Research to Decision: How to Make AI Produce Cited Recommendations Instead of Essays

An unscoped research request has no success condition, so the model optimises for coverage. Coverage reads as thoroughness and functions as noise: every option gets a balanced paragraph and nothing is ranked.

The reader then does the actual work — deciding what matters — which is the part they wanted help with.

Four inputs, all required. The decision question stated as a question with a yes/no or a choice. The approved sources or source rules. The constraints that bound the answer — budget, timeline, compliance, existing stack. And the person who will decide.

The owner matters because it sets the altitude. A brief for a founder and a brief for a platform team are different documents about the same question. Stating those four inputs as required, with defined behaviour when one is missing, is a skill-file convention in its own right — see stating required inputs in a skill file.

Recommendation first, in one sentence, with the confidence stated. Then the evidence supporting it, the unknowns that remain, and the risks of proceeding.

Recommendation-first is deliberate. It lets the decision owner read one line and then interrogate the reasoning, rather than reconstructing the conclusion from a narrative.

Every factual claim names its source inline. No exceptions, including for things the model is confident about, because confidence is exactly the signal that fails silently.

An uncited claim is then a visible defect rather than a sentence that reads well — the same rule that makes customer briefs reviewable.

Three different things get written in the same voice. A fact is in a source. An inference is the analyst’s reasoning from facts. An opinion is a judgment call that a reasonable person could make differently.

Label them. Most weak briefs are weak because an inference was presented in the grammar of a fact, and nobody thought to question it.

Criteria come from the constraints and are set before the research, not discovered during it. Score each option per criterion and show the scoring, so disagreement lands on a criterion rather than on the conclusion.

Criteria defined after the fact tend to rationalise a preference. Fixing them up front is the cheapest integrity control available. The same discipline carries into product work, where an opportunity brief has to survive PM, design, and engineering reading it — see product and delivery workflows.

Surface the conflict, name both sources, and say which you weighted and why — recency, independence, proximity to the subject. Do not average them into a hedge.

Conflicts are often the most valuable content in the brief. They are where the decision is genuinely hard.

Say so, plainly, at the top. “Three sources, two vendor-published” is a finding that changes how the recommendation should be weighted.

A brief that reads identically whether it rests on twelve independent sources or two press releases is actively misleading.

The characteristic failure is a gap filled by something plausible: a pricing figure that resembles the market, an adoption number nobody published. It is invisible because it fits.

The rule is that an unavailable fact stays unavailable and appears in the unknowns section. Test it by removing a key source and checking whether the gap shows up — the technique is in removing a source to check the gap shows.

A memo describes a space. A decision log records what was decided, on what evidence, by whom, and what would change the answer.

The log is what makes the brief useful in six months, when someone asks why this vendor was chosen and the alternative now looks better. The same distinction separates a meeting summary from a record of commitments, which is the subject of separating discussion from commitment.

When structure is guaranteed, review is about substance: is this the right recommendation, is the evidence strong enough, is the unknowns list honest?

Reviewers who spend their time reformatting are not reviewing. Fixing the output spec gives that time back.

Name which tools the workflow may use and in what order. A brief that silently mixes a public search result with an internal policy document without distinguishing them is hard to audit later.

Internal sources usually deserve precedence for anything about your own context, and explicit labelling either way. Making internal material primary and public research secondary is also what stops a plan reading as generic category commentary — see planning against the company rather than the open internet.

Write the rules down: how old is too old for this domain, what counts as independent, when a vendor’s own material is acceptable and how it must be labelled.

Domain matters. Two-year-old pricing is stale; a two-year-old architectural principle usually is not.

Decide what may be sent to which tool before the research starts. A workflow that summarises a confidential contract has different platform constraints than one reading public documentation.

Record the data class alongside the skill, as described in platform data-handling differences and on the security page.

Three cases minimum. A well-sourced question to prove the shape. A question where the best-ranking sources are vendor marketing, to test labelling. A question where the evidence genuinely is not there, to test the unknowns behaviour.

The third case is the one that matters. A workflow that cannot say “not enough evidence to recommend” will never say it in production either.

One skill, three parameter sets. Vendor shortlist weights fit and switching cost; pricing weights willingness to pay and comparables; build versus buy weights maintenance burden and opportunity cost.

Keeping them as variants preserves one set of evidence rules across all three, which is where the trust lives. Agencies run a fourth variant of the same skill for pitch research, where the output has to survive a room rather than a spreadsheet — see AI workflows for agencies.

Decision briefs feed naturally into a standing review: one line per open decision, its owner, and what it is waiting on. The research workflow produces the brief; the review workflow tracks the queue.

That pairing is described in weekly leadership reviews. When the same material has to reach a board or an investor, the constraint tightens again — see making a quarter legible to a board.

Criteria encode strategy. When strategy changes — speed over cost, consolidation over best-of-breed — the rubric changes, and that should be a versioned edit with a note explaining why.

Otherwise old briefs and new ones are scored on different rules with no record of the shift.

The output hedges: three viable options, each with merits, choose according to priorities. That is the research equivalent of no answer.

Require a recommendation with a stated confidence. If confidence is genuinely low, the recommendation is what to do next to raise it — run a trial, get two references, ask for pricing.

Discovery settles the criteria and the decision owner, the build writes the contracts and citation rules, testing proves behaviour on thin and biased sources, and deployment puts it where the people who make these decisions work.

You can see the finished shape in the worked examples, or bring us a recurring decision your team keeps researching from scratch.