From transcript to deliverable: the AI-assisted research-report workflow

The engagement involves seven stakeholder interviews. You've done six of them. Each one is a 45-minute conversation with someone whose perspective matters to the diagnostic you're producing. The transcripts are in a folder. The deliverable is due in nine days.

The trap, for almost every solo consultant who's worked on this kind of engagement, is the same: you wait until you have all seven transcripts in hand, then sit down on a Saturday and try to synthesize the whole thing in one push. By Sunday afternoon you have a 40-page draft that's structured around the interviews instead of structured around the insight. The client sees a research report; what they paid for was a recommendation.

This article is the workflow that produces the recommendation in nine days instead of fifteen, by handling the synthesis differently — and by using AI for the specific parts that are AI-tractable, while keeping AI out of the parts that aren't.

Where the typical workflow breaks

Most consultants' interview-to-deliverable workflow has three failure modes that compound.

Failure mode 1: synthesizing only at the end. All interviews finish; the consultant then sits with the pile and tries to find the pattern. By the time they're processing interview five, they've forgotten the texture of interview one. The synthesis ends up dominated by recency bias — the last two or three interviews shape the entire framing.

Failure mode 2: structuring the deliverable around the inputs. "Here's what we heard from Sales. Here's what we heard from Engineering. Here's what we heard from Product." This is a research summary, not a deliverable. The client paid for a synthesis that takes a position; the input-structured version just hands them back what they could have read in the transcripts themselves.

Failure mode 3: writing the executive summary last. The exec summary at the top of the deliverable is the only thing 70% of stakeholders will read. Writing it last means it's a paraphrase of the document — accurate, but not sharpened by the discipline of writing it first. The exec summary written first is shorter, clearer, and braver.

The workflow below addresses all three.

The four-stage workflow

The four stages happen across the engagement, not in one weekend push. The total person-time across all four is similar to the bad version above; the difference is sequencing and quality.

Stage 1 — Process each interview within 24 hours

Right after each interview, while the texture is still fresh, run a Claude prompt that takes the transcript (or rough notes) and produces three artifacts: a one-page debrief (similar to the discovery debrief format), a list of three to five direct quotes worth preserving with attribution, and a short list of open questions the interview surfaced.

Time per interview: 12-15 minutes. Total across seven interviews: ~90 minutes spread across two weeks, never more than 15 minutes in one sitting.

The debrief and the quotes go into a clients/<name>/interviews/ folder. The open questions go into a running document — clients/<name>/open-questions.md from the client-memory-across-sessions article. By the time the seventh interview is done, this file is the spine of the deliverable.

Stage 2 — Identify the spine after interview five

Five is the threshold where patterns start showing up. After interview five (not seven, not the end), run a second Claude prompt with all five debriefs as input. The prompt's job: "What's the through-line that's emerging? What position should the deliverable take? Where is the evidence strongest, and where is it weakest?"

The output is the spine of the deliverable — the position you'll be defending in writing. It's not the deliverable yet; it's the one-paragraph version of the deliverable's argument. From this point, the remaining two interviews become targeted — you go into them with questions that will either confirm or challenge the emerging spine, rather than just gathering more data.

Time: 20 minutes. The single highest-leverage 20 minutes in the entire engagement.

Stage 3 — Write the exec summary first

Before drafting any of the deliverable's body, write the exec summary. One page. Maximum 500 words. The exec summary states the position, names the three to five recommendations, and previews the structure of the deliverable.

This is the stage where you don't use AI heavily. You may use Claude to polish a draft you've already written, but the exec summary needs to come from your judgment, not from synthesis. It's the part of the work that's hardest to fake and the part the client will read most carefully.

Time: 30-45 minutes. Write it Monday morning of the deliverable week, not Friday afternoon.

Stage 4 — Draft the body sections, structured by argument

The body of the deliverable is structured around the argument, not the inputs. Each section advances the position from the exec summary. Within each section, the evidence (quotes, observations, data points) comes from the seven interviews — but the section's job is to make a point, not to summarize who-said-what.

This is the stage where AI is most useful. You give Claude a section heading, the relevant subset of interview debriefs and quotes, and the position the section needs to defend. Claude produces a draft that takes the position; you edit for voice and sharpness. Repeat for each section.

Time per section: 20-30 minutes drafting, 15-20 minutes editing. For a deliverable with five body sections: roughly 4-5 hours of total work, spread across two or three days.

Where Claude is wrong for the job

The four-stage workflow uses AI for synthesis, drafting, and pattern-recognition. There are three places it's the wrong tool entirely:

The judgment calls. Whether to recommend Path A or Path B isn't an AI question; it's a consultant question. The model can produce plausible cases for both; the choice is yours. If you let the AI pick, the deliverable reads like AI work — the client notices.

The relationship-sensitive language. When a stakeholder said something that you need to surface but that names a specific colleague critically, the AI doesn't know which way to phrase it for political safety. You do. Edit those passages by hand.

The structural decision. "Should this be a 30-page deliverable or a 12-page one?" AI defaults to longer because more text feels safer. The right answer is almost always shorter than the AI suggests. Decide the length before drafting; hold to it during.

