AI Prompt Checklist for SMB Proposals

Most founders who write proposals spend the most time on the part AI completes fastest. Noy and Zhang's preregistered experiment with 444 college-educated professionals found that participants with ChatGPT access finished comparable writing tasks in roughly 40 percent less time than those without it, while evaluators rated their outputs approximately 18 percent higher in quality. The founders grinding through a two-hour proposal draft on a Tuesday night are not doing the hard part. They are doing the slow part.

The shift that actually happened in that experiment

Noy and Zhang tracked not just time and quality but how participants allocated effort across the task. Before AI access, writers spent about half their time on rough drafting. With AI access, that share fell by more than half, while editing time more than doubled. Writers did not stop working. They stopped doing the lowest-judgment work and started doing more of the highest-judgment work.

That shift is the thing worth building toward in proposal writing. A founder who types a bare-bones prompt and pastes the output into a Google Doc is not replicating it. The editing time doubled in the experiment because writers had something worth editing. A weak prompt produces output so generic that editing it takes longer than starting over.

When unstructured prompting is good enough, and when it isn't

A reasonable objection to the checklist idea goes like this: the Noy and Zhang experiment used no checklist. Participants typed their own prompts without any template, and still produced outputs rated 0.4 to 0.45 standard deviations higher in quality than the control group. If those gains appear without structure, the structure is optional.

The objection is correct for the tasks the experiment tested. Press releases, short reports, analysis plans. These tasks share one property: the writer already holds all relevant content before typing. A press release about a product launch requires no external client context. You know the product. You write the prompt.

An SMB proposal requires something different. You need to encode a specific buyer's stated pain point, their competitive situation, their budget range, and the solution framing tied to their priorities. That information varies by client and is easy to omit under time pressure. The checklist's function is not to improve AI's prose quality in the abstract. It is to prevent you from prompting AI with incomplete information. Generic output in a proposal does not just read poorly. It signals to the buyer that the proposal was not written for them, which is the exact failure mode that loses deals.

What the checklist actually contains

The prompt template needs four inputs before you send anything to AI. First, client context: company name, industry, size, and the specific outcome they described in the last conversation. Second, the pain point in their words, not yours. If they said "we're losing deals because our turnaround is too slow," that phrase goes into the prompt verbatim. Third, the solution structure: which of your offerings maps to which part of their problem, and in what order you want to present it. Fourth, tone guidelines: formal or conversational, whether they respond to data or narrative, any terminology they used that you want mirrored back.

With those four inputs loaded, AI produces a draft oriented toward a specific buyer rather than a generic archetype. You spend your editing time on pricing logic, relationship nuance, and the one differentiating claim that no template surfaces on its own.

Who gains most from this

Noy and Zhang found that workers who scored lower on the first task gained more from AI access than strong writers, narrowing the spread of quality across the sample. [Inference] For SMB founders who are strong at sales but weaker at written persuasion, that compression effect is the most useful property of the whole system. A founder who closes deals in conversation but freezes at a blank document gets a first draft that reflects the conversation they already had, provided the checklist captured it accurately.

The checklist is a forcing function for that capture. It does not make AI smarter. It makes the founder less likely to skip the inputs that separate a proposal written for this buyer from one written for any buyer.

Run the checklist once, save the template, update the four fields per client. The editing session that follows is the work. Everything before it is setup.