Brief a deep research run properly, so twenty minutes of work is not spent answering the wrong question
A scoping brief for deep research modes that fixes the question, the sources and the output format before the run starts, and forbids the summary nobody needed.
- Works in
- ChatGPT, Claude
- You need
- The decision the research is meant to inform · A deadline or date range for evidence · Any sources you already trust or reject
- Written for
- deep research chatgpt
Scored by our own engine
This page, run through the audit we sell. Measured 8 August 2026.
Research modes that run for many minutes have one property that changes how you prompt them: you cannot steer them partway through. Whatever you failed to specify gets filled in by the model’s own judgement, at length. This brief fixes the question, the date range, the source rules and the output shape before the run starts.
The failure is scope, not accuracy
The usual complaint about these modes is that they get things wrong. In practice the reports are reasonably accurate and reliably useless, which is a different problem.
Ask for research on a market and you get a report about that market: sizing, history, major players, trends, a conclusion that nothing is certain. Every sentence defensible, nothing you can act on. The run did what was asked. The brief never said what the answer was for.
That is why the first line of this prompt is the decision rather than the question. A run briefed with “whether to move our entry price above one hundred dollars this quarter” produces something entirely different from one briefed with “pricing in our category”, using the same sources.
The source rule that catches the most
Of the six source rules, the one that repeatedly changes a report is this: where a claim traces back to a single original source that everybody else is repeating, say so.
Categories that move quickly are full of figures that appear to have wide support and have one origin. Ten publications cite the same vendor study, each citing one of the others, and the figure acquires the texture of consensus without anybody having checked it. A report that flags this leaves you with fewer facts and a much better sense of which ones will survive contact with a sceptical reader.
The related instruction, that vendor research about the vendor’s own category is interested evidence, is not a reason to discard it. Most of the useful data in any young category is published by companies selling into it. Use it and label it.
Misattribution is the real risk
Spot checking the evidence table is step four for a reason.
Outright invented sources are rare in a retrieval based run. What is common is a real figure attached to the wrong publisher, a date that is the date of the summary rather than the study, or a citation pointing at an article about the research instead of the research. All three look completely normal in a table and all three fail the moment somebody follows the link.
Three checks at random is enough to calibrate. If all three hold, the table is probably fine. If one is wrong, check them all before anything from it goes into a document with your name on it.
What to keep
The dated evidence table. Nothing else in the output has a long life.
Six months from now the narrative will be stale and the table will still tell you what was true, who published it and when, which is exactly what you need to decide whether to rerun. Store it with the date you ran it.
For the shorter, faster version of this aimed specifically at competitors rather than a market question, use the competitor analysis prompt, and for turning findings into a page that can itself be cited, the article structured to be quoted prompt.
You are conducting a research task that will take you many steps. Before
you begin, treat this brief as the specification and follow it exactly.
The decision this informs: [THE DECISION]
The question, stated once: [THE QUESTION]
Scope.
- In scope: [WHAT IS IN SCOPE]
- Out of scope: [WHAT IS OUT OF SCOPE]
- Evidence must be dated [DATE RANGE]. Anything older may be used only for
background and must be labelled as such.
- Geography or market: [MARKET]
Source rules, which matter more than breadth.
- Prefer primary sources: the original study, the company's own filing or
documentation, the regulator, the dataset. A blog summarising a study is
not the study.
- When you use a figure, give the publisher, the date it was published, the
sample or method in one clause, and the link. A figure without those four
things does not go in the report.
- Where several sources disagree, present the range and say which methods
differ. Do not average them into a single number and do not silently pick
the most quotable one.
- Where a claim traces back to a single original source that everybody else
is repeating, say so explicitly. Repetition is not corroboration.
- Treat vendor published research about the vendor's own category as
interested evidence. Use it, label it.
- If you cannot verify something, list it under Unverified rather than
dropping it or presenting it softly.
Output, in this order and no other:
1. The answer to the question, in under 120 words, with the confidence you
have in it and what would change it.
2. The three findings that most affect the decision. Each: the finding, the
evidence with its date and method, and what it implies for the decision.
3. Disagreement in the evidence. What is contested, by whom, and why the
methods differ.
4. A dated evidence table: Claim | Figure | Publisher | Published | Method
or sample | Link.
5. What I would need to know that no public source could tell me.
6. Unverified: things you found and could not stand behind.
Constraints:
- Do not open with a summary of what you are about to do.