The brief that runs all four stages

Each Claude prompt across the four stages benefits from the one-pager brief format — five sections, 90 seconds to write, transforms generic output into specific. For the deliverable workflow specifically, the brief should be written once for the engagement (the kickoff-context plus the spine that emerged at Stage 2) and reused as the input prefix for every Claude prompt across stages 1, 2, and 4. Three changes from the standard brief: it includes the engagement's working hypothesis after Stage 2, it includes the voice rules calibrated to the client's stakeholders, and it includes any "don't go there" instructions (sensitive topics, people to not name, etc.).

The brief discipline is what keeps the workflow's outputs consistent across two weeks of separate Claude conversations. Without it, the prompts each produce slightly differently-voiced output, and the deliverable ends up sounding like seven authors wrote it.

What this lets you actually ship

A consultant who runs the four-stage workflow on a seven-interview engagement ends the deliverable week with a draft that's roughly 80% there. Monday: exec summary written. Tuesday-Wednesday: body sections drafted with Claude. Thursday: editing pass, sharpening, polishing. Friday: send.

Compare to the typical version: Friday before the deadline, sit down with seven transcripts, panic-synthesize for two days, send Sunday night. Same number of hours, much worse output, almost certainly missing the spine that would have made the deliverable land.

The deliverable shape that comes out of the four-stage workflow has three observable traits the typical version doesn't: an exec summary that takes a clear position, a body that advances an argument rather than summarizes inputs, and quotes used to evidence the argument rather than carry the argument. Clients notice the difference. Some can name it; most can't. But the consultant who delivers this version gets the follow-on engagement.

A two-week timeline, day by day

Concrete shape. A seven-interview diagnostic, two-week deliverable window, four-stage workflow distributed across the calendar instead of crammed into the last weekend.

Week 1

• Mon: Interview 1 (45 min). Same day: Stage-1 processing (12 min) → debrief filed, quotes preserved, open questions added.
• Tue: Interview 2. Stage-1 processing.
• Wed: Interview 3. Stage-1 processing.
• Thu: Interview 4. Stage-1 processing.
• Fri: Interview 5. Stage-1 processing. End of day Friday: Stage 2 — run the spine prompt (20 min). Spine emerges; targeted questions queued for interviews 6 and 7.

Week 2

• Mon: Interview 6 (now targeted toward spine validation). Stage-1 processing.
• Tue: Interview 7 (targeted). Stage-1 processing. By end-of-day Tuesday, all seven interviews processed and the spine has been tested twice.
• Wed: Stage 3 — exec summary written (45 min). Position locked, recommendations named.
• Thu: Stage 4 — body sections drafted with Claude using the brief. ~5 hours of focused work.
• Fri morning: Editing pass, sharpening, polishing.
• Fri afternoon: Send.

Total human time across the two weeks: roughly 18-22 hours (interviews + processing + drafting + editing). Same as the panic-synthesis version, distributed differently. The output difference comes from the distribution, not from working harder.

Stage-1 mistakes that compound by Stage 2

The most common reason the workflow fails isn't anything in Stages 2-4 — it's failure to actually run Stage 1 within 24 hours of each interview. Three patterns cause this.

Batching interviews and trying to process them together. "I'll do Tuesday and Wednesday's interviews, then process all four on Thursday." Wrong. By Thursday, Tuesday's texture is decayed. Each interview gets processed the same day, period.

Half-processing. Skipping the quotes step because it feels redundant. The direct quotes preserved at Stage 1 are what carry the deliverable's voice at Stage 4. Without them, the body sections sound like AI summaries instead of stakeholder evidence.

Letting Stage 1 expand into synthesis. The Stage 1 debrief is one interview's take; it doesn't try to find the through-line. Trying to synthesize at Stage 1 produces narrowed framing that biases Stage 2's pattern-recognition.

If Stage 1 happens cleanly, Stages 2-4 work. If Stage 1 slips, the whole workflow degrades to the panic-synthesis pattern.

Where this fits in the broader practice

The deliverable-drafting work above is covered in more depth in Chapter 6 of The Solo Consultant's AI Playbook — Deliverable Drafting, which makes the broader argument about where AI helps and where it kills your reputation, with the right workflow for each. The Playbook chapter is the framing; this article is the specific drill-down for the interview-to-deliverable variant of that work.

The interview-processing piece in Stage 1 is closely related to the client meeting workflow — same shape of work, slightly different output structure (interviews are evidence-gathering; meetings are decision-tracking). Both run the same 12-minute processing routine per session. And once the deliverable ships, the same engagement materials feed the case-study harvest — one wrap, several marketing assets from the artifacts you already produced. If you'd rather run this whole workflow in files than in chat — transcripts dropped into a folder, the deliverable landing as a dated file — point Claude at a folder is the workspace pattern built for exactly that.

The single move that produces the biggest output-quality change in this workflow isn't any one stage. It's the discipline of processing each interview within 24 hours instead of letting them stack up. That habit alone shifts the entire workflow from end-of-cycle panic-synthesis to gradual emergence of the spine. If you take one thing from this article, take that.

The deliverable is an argument. The interviews are evidence. AI is leverage. Your judgment is the work.