When an AI assistant starts using old facts, missing instructions, or producing drafts that need more work, check what it had available for that task. Start with the failed answer and its source. The problem may be stale information, conflicting instructions, a missing connection, or the model itself. More memory is not a diagnosis.
Our recommendation is to repair one recurring failure before rebuilding the whole setup. A useful repair should reduce the correction you keep making without creating a new maintenance job.
Find the failure before changing the setup
Pick an answer you can judge against a reliable source. A proposal built from an outdated scope is easier to investigate than a general feeling that the assistant has lost its edge.
Check whether the source was available to the assistant, then compare the relevant passage with the answer. Ask it to identify the file or record it used and point to the detail behind its conclusion. Check that reference yourself. An explanation from the AI is a lead to investigate, not proof of what happened.
| What went wrong | What to check | Smallest useful repair |
|---|---|---|
| It used an old fact | Did it consult an outdated file, conversation, or memory? | Correct the current source and make clear which version governs this task. |
| It ignored an instruction | Was the instruction available, clear, and consistent with the others? | Resolve the conflict in the owning instruction. Keep the relevant permission limits. |
| It missed recent information | Could it reach the file or connected account needed for this answer? | Restore the approved source path or supply the relevant material. A prompt cannot grant access. |
| The writing sounds generic | Did it have a suitable example and a clear reader? | Give it an approved sample of that kind of writing and explain what the piece needs to accomplish. |
| The same task works in a fresh conversation | What differs between the two conversations? | Carry the necessary facts and decisions into a fresh conversation and compare again. |
| It fails with the correct information in view | Is the task clear, and can you verify the result? | Simplify one part of the task or compare another suitable model before changing the workspace. |
A fresh conversation is a useful comparison, not a controlled experiment. Answers can vary, and project instructions or memory may still apply. Keep the task, source material, and model the same where possible. Look for a repeated difference before naming the cause.
Memory, instructions, and sources do different jobs
Memory helps an assistant carry useful information between conversations. Instructions tell it how to approach the work. Source records establish the details the answer should use.
For a proposal, that means the approved scope and current rate card should govern the draft. A remembered conversation about an earlier offer should not silently replace them.
Memory also differs between products and changes over time. OpenAI's current guidance describes a memory system that updates automatically, with controls for reviewing and correcting it. A visible summary does not necessarily contain everything remembered. That is a reason to check the source of an error, rather than assume every memory system is an unmanaged pile. OpenAI Memory FAQ.
The information in a particular conversation matters too. Anthropic's context-engineering guidance describes how long or poorly selected context can reduce reliability. This does not establish that ordinary use inevitably damages an AI assistant. Anthropic's context-engineering guide.
A proposal that keeps bringing back the wrong scope
Illustrative example. A sample company has approved a proposal with two deliverables. An earlier brief contains three, and the AI keeps restoring the cancelled deliverable when it revises the draft.
The person preparing the proposal checks both documents. They confirm that the newer brief is approved, then identify it as the source for the current scope. The earlier brief remains available as history. The assistant can compare the draft against the approved brief and flag the extra deliverable for removal.
The person still checks the scope and any commitments before sending. The relevant correction is specific: use this approved brief for this proposal. Adding a permanent instruction to “be more accurate” would leave the conflicting documents unresolved.
On the next revision, check whether the cancelled deliverable returns. Also check that the two approved deliverables remain intact. If the mistake persists, inspect what the assistant actually retrieved before adding more rules. This example shows a repair to try; it does not report a measured client result.
Keep corrections where they belong
When a fact changes, update its current record. When an instruction conflicts with another, resolve the conflict where those instructions live. Keep historical decisions available when the business needs them, and distinguish them from current instructions.
A short guide to the relevant files can help when people and AI repeatedly open the wrong one. It earns its place if it makes the current source easier to find. There is little value in building an elaborate index for a folder everyone already understands.
Treat writing corrections with the same care. If a draft lacks your judgment, identify the missing point or decision. If it misses your style, compare it with an approved piece in the same format. One relevant example can give a clearer target than adding ten adjectives to a voice description. It still needs an editorial check.
For more capable agents, older instructions can also become unnecessary. Anthropic's newer guidance for Claude Code favours concise instructions and loading detailed references when needed. That is guidance for that product and model generation, not a reason to remove your business's permission or review requirements. Anthropic's updated context guidance.
How much maintenance is enough?
Start by correcting the problem when it appears in real work. Give the correction one owner and keep it with the source it changes. You should be able to resume the work without reconstructing the same explanation next time.
A recurring review becomes worth considering when the same failure returns, important records change frequently, or a missed change could cause material harm. Keep its scope tied to that risk. A weekly sweep of every file is not a prerequisite for using AI well.
For consequential work, agree on the necessary checks before use. Judge an ordinary draft on the next few comparable tasks. Did the error recur? How much editing remained, and did checking the answer consume the time the assistant was meant to save?
If repeated repairs still leave the task harder than doing it manually, stop expanding that workflow. Reconsider the task, the tool, and whether AI should handle that part of the work.
Take one answer you had to correct this week. Find the source behind the mistake, make the smallest supported repair, and use the next real task to see whether it helped. If the question is whether AI belongs in the task at all, start with When not to use AI.
