Scoping the research question.

Why scope is the research equivalent of “read the plan” CORE

In Module 3 you learned four moves for a directed edit: locate, plan, write, verify. You checked locate and plan before the agent wrote a single line of code, because a well-written change to the wrong file is still wrong.

Research has the same structural leverage point, earlier in the process. Scope is where the research task gets fixed in place. Get it right and every move downstream — retrieval, triangulation, synthesis — has a target to aim at. Get it wrong and the agent does beautiful work on the wrong question. A research agent running on a vague prompt produces a vague brief. You cannot fix it downstream. You have to fix it before retrieval begins.

Here is why scope cannot be the agent’s job. The agent does not know why you are asking. It does not know who is going to read the output, what decision the output is feeding, what your time budget is, what you already know, or what a wrong answer would actually cost you. All of that is in your head. The best the agent can do is ask clarifying questions — and clarifying questions are only as good as the answers you give them. So scope lives on your side of the line. The agent helps you sharpen scope; it does not produce scope.

This is the same move, structurally, as writing a directed-edit prompt in Module 3. You named the goal, the suspected location, the required plan step, the required verify step. A scoping brief does the same work — it tells the agent what counts as a successful run before any work begins.

The five parts of a scoping brief CORE

A Module 4 scoping brief has five parts. Each part exists because at least one failure mode happens when it is missing.

1. The question, in one sentence. Not the topic. The question. “Solar panels” is a topic. “Given our roof, our electricity usage, and Ohio utility rates, does installing solar panels at 1234 Main Street pay back within ten years?” is a question. The difference is that the second one can be answered. If you cannot write your question in a sentence that ends with a question mark, scope is not done.

2. The reader and the decision. Who is going to read the output of this research, and what are they going to do with it? “My mom, before she commits to the quote from Solar Co.” “Me, as the topic sentence of my AP US History paper on Reconstruction.” “Our homeschool co-op board, deciding whether to standardize on one writing curriculum next year.” The reader and the decision together tell the agent what a useful answer looks like — a useful answer to your mom is different from a useful answer to your co-op board.

3. What would count as an answer. One to three sentences that sketch the shape the answer will take. “A two-paragraph comparison of the two curricula on price, scope-and-sequence, and parent time cost, with one sentence at the end saying which fits our situation best.” Or: “A list of the three most-cited arguments on each side, with a representative quote for each.” If you cannot sketch the shape of the answer, you will not know when the agent has produced it.

4. What is out of scope. Name two or three things the agent should not spend time on. “Not comparing the environmental impact of the two curricula.” “Not the history of homeschooling law.” This is the line that most students skip and that most prevents agent drift. A research agent left to its own judgment on what is in-scope will usually produce a wider brief than you wanted, because producing more looks more helpful.

5. Time and source constraints. How much time are you willing to spend on this, and what sources are preferred or required? “Two to three hours of my time including reading. Prefer sources from the last three years. Primary sources on the curricula’s own websites are required; reviews from homeschool bloggers are fine; social-media threads are not.” Constraints shape retrieval. Without them the agent will default to whatever is easy to find.

That is the scoping brief. Five parts. One page. Written before retrieval begins.

Four vague research requests, converted CORE

A good way to learn scoping is to take bad research requests and rewrite them. Here are four real ones, with the scoping work done on each.

Vague: “Research homeschool curricula.”

Bounded: Question: For a family of two students in 4th and 7th grade, with a working parent who can spend about 90 minutes a day on direct instruction, which of three specific math curricula — Beast Academy, Singapore Math, and Saxon — is the best fit? Reader & Decision: My mom, choosing our curriculum for next year by end of May. Answer Shape: A one-page comparison table on three axes — parent time cost, scope-and-sequence fit for both grade levels, and price including consumables — with a one-paragraph recommendation at the end. Out of Scope: other math curricula; curriculum comparisons older than two years; religion-specific considerations (handled separately). Constraints: four hours of my time total; primary sources on each publisher’s website are required; two independent reviews from homeschool parents per curriculum are preferred.

Vague: “Find out about the Reconstruction era.”

