language the audience cannot follow. Use the other options when they address a different need. If the answer already works, you may still try a follow-up as an experiment, with the original kept for comparison.

Each follow-up should address what the revision still needs

Choose the next follow-up from the previous response. A sequence is useful when each change informs the next. A clear standard also supports asking the assistant for a complete revision pass. The example below isolates the steps to show what each correction changes.

A follow-up replaces a vague claim with the facts

The request: 'Help me write a quarterly review for my manager. I lead a team of four. We shipped the new onboarding flow. The API integration, the connection to another system, is delayed because two password-reset tests still fail. My manager wants direct language and a next action. Use only these facts.' A plausible weak line is: 'Despite integration challenges, we continue making strong progress.' It sounds positive without saying what needs attention.

Critique: 'Compare that line with the notes. What does my manager still need to know?' A useful answer identifies the omitted test failures and notes that no next action was supplied. It should ask for that action rather than invent one. Check that those criticisms follow from the notes before asking for a revision.

Reframe and tighten: 'State the failed tests directly. Add that I will send the retest results by noon tomorrow. Remove the general reassurance.' A revised line could read: 'The API integration is delayed because two password-reset tests still fail. I will send the retest results by noon tomorrow.' The follow-up supplied a new commitment; the assistant did not invent it.

The change is visible in the sentences. The revision names the problem and the next update. It has improved on those criteria. The turn count establishes nothing by itself, and a later edit that removes the test failures would make the draft worse.

Continue while a specific correction improves the result. Repeated failure after a clear correction suggests that another follow-up may not address the cause. Check whether the assistant has the needed source, whether two instructions conflict, or whether the task needs a different tool.

Know when a fresh start beats one more correction

A fresh conversation can help when obsolete instructions keep affecting the answer. Repeatedly returning to a rejected approach or losing the current goal among corrections are reasons to consider a reset. They are symptoms to investigate, rather than a rule that two failed turns require restarting.

For a reset, preserve the current goal, settled decisions, source locations, and unresolved problem. Ask the assistant to draft that handoff, then check it before giving it to a fresh conversation. The rejected approach should be clearly marked if it remains relevant. A useful summary carries the decisions forward without asking the next session to reconstruct every correction.

Save a recurring correction in the instructions for the next task

The agent saves a correction when it applies to future requests. If a project update needs a next action, the agent adds a requirement to include it when the source provides one and flag it when missing. It tests that instruction on another update and uses the result of your review to refine the saved version.

In an assistant that supports skills, create a skill as soon as the recurring task is clear. Ask it to save the drafting instructions, current requirements, and checks. Give the skill a name you will use to request it and a short description of when it applies. For project updates, include the next-action requirement now. The agent maintains the skill as you develop a useful improvement, requirement, or quality standard, and tests the revised instructions on relevant examples.

Practice the revision loop on something you recently accepted as-is

Evaluate and revise a previous AI output

Claude reviews and revises a previous AI output, then proposes reusable instructions based on the changes that helped. Compare the versions against your goal.

Copy the promptUse the copy button above the box.
Paste it in Claude or CodexUse the chat or project where you want the assistant to do the work.
Answer the questionsGive clear answers so it understands what you want made.
Test it to see if it worksTry the output and refine until it does what you need.

Problem: I want to check whether an AI-generated draft I accepted serves its intended purpose. Intent: I want to preserve what works, repair supported weaknesses, and identify instructions to test again. Ask me for the draft, its original request, intended audience and goal, and the source material needed to check its claims. Use anything I have already supplied. Ask about a missing detail when guessing would change the revision. Work through these checks: 1. Critique: Compare the draft with the goal, audience, and sources. Quote each passage that needs improvement and explain the consequence. State where the draft already meets the criteria. Distinguish an unsupported claim from one contradicted by evidence; do not invent a weakness or replacement fact. 2. Revise: Address supported problems in order of their effect on the task. Preserve sound passages, facts, citations, attribution, and qualifications. Identify unresolved source questions separately. 3. Tighten: Remove repetition or filler only where doing so preserves the explanation, examples, and steps the reader needs. Keep the previous version if a proposed cut would weaken it. If no revision is warranted, return the original with your assessment. Compare the original and final versions against the same criteria. Describe improvements, trade-offs, or an unchanged result without assuming iteration helped. If the comparison suggests a reusable correction, draft a prompt based on that evidence. It should ask for changing task details rather than contain unfilled placeholders. Label it as a candidate to test on another case, and ask me to confirm it before saving any personal instructions or preferences.

References

1 source

1
Economic Index report: Learning curves View source
Anthropic · 2026 · March 24, 2026
Users whose accounts were at least six months old showed more task iteration and less directive use in the sampled Claude.ai conversations.
The observed association supports examining how people collaborate, without making interaction length a measure of expertise.
This is an observational comparison based on account tenure, not measured practice hours or a randomized test of iteration. Cohort differences and which users remain active can affect the result.