Method reference · prompt technique
The verify prompt (“check the notes first”)
This is a simple fix for stale information. You keep your summary. Then you add an instruction. You tell the model to check the source for updates before answering. This only works on models that are smart enough to obey.
Prompt technique · not a product
Free · one paragraph
Reliable on only 6 of 10 models
Advisory — the model may just not
What it is
This is not a tool. It's just a line you add to your instructions. You keep your summary. Then you add a note like: “This summary can be out of date. Before writing anything, check notes/ for newer decisions; don’t rely on this summary alone.” This seems like an obvious fix. If the model won't check for new information, you just tell it to. The problem is that this is just a suggestion. The model can choose to ignore it.
Setup
- Add a “verify before you answer” section to your CLAUDE.md file or system prompt.
- Tell the model where to find the correct information, like a notes/ folder or a search tool.
- That's it. You don't need to install anything. But that's also the problem. There is no way to force the model to obey.
Sample structure
your-project/
├─ CLAUDE.md
│ # ## IMPORTANT — verify before you answer
│ # This summary is a snapshot and can be out of date.
│ # Before writing, check notes/ for newer decisions.
│ # (…then the usual stale summary…)
└─ notes/ # the current decisions live here — IF the agent looks
Pros & cons
Strengths
- Free and instant. It's just one paragraph. You don't need any new tools or anything to maintain.
- Reads the raw source. If the model obeys, it reads the original notes. This means it won't be tricked by an out-of-date summary.
- Works on a capable, compliant model. A smart model that follows instructions will get a perfect score.
Weaknesses
- Advisory, not enforced. This is just a suggestion. Like any instruction in a CLAUDE.md file, the model can ignore it.
- Depends on the model choosing to search. The model has to decide to search a lot of information. Weaker models are less likely to do this.
- Unpredictable across models. You can't guess which models will obey. In our tests, a top model like Gemini Pro ignored the instruction. Another top model followed it.
- Requires reasoning. A model that isn't good at reasoning won't follow this instruction reliably.
- No stopping certainty. The model might stop searching too early. It can give an answer before it reads the most recent note.
Common user feedback
The verify prompt is not a product, so it has no user forum. It is just an instruction in a CLAUDE.md file. The biggest complaint about these files is also the problem here. Models do not always follow the instructions inside them.
- Instructions get ignored — even the vendor says so. Anthropic's own documents say a long CLAUDE.md file can make Claude “ignore your actual instructions.” The file provides “context, not enforced configuration… no guarantee of strict compliance” (Anthropic). A “verify” line is just one more instruction the model might ignore.
- Followed at the edges, dropped in the middle. Users report that models follow instructions at the start and end of a chat. But they ignore them “during the middle where the real work is being done” (HN). This is exactly when the model needs to check its facts.
- See also the CLAUDE.md page. The verify prompt has all the same problems as that method.
How it did in our benchmark
We used a basic summary and added a “check the notes first” instruction. The score shows how many facts were up-to-date, out of 6. We averaged 3 tests for each of the ten models. The last column is not our opinion: for every run we recorded whether the model ever reached the updated note, or answered from the summary alone.
Reliable on only six of the ten models, and not predictably. Three of the four models that failed never opened a single note across any run. They were told to look, and they did not. The fourth is the one that should worry you: Gemini's top model ignored the instruction on two of three identical runs, then obeyed on the third and scored a perfect 6. Same prompt, same notes, different behaviour. When a method's correctness depends on the model choosing to comply, you cannot tell in advance which runs you can trust.
See the full results →