Bounded: Question: What were the three most significant legal changes during Reconstruction (1865–1877), and how did each change unravel by 1900? Reader & Decision: Me, writing an AP US History essay due April 30; this is the topic sentence. Answer Shape: One paragraph per change, naming the legal instrument (amendment, statute, court case), the intended effect, and the mechanism by which the effect was reversed or gutted; each paragraph has at least one primary-source quotation. Out of Scope: military history of Reconstruction; cultural history; figures not directly involved in the three changes. Constraints: five hours of my time; primary sources (Congressional record, Supreme Court opinions, state statutes) required; textbook-level secondary sources fine; Wikipedia only for initial orientation.

Vague: “Research whether my dad should switch accounting software.”

Bounded: Question: For a one-person tutoring business invoicing roughly 30 clients/month with no employees, does switching from QuickBooks Self-Employed to Wave save enough money and time over a year to be worth a two-week transition? Reader & Decision: My dad, deciding whether to switch before his QuickBooks renewal date on June 1. Answer Shape: A single page with three sections — (a) one-year total cost comparison at his volume; (b) feature delta for the things he actually uses (invoicing, 1099 tracking, sales tax); (c) transition cost estimate in hours — ending with a switch / do not switch recommendation and the confidence level. Out of Scope: accounting software for larger businesses; any software neither of the two above. Constraints: three hours of my time; both products’ current pricing pages required; two recent third-party reviews per product preferred; no paid review sites.

Vague: “Write something about AI in education.”

Bounded: Question: What do public high school English teachers actually say (in trade press, surveys, and interviews since 2024) about grading student work in the presence of AI, and what are the three most commonly proposed policy responses? Reader & Decision: Me, for a personal essay I’m submitting to the school paper by May 10. Answer Shape: Two paragraphs of teacher-voice synthesis (what the pattern sounds like in their own words) followed by one paragraph on the three most commonly proposed policies; total length about 700 words; every teacher quote attributed. Out of Scope: university-level teachers; cheating-detection technology; AI-tool vendor claims. Constraints: four hours of my time; published pieces from 2024 or later; at least two trade-press sources (Education Week, Edutopia, Chalkbeat) and one survey-based source required.

Notice what scoping did. It turned each request from a topic the agent could wander inside into a target the agent can aim at. A research agent run against any one of the bounded questions will produce usable output. A research agent run against any of the vague starting points will produce a pile.

Using the agent as a critic of your scope CORE

There is one good use of the agent in the scoping move: ask it to critique your scope, not write it.

After you have written your scoping brief, paste it into Cowork or Claude Code with a prompt like: “This is my scoping brief for the research task I am about to hand to a research agent. Do not run the research yet. Instead, read the brief and tell me: (a) any question I left ambiguous; (b) any place where ‘what counts as an answer’ is not specific enough; (c) anything I probably meant to exclude but did not list in Out of Scope; (d) any constraint that will be impossible to satisfy.”

This is a genuinely good use of the agent. It is reading fresh eyes over your scope before the expensive work begins, and it costs almost nothing. Expect three to six edits per scoping brief from this pass. Apply the edits you agree with; reject the ones that do not match your actual intent.

What you do not do here is let the agent rewrite the brief for you. If you ask the agent to “improve the scoping brief,” you get back a version that sounds smoother and almost always drifts toward what sounds impressive rather than what you actually need. Keep the author’s pen in your hand.

One and only one scope per session CORE

Research agents — and most AI agents — degrade when the session tries to serve two scopes at once. A common student failure mode: start with “compare the two math curricula,” drift into “actually, also tell me about writing curricula,” and end with a brief that is a bad answer to both questions.

The discipline is one scope per session. If you find yourself wanting a second research question, finish the first one (or explicitly park it), close the session, and start a new one. This is the same principle from Module 1’s conversation-as-state — each session carries its context forward, and muddying the scope midway through costs you clean answers.

In practice: if your scoping brief is strong, you will rarely be tempted to drift. If you are tempted to drift, it is usually a sign the original scope was too narrow for what you actually wanted — in which case stop, revise the scope, and restart.