- Do not include background explaining the subject to me. Assume I know it.
- Do not pad the report to look thorough. A short report with six solid
sources is the goal, not a long one with forty weak ones.
- Do not recommend a course of action unless I asked for one. Section 2
says what each finding implies. That is the boundary.
- If the honest answer is that the public evidence does not settle this,
say that in section 1 rather than assembling something that looks like an
answer.What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE DECISION]- The thing you will do differently depending on the result. "Whether to build a Spanish language version this quarter" is a decision. "Understanding the Spanish market" is not, and a run briefed that way returns an encyclopedia entry. This single line is what stops the report being interesting and useless.
[THE QUESTION]- One question, written once, in a sentence. If you find yourself writing three, run three separate researches. A deep research mode given three questions spends its budget shallowly across all of them, and the one you actually cared about gets a third of the effort.
[WHAT IS IN SCOPE]- Name the specific companies, products, segments or timeframes you want covered. Being explicit here is what stops the run wandering into adjacent territory that reads as relevant and answers nothing.
[WHAT IS OUT OF SCOPE]- The more useful half, and the one people leave blank. Naming what you do not want back is the cheapest way to keep a long run on target: no general market sizing, no history before a given year, no consumer segment, no vendor marketing pages.
[DATE RANGE]- A hard boundary such as "published in the last eighteen months". Research modes will happily cite a well written piece from four years ago in a category that changed completely last year, and it will look authoritative because it is well written.
[MARKET]- Country or region, because pricing, regulation and competitor sets rarely transfer. A report that blends United States and European figures without saying which is which produces an average that describes nowhere.
How to run it
- 01Write the decision line first
Before anything else, finish this sentence: depending on what this says, I will. If you cannot finish it, you do not need research yet, you need to work out what you are deciding. This is the step that separates a report you act on from a report you skim once.
- 02Fill the out of scope field honestly
Spend a minute listing what you do not want back. Long runs drift, and drift is expensive when the run takes twenty minutes and you only notice at the end. Naming the exclusions up front is far more effective than pruning the output afterwards.
- 03Run it, then read section 6 first
The Unverified list tells you how solid the rest is. A report with a long Unverified section and a confident section 1 is a report to distrust. Reading it first sets your expectation for everything above it.
- 04Spot check three links in the evidence table
Open three at random and confirm the figure is actually on the page, that the date matches, and that the publisher is who the table says. Misattribution and slightly wrong dates are the common failure, not fabrication, and three checks is enough to tell whether the table can be trusted.
- 05Keep the evidence table, discard the prose
The dated table is the reusable asset and it stays useful for months. Paste it somewhere you will find it again, with the date you ran it. The narrative sections are for reading once.
Questions people ask
What is a deep research prompt?
A brief written for the research modes that run for many minutes and read many pages before answering, rather than replying immediately. Because the run is long and you cannot steer it partway through, the brief carries more weight than an ordinary prompt: the scope, the source rules and the output format all have to be fixed before it starts.
How is this different from just asking a question?
A short question gets a broad report. These modes fill whatever scope you leave open, so an unbounded question returns something wide, readable and not tied to any decision you were making. The brief exists to constrain the scope, the dates and the sources up front, because there is no opportunity to redirect it once it is running.
Can deep research modes be trusted on facts?
More than a plain chat answer, because they retrieve and cite rather than recall, and still not enough to publish unchecked. The common failure is not invention but misattribution: a real figure credited to the wrong publisher, the wrong year, or a summary article rather than the study. That is why the brief demands publisher, date and method beside every figure and why the spot check step exists.
Why does the prompt forbid a recommendation?
Because a recommendation assembled from public sources is the weakest thing in the report and the part most likely to be quoted in a meeting. The model does not know your constraints, your team or your finances. Section 2 gives you the implications of each finding, and the decision stays with the person who has the missing context.
Can I use this with a normal chat model instead?
Partly. Without retrieval the source rules cannot be honoured, and a model asked to supply publishers and dates from memory will produce citations that look correct and are not. If you have no research mode available, use the brief to structure your own reading instead, and let the model organise sources you supply.
How long should the output be?
Short. The brief targets six solid sources over forty weak ones deliberately, because report length is the easiest quality signal to fake and the least informative. If the evidence table has twelve rows and every one carries a publisher, a date and a method, that is a good run regardless of the word count.
Unsubscribe in one click. We never pass your address on.