What to do when ChatGPT does not follow your instructions

Identify what 'not following instructions' means here

When ChatGPT misses your request, the first useful step is to name the mismatch. Did it use the wrong format, invent a fact, ignore the audience, omit a step, or claim an action happened? These problems can look similar in a frustrating answer, but the repairs are different.

'You ignored everything' gives little information to work with. 'I asked for three bullet points, but you returned six paragraphs' names a visible failure. 'The source does not include a deadline, but you added Friday' names a factual failure. A clear correction lets you judge the next answer.

Keep the current task visible. Say what finished result you need, what material controls the facts, and which requirement matters most. You do not need to explain every earlier conversation if one short brief captures the current job.

Avoid assuming the problem is always the prompt. Sometimes the assistant could not read a file, lacks a tool, or is responding to conflicting shared instructions. Diagnose the cause before spending time making the request longer.

Separate a writing mistake from a missing capability

A writing mistake can often be repaired in the next turn. If the answer is too formal, show the phrase you want simplified. If it has the wrong headings, provide the required headings. If it uses unsupported numbers, identify them and point back to the allowed source.

A missing capability needs a different response. ChatGPT cannot analyze a file it did not receive, read a private page without an available authorized path, or schedule something merely by writing a timetable. Ask what source or tool was actually used and what could not be accessed.

For current product features, check the official documentation and the controls available in your account. Platform, workspace, and access conditions can matter. Demanding a feature harder does not make it available, and a fluent substitute answer does not mean the requested operation happened.

Use a practical fallback when needed: paste a clean excerpt, upload a readable copy, request a draft instead of a sent message, or perform the task with an available tool. Keep the fallback labeled so you do not mistake it for the original completed action.

Find conflicting instructions and examples

A conversation may contain an old request for a formal tone, a later request for friendly language, and an example that is full of formal phrases. The assistant has several signals to reconcile. Write one current instruction that resolves the conflict: 'For this reply, use friendly plain English. The older formal example is no longer the style guide.'

Check any shared instructions that apply to the chat, such as Project instructions or personal preferences. If a Project says every output must be a table, a request for a short email may keep receiving a table. Remove or narrow the shared rule when it no longer fits the work.

Examples can conflict with your written rules. If you ask for short sentences but provide a sample with long dense paragraphs, say which aspect of the sample to follow. 'Use this sample's warmth, but keep sentences shorter and do not copy its structure' is clearer.

Keep only rules that change the output in a useful way. Long lists of overlapping instructions can make it harder to see which one matters when two requirements cannot both be satisfied.

State a requirement you can actually check

Words such as 'better,' 'professional,' and 'complete' can mean different things to different readers. Add a concrete description. A professional customer reply might mean calm language, a clear next step, and no promise the business has not approved. A complete data summary might mean one row per source record, not a long explanation.

Choose checks tied to the task. For a rewrite, check meaning, tone, and length. For extracted records, check row count, field values, and missing data. For a recommendation, check the stated requirements and the evidence for decisive claims. A generic request to 'double-check everything' is less useful.

A visible rule also helps with revision. 'Use Not set when the notes give no deadline' can be checked row by row. 'Do not make assumptions' is broad enough that the assistant may not know which inference caused the problem.

Do not create unnecessary restrictions. If a useful explanation needs more than 100 words, a strict 100-word limit can cause omissions. Resolve the trade-off explicitly rather than asking for maximum detail and extreme brevity at the same time.

Teaching example: fix invented meeting commitments

Use this fictional note: 'Nora will send the draft on Thursday. We could ask Lee to check the figures. No review deadline agreed.' You ask for an action table using only assigned work. A weak answer lists Nora's draft, then adds 'Lee — check figures — Friday.'

The failure has two parts. Lee was suggested, not assigned, and Friday was not in the source. The repair should remove the invented commitment while preserving Nora's supported action. You can ask for the idea to appear separately under open questions if that helps the meeting lead.

A focused correction is: 'The notes do not assign Lee or give a review deadline. Keep Nora's row. Move the possible figure review to Open questions with owner Not assigned and deadline Not set. Do not add dates.'

The corrected teaching output below follows that rule. It illustrates what to check; it is not a claim that this exact interaction was tested in an account.

