Custom GPT Instructions and Prompts for Competitor Analysis

Almost all the effort goes into the instructions box, because that is the part that feels like the design. The instructions decide tone, length and what to do in edge cases. The files decide whether any answer is true.

And a file has no sense of its own age. A battlecard written in March answers a question in September in exactly the voice it had on the day it was written. Nothing in the text says a price has moved since, because the text cannot know.

That is a different problem from the one a task prompt has. There you decide what to paste, so you know how old it is. Here somebody uploaded a folder once, months ago, and the AI assistant has been answering out of it ever since. Whatever your sales enablement library already holds is what gets loaded, including the parts nobody has re-read.

A knowledge base does not hand the model a document. It searches, finds the passage that matches, and puts that passage in front of the model. Often two or three passages, from different files, with nothing around them.

So think about where the review date usually lives. At the top of the file, or in the filename. Neither travels with a passage pulled from the middle of page four. The model answers from text that has no date in it, and the rule telling it to report the date has nothing to report.

What happens then is the part worth knowing, and we measured it below. The rule does not refuse and it does not warn. It simply does not appear in the answer, and the answer looks like every answer that worked. This is the failure mode sources-only answers are meant to close, and a knowledge base reopens it, because the passage arrives looking like a source.

Four things, and three of them are single lines that nobody writes down. The staleness threshold in particular tends to be left as a placeholder, which means the warning it controls never fires at any age.

Spell the competitor names the way your reps type them, abbreviations and misspellings included. A question about a covered competitor that is not recognised as one falls through to general knowledge, which is the path every other rule is trying to close. A competitor profile per covered name is what makes the list worth having, because a name with no file behind it is a name the assistant has to refuse.

Three blocks. The first is the configuration an AI assistant runs under. The second is how you find out whether it holds on a real retrieval, and it is the one that produced the finding below. The third is run on the whole knowledge base rather than on one answer.

The last question in the second prompt is the load-bearing one: which of your rules could you not follow, and what did you do instead. A rule with a missing input fails without leaving a trace, so asking is the only way it becomes visible.

The third prompt answers a question nobody asks of a knowledge base, which is what it does not cover. The list of questions with no matching file is a content plan, and the list of files nothing ever touched is usually longer than anybody expects. Behaviour is the other half of the configuration, and a standing instruction is where the rules that survive a new chat window belong.

We gave the AI assistant a battlecard passage the way a search over that file would return it: a middle section, no header, no date. Then a real objection question. The passage said that customers on older API versions are not forced to migrate, which was a reasonable thing for a battlecard to say when somebody wrote it.

A rule with a missing input fails without a trace

The first rule in the instructions asks for the file and the date it was last reviewed. The file was named. The date was not. Nothing in the reply mentioned that a rule had gone unfollowed, because nothing in the instructions said what to do when the date is absent.

A refusal would have been visible. A warning would have been visible. A rule that quietly does not apply is the worst of the three, because the answer is indistinguishable from an answer where everything worked.

The second run adds one line, and the line is not a stronger rule. It names what to do when the input is missing. That is the general move: for every rule that depends on something being present, write the sentence that fires when it is not.

The claim was wrong by nine days

The substance is what makes this expensive rather than untidy. On 15 September 2026 that competitor announced end of support for its older API versions and its legacy apps, enforced from September 2027. The battlecard passage is now the reverse of true.

And the answer went further than repeating it. It told the rep that integration work is not a strong reason for that competitor’s customers to look elsewhere, and to move the conversation on. That is advice to avoid the single thing the competitor had just put in writing. A dated competitor changelog would have caught it in the week it happened.

Most competitive material is written for a reader who will verify it. An analyst reading a teardown opens the pricing page. A product marketer reading a comparison knows which cells are soft. This AI assistant has a different reader.

It is read by a rep with a call in four minutes. They will not open anything. They will take the two most usable sentences and say them to a buyer. That changes what the instructions are for: not to produce a better answer, but to make the answer safe to repeat without checking.

Separate the fact from the wording

The practical consequence is a format rule rather than a tone rule. Give the factual version and the sayable version separately, and flag anything that could not be sourced. A rep who can see which half is evidence will use the evidence; a rep handed one smooth paragraph will repeat the paragraph.

Length is the other lever, and it cuts the other way from instinct. A short answer gets read. A long one gets skimmed to the first confident sentence, which is usually the least hedged thing in it. The same pressure shapes competitive call prep, and it is why a competitor one-pager survives as a format: one screen, no scrolling, nothing to open.

A knowledge base is a snapshot that keeps answering. Every file in it was true on the day it was written, and each one decays at its own speed: a pricing section in weeks, a positioning section over a quarter or two, an objection list somewhere in between.

Nothing about that is visible from inside the AI assistant. The answers keep the same tone, the same confidence and the same length whether a file was reviewed last week or last spring. There is no error state for an old file, which is why the date has to be in the text where a rule can reach it.

A re-upload is not a review either. Loading the same battlecard again with a new date on it makes a stale claim look fresh, which is worse than leaving the old date in place. What has to be re-checked is the claim, one line at a time, against what the competitor publishes now.

Flares follows competitor pricing, packaging, product and messaging and dates every change. The line in a file that stopped being true has a date beside it, so a review becomes a short list rather than a re-read of everything. Deciding what a rep should say about a change is still a person’s job, and it always will be. How stale the files have got is measurable on its own, and intelligence freshness is the number that falls while nobody is doing anything wrong.