Chapter 3 · Lesson 3

Know When to Stop Prompting

Recognise when the problem is missing evidence, unsuitable tools, or a decision that belongs to a person.

7 min · Reviewed July 2026

Scenario

You have rewritten the prompt six times, but the answer still lacks the certainty and detail you need.

Prompting cannot create missing evidence

A clearer instruction can improve how available information is used, but it cannot produce a fact that is absent from the sources. If the answer requires current data, internal records, specialist knowledge, or a primary source, obtain that input first.

The tool may not fit the job

A general chat model is not automatically the right system for calculation, database retrieval, legal review, image editing, or current research. Switch to a suitable tool or conventional method when the task requires capabilities the chat does not have.

Some uncertainty belongs in the decision

High-stakes choices often remain uncertain even after good analysis. Do not keep prompting until the model produces artificial confidence. Record assumptions, consult the right people, and make the decision with appropriate accountability.

Build the mental model

A stop rule prevents prompt iteration from hiding missing evidence, access, or accountability.

Classify the failure before retrying

Ask whether the issue is instruction, input, capability, or judgment. Only an instruction failure is primarily solved by another prompt.

Set an escalation condition

Decide in advance when to fetch a source, switch tools, ask a specialist, or make the human decision. This prevents endless low-value regeneration.

Worked example

A model cannot provide a reliable current sales forecast.

Weak approach

Think harder and give me a more accurate forecast. Be certain.

Improved approach

Classify what prevents a reliable forecast: missing data, unavailable tool access, unsuitable method, or ambiguous decision criteria. List the minimum evidence or capability needed for each gap. Stop without forecasting if current pipeline data is unavailable, and route the task to the approved forecasting system.

Why it works

The improved request diagnoses the blockage and includes a stop condition instead of pressuring the model to manufacture confidence.

The four-way stop check

Instruction: the model misunderstood an available task. Input: required evidence is absent. Capability: the tool cannot perform or access the work. Judgment: the decision needs accountable human authority.

Record which class caused the stop. Repeated stops of the same type reveal a process problem, such as missing data access or a task that should never have been assigned to chat.

Apply it to your work

Write a stop rule for one high-friction task

Name the evidence, capability, and approval the task requires. Define what happens when any one is missing.

Starter templateRetry only if the failure is [...]. Stop when [...]. Then obtain/switch/escalate to [...].

spot problem

Which situation is least likely to be solved by improving the prompt again?

Find the choice that creates the clearest risk or quality problem.

Takeaway

Stop prompting when the missing ingredient is evidence, capability, or accountable human judgment.