Darkmind / Issue 039

Name the Unknowns Before You Start AI Research

Turn a broad request into a short decision frame that separates fixed requirements from open variables, so an AI researches the question you actually need answered instead of a plausible substitute.

Nov 15, 20264 min readResearch workflows

A research agent is very good at making an incomplete question feel complete. Ask it to “find the best project-management tool for our team,” and it can return a convincing comparison of price, features, and reviews. But the recommendation may silently assume the team size, security requirement, existing stack, budget, geography, and definition of “best.” The problem is not that the search failed. It is that the system converted missing decision inputs into invisible assumptions.

The practical move is to write a decision frame before retrieval. It separates what is fixed, what is still unknown, what evidence must establish, and which unknowns are safe to leave open. The researcher receives that frame—not only the original sentence—and must label any unresolved variable in the final answer rather than deciding it on the reader’s behalf.

This matters even when a research model can plan its own searches. OpenAI’s deep-research API example says the model decomposes high-level questions and synthesizes a citation-rich report, but its API path starts researching from the supplied input; developers are responsible for providing scope, comparisons, metrics, regions, preferred sources, and output expectations.[1] Citations can make a claim traceable, but they do not prove that the claim answers the right decision. Anthropic’s citations documentation distinguishes source locations and cited text from the surrounding prompt context, and notes that applications still need to check that required fields have both a value and a citation.[2]

Turn the request into a decision frame

Begin with the action the research will enable. “Compare collaboration tools” is a topic. “Select one tool for a 40-person regulated support team that must keep customer data in the EU” is a decision. The difference tells the researcher which facts could disqualify an option and which details are only nice to know.

Then divide the frame into three parts:

An open variable is not a blank the model should fill. It is a routing decision. If knowing it would change the search corpus, eligibility rules, or recommendation, ask for it before expensive research when an interactive user is available. If work must proceed, choose an explicit conservative default, state it, and return conditional findings rather than a single unqualified recommendation.

For example, “Which AI transcription tool should we buy?” may have open variables including languages, meeting volume, whether recordings contain regulated data, required integrations, and the organization’s approved vendors. A comparison based on public pricing can still be useful, but it cannot settle compliance or procurement fit. The final result should say exactly which conclusion depends on each unresolved input.

Copy this research decision frame

Use this as the handoff between an intake step and a web, document, or internal-knowledge researcher.

DECISION TO SUPPORT
- action after research: [choose | approve | explain | plan]
- decision owner / audience: [who will use the result]
- deadline or AS-OF date: [date]

LOCKED FACTS (do not reinterpret)
- target: [product, market, policy, problem]
- scope: [region, team, tenant, version, time range]
- hard constraints: [budget, security, compatibility, exclusions]
- acceptable source tiers: [primary docs | official filings | research | etc.]

OPEN VARIABLES
For each unknown, record:
- variable: [e.g., required data residency]
- changes research? yes | no
- safe default if proceeding: [explicit value | none]
- final-answer rule: [ask first | provide conditional result | omit recommendation]

EVIDENCE REQUIREMENTS
- R1: [claim the decision needs] — preferred proof: [source type]
- R2: [claim the decision needs] — preferred proof: [source type]
- R3: [claim the decision needs] — preferred proof: [source type]

RESEARCHER OUTPUT
1. Restate locked facts and AS-OF date.
2. List open variables and the treatment used for each.
3. Map every load-bearing claim to a source and locator.
4. Mark findings as confirmed | conditional | not established.
5. Return: recommendation | conditional options | needs input.

The phrase changes research? is the useful test. A preferred visual style may change presentation but not which evidence you need. A required region, data class, or integration can change the candidate set itself. Treat the latter as decision-changing unknowns. Do not hide them in a generic “assumptions” paragraph after the model has already ranked options.

Give unknowns a visible outcome

When an answer must go forward before every variable is resolved, make the output conditional by design. Instead of “Tool A is the best choice,” write: “Tool A is the leading option if EU-only processing and the named CRM integration are confirmed; otherwise the comparison is incomplete.” Put the missing proof beside the conclusion, with a next safe action such as “obtain the vendor’s current data-processing terms” or “confirm the team’s monthly meeting volume.”

This also improves citation review. A citation can support “the vendor documents SSO on its Enterprise plan,” while the decision frame reveals whether Enterprise pricing, the team’s identity provider, and its contract region are still open. The source earns a bounded claim; it does not inherit authority over the whole recommendation.

Failure mode: turning the frame into a questionnaire

A decision frame should reduce waste, not delay every inquiry until a person completes a 30-field form. Ask only about unknowns that change the decision, the evidence source, or the safety of acting. For low-stakes exploration, mark broad defaults and proceed. For a consequential purchase, policy answer, or deployment decision, stop when a missing variable could reverse the result.

The habit to keep

before AI searches, make the missing inputs visible. Freeze the facts you know, name the variables that could change the answer, and let the final recommendation show its conditions instead of disguising them as certainty.

Sources

  1. OpenAI, “Introduction to deep research in the OpenAI API”
  2. Anthropic, “Citations”