Field Guide

Turn a recurring AI mistake into a reusable instruction

Research-based guidance; hypothetical example

You ask an AI assistant for a marketing plan you can finish this week. It suggests a new landing page, paid ads, and a video series. You explain that you have two hours and an existing email list. The next plan has the same problem.

That correction is worth capturing. Write down the behavior you need, where it applies, and how you will check the next result. You can then reuse the instruction on a similar task instead of reconstructing the correction from memory.

This guide offers a method to try. The example is hypothetical, and the method has not yet been validated through repeated use by Today’s Worker. A saved instruction can still be ignored, misunderstood, or applied in the wrong situation.

Start with the mistake you can actually identify

“Give me better answers” leaves too much open. Describe the recurring behavior: recommendations exceed the available budget, summaries introduce unsupported facts, or drafts omit a required section.

Then identify what needs to change. A factual error may need a source check. A missed requirement needs a visible requirement and an acceptance check. A preference, such as the tone used for one client, needs a narrow scope. Optional taste should remain a preference rather than becoming a rule for every task.

Check the inputs, too. If the assistant never received your available hours, supply them. If it cannot access the report, provide an authorized copy or restore access. An instruction to “be accurate” cannot replace the missing source.

Write the behavior you want, including its limits

Use this small correction card:

• Observed problem: What keeps going wrong?
• Desired behavior: What should the assistant do instead?
• Applies when: Which tasks need this rule?
• Check: What would show that the output meets it?
• Exception: When should the rule stay out of the way?

Consider a hypothetical independent marketer with a $150 cash ceiling, two available work hours, and an existing email list. Those numbers are teaching inputs, not a recommended budget or a client result. The recurring problem is that suggested promotions require channels and production capacity the marketer does not have.

The desired behavior is more specific than “be realistic”: recommend an executable action, identify necessary inputs, and make uncertainty visible. A reusable instruction could read:

For short marketing plans with stated resource limits, check each recommendation against the cash ceiling, available work time, and channels already available. Include estimated cash requirements, estimated work time, and required inputs. Label estimates. If a necessary constraint is missing, ask one focused question or state the assumption before recommending an action. Do not invent reach, conversion, or revenue forecasts. If nothing fits, explain what prevents a workable recommendation. This rule does not limit a separately requested long-term strategy or a task with a newly approved budget.

The exception matters. A two-hour execution constraint should not silently restrict a later request to explore next year’s strategy. Keep the instruction attached to the work it was written for.

Add checks you can use on the output

For this example, review four things:

1. Does the proposed action fit the stated cash ceiling and available time?
2. Can it use the stated channels and inputs?
3. Are estimates labeled, with unsupported outcome forecasts omitted?
4. Are missing requirements visible instead of silently filled in?

A confident time estimate does not prove that an action fits. Inspect the work steps: writing, design, setup, approval, and sending may each consume time. If the estimate looks doubtful, revise the scope before treating the recommendation as executable.

Some checks can be mechanical, such as confirming a required section exists or adding listed costs. Others require judgment, such as whether an offer makes sense for the audience. Keep those judgments visible.

Test the instruction on another relevant task

Save the corrected example, then try a comparable request. Include the reusable rule and the actual constraints. For the hypothetical case, a test prompt is:

Suggest one promotion for an existing email list. The cash ceiling is $150 and there are two work hours available. There is no permission to start paid ads. State the inputs needed, likely work steps, and estimated effort. Apply the saved resource-limit rule.

Check the resulting recommendation yourself. Do not count “I followed your instructions” as a pass. Anthropic’s guide to agent evaluations distinguishes the assistant’s record of activity from the actual outcome and discusses variation across attempts. For this small exercise, that means inspecting what the assistant produced.

Record the task, rule version, assistant and date, whether the named mistake recurred, and any new problem. Add the time you spent reviewing and correcting. Where practical, compare similar tasks using your ordinary instructions and the new rule. Differences in task difficulty still matter.

Also try an exception: ask for a six-month strategy with a future budget still undecided. Does the rule allow exploration while keeping assumptions clear? One passing attempt is an observation. It does not establish that the correction will hold across different tasks or tools.

Save the smallest instruction that helps

A short rule can live with the project it belongs to. A repeated procedure with several steps, examples, and resources may warrant a skill. In Anthropic’s explanation of skills, project context and reusable procedures serve different purposes. The available setup depends on your assistant; check its current documentation before installing anything.

Keep the dated test record beside the instruction. If the rule creates more review work than it removes, simplify it. If the same failure persists, investigate the input, source access, task scope, or checking method before adding another paragraph.

That approach follows the broad principle in Anthropic’s Building effective agents: start with a simple solution and add complexity when the results justify it. The source’s tooling guidance is older; this is a design principle, not an installation recipe.

Make a correction you can reuse

Fix It Once helps format your correction into a scoped instruction, checks, and a test prompt. Bring one recurring mistake and the behavior you want instead. Use a sanitized example if you include one.

The next useful milestone is applying that instruction and checking whether the named mistake recurs. Copying it is preparation. Later use supplies the evidence for keeping, changing, or retiring it.

Read our editorial standards for how Today’s Worker distinguishes research-based guidance, hypothetical examples, and tested methods.