Ask for a focused revision and preserve good work

When most of the answer works, name the part that should change. 'Keep the structure and the first two paragraphs. Rewrite the final paragraph in plain language and remove the unapproved refund promise.' This gives the assistant a boundary and reduces the chance of losing good material.

Quote a short wrong passage when necessary. For a long answer, the exact sentence is easier to identify than 'the bit near the end.' Explain why it is wrong and state the replacement rule. A factual correction should point to the fact source or say the fact is missing.

If the format failed, provide a small model of the desired format. For example: 'Return three columns: Task, Owner, Due. One row for each confirmed action. No paragraphs before the table.' Use the sample to show structure, not to smuggle in invented data.

Read the revision against the source again. A correction can fix one mistake while introducing another, such as changing a recorded date or dropping a relevant condition. Preserve meaning as well as appearance.

Break a broad task into useful stages

A broad request can combine reading, deciding, writing, and formatting. If the answer keeps drifting, separate the steps that need different checks. For meeting notes, first extract decisions and assignments; then turn the checked list into a follow-up message. For a resume, first map evidence; then rewrite bullets.

This does not mean every small task needs a long approval process. Use stages when they expose a real source of error. A simple spelling correction can happen directly. A large collection of records may benefit from checking a small sample before processing the rest.

Give each stage a clear result. 'List the fields present in the source and mark missing ones' is a useful first step. 'Think carefully before answering' does not tell you what changed or provide something you can inspect.

When the input is large and structured, consider a script or spreadsheet operation for mechanical work. ChatGPT can help design the rule and inspect exceptions, while deterministic processing handles repeated sorting, counting, or copying. That often makes errors easier to locate.

Start a fresh chat when old context keeps pulling the answer away

A fresh chat can help when the conversation contains many abandoned versions, conflicting instructions, or several unrelated tasks. Bring a compact current brief: the outcome, source, constraints, and one example. Do not paste the whole history that caused the confusion.

Include useful corrections explicitly. If the current fact is that no deadline was agreed, put that in the new brief. If the required tone is simple and friendly, state it. A clean chat should retain the important decisions without carrying forward every old draft.

Remember that shared settings or Project instructions may still apply. If the same conflict follows you into a new chat in the same Project, inspect those shared instructions. Starting over inside the same context does not necessarily remove a durable rule.

Do not reset for every minor miss. A focused correction is usually faster when the task and source are still clear. Choose a fresh start when it reduces confusion, not as a ritual you repeat after every answer.

Check claims about actions and evidence

Some failures are more than style. An answer may say it read a document, ran code, created a reminder, or sent a message. Ask for the evidence appropriate to that action. A summary should refer to the source. A code test needs observed output. A scheduled task needs a saved task and a run when due.

Keep draft, attempted, completed, and verified states separate. 'Here is an email you can send' is a draft. 'The send action succeeded' is an action report. Confirmation that the intended recipient received it is another question. Do not accept the strongest status merely because the wording sounds confident.

If the source could not be read, ask the assistant to remove source-dependent claims and state the limitation. Then provide an accessible source or choose another method. A plausible general answer can still be useful, but it should not be labeled a finding from your document.

For current external facts, use suitable primary sources and open the decisive links. The instruction 'be accurate' cannot replace the evidence needed for a claim that may have changed.

Save a short pattern for the next time

After you get a useful result, save the brief that made the task clear: result, source, limits, format, and check. Keep it small enough to adapt. A reusable prompt should help you give the relevant facts, not contain a growing list of every mistake from every old chat.

You can also keep a simple repair pattern: 'The missed rule is X. The wrong part is Y. Replace it using Z. Keep A and B.' This works for many writing, extraction, and formatting tasks because it names the failure and protects what already works.

If the same issue recurs, look upstream. Missing facts may need a better source note. Repeated tone problems may need a clearer sample. Wrong assignments may need a rule distinguishing proposals from commitments. Fix the cause once when possible.

The aim is a result that passes the task's real checks, not a conversation where the assistant agrees that it should do better. End when the answer is useful, accurate enough for its purpose, and honest about anything still unverified.

Keep the useful material while correcting the missed requirement. Repair only the failed part →