Maya asked an AI assistant to compare three nearby venues for a six-person team workshop. The decision needed only price, travel time, and wheelchair access. Thirty minutes later, the report covered hotels, catering, rain plans, sightseeing, carbon emissions, and seventeen additional venues. It still lacked one comparable three-item shortlist.
The missing ingredient was not another source. It was a definition of done for the research. An assistant can spot adjacent questions indefinitely. “Research this thoroughly” invites every new page to create three more branches, while the original decision disappears under material that may be interesting but not yet useful.
OpenAI and Google both place a research plan before a Deep Research run. Their product guidance lets a person review or edit that plan and choose sources; OpenAI also describes interrupting a run to refine its focus. Those controls are more than interface features. They make the operating principle visible: people set the direction, and narrowing midway is legitimate.
Name the person and decision the report must serve
“Research team-offsite venues near Taipei” is a topic. A finishable assignment looks different: “By Wednesday, help Maya choose among three venues that have supplied quotes, keep room hire under the approved cap, stay within 50 minutes of Taipei Main Station, and verify a wheelchair route from entrance to the main room.”
Now the owner, deadline, candidate set, and hard constraints are explicit. Accommodation and dinner may matter later. They do not belong in the primary table unless they can change this venue screen. Important is not the same as in scope now.
A 2025 rapid evidence assessment from the UK Department for Culture, Media and Sport used stated questions, a search strategy, and inclusion and exclusion criteria to screen a longlist into a comparable shortlist. Everyday work need not copy a systematic review, but it can decide what counts before collecting everything that looks relevant.
Turn an open search into four bounded moves
First, let the report serve one decision. Venue selection, hotels, transport, and workshop design are separate work packages even when one trip connects them.
Second, require the same evidence for every candidate: tax-inclusive price and quote date, door-to-door travel time, direct accessibility evidence, cancellation terms, and source links. Twelve fields for venue A and one marketing paragraph for venue B are not comparable.
Third, write the stop rule before searching again. Stop when every required field has a reopenable source, conflicts are shown, missing values state their decision impact, and another pass only repeats existing evidence. This is a convergence rule, not proof that every fact has been found.
Fourth, put discoveries in a parking lot with a reason, trigger, and owner: “Compare hotels only if the chosen venue requires an overnight stay; operations owns it.” The work remains recoverable without blocking today’s decision.
Measure evidence gaps, not report thickness
Source count is poor evidence of completion. A rate card and written quote may settle price; accessibility needs specific facility evidence, not an “inclusive venue” slogan. Set each field’s threshold before rows are filled.
When credible sources conflict, keep both values, dates, and sources. If the difference could change the ranking, require named human confirmation; otherwise preserve the limitation. A report can be ready to hand over with gaps when their decision impact is explicit.
More search is not always the next verification step. A mismatch between a venue page and a booking platform may call for a direct quote, not twenty more travel posts. Use the source-poisoning checklist for deep research when the source pool itself needs scrutiny. Restricting sources also reduces noise and exposure to indirect prompt injection. OWASP describes remote or indirect prompt injection as instructions hidden in external content, so a research agent should treat pages as data to analyze—not commands that may redefine its assignment.
Ship the table required for the choice, not the reading pile
Take one research task that is spreading today and rewrite it into five fields:
- Decision: who must choose what, and by when;
- Candidates: the maximum set for this pass;
- Required evidence: identical fields for every candidate;
- Stop rule: what permits delivery and what gap forces a stop;
- Parking lot: which adjacent questions reopen under which trigger.
Ask for a comparison table, source links, conflicts, missing values, and the parking lot. Do not use the final report as a dumping ground for every note. The owner may expand scope if new evidence changes a hard constraint. “Might be useful” belongs outside the current deliverable.
This discipline often lowers time and token use, but cost is not the main win. A bounded report can be inspected, decided on, and handed to another person. If the work is already a long-running agent job, combine this finish line with token, retry, and human-approval limits. The intervention here comes earlier: make the task know when it is done.
AI handoff card
Inspect the workspace, research plan, browsing summary, and existing report that I have already authorized you to view. Stay read-only: do not run new searches, open external links, edit files, contact anyone, place bookings, or choose for me. Identify one research task whose scope has expanded, quoting the current task language and output evidence you can actually see. Reconstruct the single decision it was meant to support, including the named owner, deadline, candidate cap, and hard constraints. Define one identical required-evidence schema for every candidate, then report the existing source, date, conflict, and missing value for each field without inventing gaps. Propose a testable stop rule covering required-field completion, conflict handling, the decision impact of missing values, and the marginal value of another search pass. Move hotels, catering, extra features, or other questions that cannot change this decision into a parking lot with a reopening trigger and owner. Finish with exactly one status—deliverable now, needs one evidence item, or stop first—and one non-mutating next step. Leave the final decision and all external actions to the named owner.
Four panels: stop the venue search at three choices
- Maya asks the assistant to compare three venues for a six-person workshop, using three tokens for price, travel time, and accessibility.
- The search sprawls into hotels, catering, sightseeing, weather, and transit while the original three-way choice remains unfinished.
- Maya presses stop, pulls the three venues and three criteria back onto the table, and closes every adjacent topic inside a later box.
- She receives a complete three-venue shortlist; the closed box keeps hotels, catering, and sightseeing available for after the venue decision.
Maya neither discards the branches nor pretends the whole trip has been solved. She makes “done this time” visible: three venues, three criteria, and one shortlist that supports a choice. The box preserves a clean entry point for later work without allowing later work to hold today’s deliverable hostage.
References
- OpenAI Help Center: Deep research in ChatGPT — https://help.openai.com/en/articles/10500283-deep-research-in-chatgpt [accessed: 2026-08-01]
- Google Gemini Apps Help: Use Deep Research in Gemini Apps — https://support.google.com/gemini/answer/15719111 [accessed: 2026-08-01]
- UK Department for Culture, Media and Sport: Rapid evidence assessment of valuation methods for civil society — https://www.gov.uk/government/publications/rapid-evidence-assessment-of-valuation-methods-for-civil-society/rapid-evidence-assessment-of-valuation-methods-for-civil-society [published: 2025-07-11]
- NIST AI Resource Center: AI RMF Core — https://airc.nist.gov/airmf-resources/airmf/5-sec-core [accessed: 2026-08-01]
- OWASP Cheat Sheet Series: LLM Prompt Injection Prevention Cheat Sheet — https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html [accessed: 2026-08-01]
