Assistant Misstatement Repair
The instinct is to report the answer to the model vendor and wait. That rarely works alone, because most false claims are faithful reproductions of something still published: your own retired pricing page, a review of a version from two years ago, an old PDF, a press release nobody removed. While the source stays live the answer regenerates, and reporting a symptom whose cause is still on your own domain spends the one channel you have.
FORMAT
assistant-misstatement-repair.md
WHEN TO REACH FOR THIS
Use when an assistant repeatedly states something false about your product, such as a feature you never shipped, a retired plan or a price that changed.
The skill file
assistant-misstatement-repair.md
---
name: assistant-misstatement-repair
description: Use when an assistant repeatedly states something false about your product, such as a feature you never shipped, a retired plan or a price that changed.
---
# Assistant Misstatement Repair
The instinct is to report the answer to the model vendor and wait. That rarely works alone, because most false claims are faithful reproductions of something still published: your own retired pricing page, a review of a version from two years ago, an old PDF, a press release nobody removed. While the source stays live the answer regenerates, and reporting a symptom whose cause is still on your own domain spends the one channel you have.
## What you need first
- The exact prompt that produces the claim, with the assistant and date, since answers vary between runs
- Any URLs the answer cites, plus the answer text verbatim
- A dated record of what actually changed and when: pricing history, removed features, renamed plans
- Access to your own archived and unlinked pages, including PDFs, changelogs and old campaign landing pages
## Method
1. Reproduce the claim three times in clean sessions before doing any work. A single odd answer is not a pattern and does not justify the effort that follows.
2. Follow the citations the answer gives, but treat the stated source as a lead rather than a fact. A model can attribute a claim to a page that does not contain it.
3. Search your own estate for the wording before looking outward. Retired pricing pages, stale help articles, old PDFs and press releases are the most common cause and the only one you can fix outright.
4. Correct the source itself rather than publishing a rebuttal elsewhere. A new page saying the claim is false leaves two contradictory documents in the corpus, and the older one usually carries more links.
5. Remove or redirect obsolete pages instead of leaving them live under a correction banner, since the banner is prose that extraction may simply not carry.
6. For third-party sources, request a correction with the evidence attached, and accept that a review site may only annotate rather than rewrite.
7. Use the vendor feedback control as well, but after the source is fixed, so that a re-crawl finds a page that agrees with your complaint.
8. Re-test the same prompt monthly and record the date the answer changes. Correction lag runs from weeks to months, and without a dated log nobody can say whether the fix worked or the model simply moved on.
## What this produces
A short correction record per false claim holding the prompt, the traced source, the fix applied, the date, and monthly re-test results until the answer changes.
## Where this goes wrong
- Reporting the answer to the vendor while the page that caused it is still live, so the claim returns after the next crawl
- Publishing a correction page and leaving the wrong page indexed, which gives the corpus two answers and no basis to choose
- Judging the fix on one re-run, when answers vary between runs and a single clean result proves nothing
- Assuming the source must be external when it is most often your own retired content
---
From the QuQi skill library - https://www.quqi.io/skills/assistant-misstatement-repair
Free to download · no account, no email
What you need first
-
The exact prompt that produces the claim, with the assistant and date, since answers vary between runs
-
Any URLs the answer cites, plus the answer text verbatim
-
A dated record of what actually changed and when: pricing history, removed features, renamed plans
-
Access to your own archived and unlinked pages, including PDFs, changelogs and old campaign landing pages
Method
-
01
Reproduce the claim three times in clean sessions before doing any work. A single odd answer is not a pattern and does not justify the effort that follows.
-
02
Follow the citations the answer gives, but treat the stated source as a lead rather than a fact. A model can attribute a claim to a page that does not contain it.
-
03
Search your own estate for the wording before looking outward. Retired pricing pages, stale help articles, old PDFs and press releases are the most common cause and the only one you can fix outright.
-
04
Correct the source itself rather than publishing a rebuttal elsewhere. A new page saying the claim is false leaves two contradictory documents in the corpus, and the older one usually carries more links.
-
05
Remove or redirect obsolete pages instead of leaving them live under a correction banner, since the banner is prose that extraction may simply not carry.
-
06
For third-party sources, request a correction with the evidence attached, and accept that a review site may only annotate rather than rewrite.
-
07
Use the vendor feedback control as well, but after the source is fixed, so that a re-crawl finds a page that agrees with your complaint.
-
08
Re-test the same prompt monthly and record the date the answer changes. Correction lag runs from weeks to months, and without a dated log nobody can say whether the fix worked or the model simply moved on.
What this produces
A short correction record per false claim holding the prompt, the traced source, the fix applied, the date, and monthly re-test results until the answer changes.
Where this goes wrong
-
Reporting the answer to the vendor while the page that caused it is still live, so the claim returns after the next crawl
-
Publishing a correction page and leaving the wrong page indexed, which gives the corpus two answers and no basis to choose
-
Judging the fix on one re-run, when answers vary between runs and a single clean result proves nothing
-
Assuming the source must be external when it is most often your own retired content
Use this skill in your own AI
The download is a plain markdown file with the name and trigger in its frontmatter. Where an assistant supports skills it can load itself, that frontmatter is what it reads to decide this one applies.
Claude Code
Save it as ~/.claude/skills/assistant-misstatement-repair/SKILL.md and Claude loads it on its own when what you are doing matches the trigger line. Put it in .claude/skills inside a project instead if the whole team should have it.
Claude
Upload the file in the skills section of your settings. Once it is there it applies itself in any conversation where the trigger fits, so you do not have to remember it exists.
ChatGPT
There is no skills format to install into, so paste the file contents into a Project instruction or a Custom GPT instead. It then applies to every chat in that project rather than only the one you paste it into.
Anything else
Paste the markdown into the chat before your question. It works in any assistant, it just has to be pasted again each time.
Questions about this skill
When do I trace the source rather than report the answer to the vendor?
When an assistant repeatedly states something false about you: a feature you never shipped, a retired plan, an old price. Reporting and waiting is the instinct, and it rarely works alone, because most false claims are faithful readings of something still published. While that source stays live the answer regenerates after the next crawl.
What do I need in hand before starting?
The exact prompt with the assistant and date, the answer text verbatim with any URLs it cites, a dated record of what actually changed and when, and access to your own archived and unlinked pages, including PDFs, changelogs and old campaign landing pages. Without the change history you cannot separate a false claim from one that was true two years ago.
What do I end up with, and which part gets used?
A short correction record per false claim, holding the prompt, the traced source, the fix applied, the date, and monthly re-test results until the answer changes. The dated re-test log is the part that gets used. Correction lag runs from weeks to months, so without it nobody can say whether your fix worked or the model simply moved on.
What ruins this most often?
Publishing a rebuttal instead of correcting the source. The corpus now holds two contradictory documents, and the older one usually carries more links, so the wrong claim keeps winning. Leaving an obsolete page live under a correction banner does the same thing, since a banner is prose that extraction may not carry. Remove the page or redirect it.
More in AI search (GEO)