Evidence layer: academic literature + grey literature from Big Consulting
Who This Is For #
This guide is written for high-intent operators who produce bylined thought leadership at a professional standard: agency leads ghostwriting for client executives, in-house comms teams building an executive’s publishing program, and executives who write under their own name. The workflow is the same in all three cases — only the voice-capture step changes.
The premise: a byline lives or dies on two things editors can smell instantly — a defensible original argument and evidence that isn’t recycled from the same three blog posts everyone else cites. Claude compresses the labor between those two things. It does not replace the argument (that’s yours or your principal’s), and it must never be trusted as the source of the evidence (that’s what the sourcing protocol below is for).
Operating assumptions:
- You’re producing 700–1,200 word bylines for trade press, business press contributor programs, or industry publications — not peer-reviewed papers.
- You have access to Claude (claude.ai with Projects and Research mode, or Claude via Cowork/Claude Code for file-heavy workflows).
- You want a repeatable pipeline, not a one-off trick. Every prompt in this guide is a template you can standardize across clients or executives.
Why Grey Literature from Big Consulting Is Your Unfair Advantage #
Most bylines cite whatever surfaces on page one of Google. Operators who systematically mine consulting grey literature (McKinsey Global Institute, BCG Henderson Institute, Bain, the Big Four insight arms) and academic working papers (SSRN, NBER, Semantic Scholar) get three edges:
Editor-credible authority. A stat sourced to MGI’s latest workforce report or a PwC CEO Survey carries institutional weight that a vendor blog post never will. Editors don’t have to defend the source to their readers.Freshness. Flagship consulting surveys refresh annually or quarterly — McKinsey’s State of AI survey, Deloitte’s quarterly State of GenAI in the Enterprise, PwC’s AI Jobs Barometer and Global CEO Survey, Accenture’s quarterly Pulse of Change. Working papers on SSRN and NBER surface findings 12–24 months before journal publication. You can cite evidence your competitors haven’t seen yet.Specificity. Consulting research quantifies things executives actually argue about (adoption rates, ROI distributions, workforce effects), which gives your byline the concrete numbers that separate an argument from an opinion.
The catch: grey literature is unrefereed, methodologically uneven, and produced by firms with commercial agendas. That’s why Stage 3 of this pipeline includes a formal vetting protocol (AACODS) rather than “Claude said it’s from McKinsey, ship it.”
The Pipeline at a Glance #
| Stage | What Happens | Claude’s Role | Your Role |
|---|---|---|---|
| 0 | Infrastructure setup | Holds voice, style, standing rules in a Project | Build the corpus once |
| 1 | Raw material capture | Interviews you / structures the brain dump | Supply the raw thinking |
| 2 | Thesis extraction | Pressure-tests angles, finds the sharpest claim | Choose the argument |
| 3 | Evidence layer | Researches, retrieves, summarizes sources | Verify every citation |
| 4 | Outline + draft | Drafts in the captured voice | Direct, redline, re-voice |
| 5 | Editorial QA | Red-teams, fact-checks, scrubs AI-tells | Final judgment |
| 6 | Submission prep | Pitch email, headline options, disclosure | Send it |
Total operator time for a first byline through this pipeline: 2–4 hours. For subsequent bylines with the Project already built: 60–90 minutes. Compare that to the 6–10 hours a quality ghostwritten byline traditionally takes, with most of the savings coming from research retrieval and first-draft mechanics — not from skipping verification, which stays fully manual.
Stage 0 — Build the Infrastructure Once #
Do this in a Claude Project (one per executive/byline author). Projects persist documents and instructions across every conversation, so you stop re-explaining the voice every time.
0.1 The voice corpus
Upload 3–8 samples of the author’s authentic writing or speech:
- Previously published bylines or op-eds (best signal)
- Long-form LinkedIn posts written personally, not by the team
- Interview or podcast transcripts (excellent for executives who talk better than they write)
- Internal memos or all-hands notes, if available
Transcripts matter more than most operators realize: an executive’s spoken cadence is usually closer to their “true” voice than their edited prose, and editors increasingly reject bylines that read like committee output.
0.2 The voice profile
Run this once and save the output as a Project document:
PROMPT — Voice Profile ExtractionAnalyze the attached writing samples from [NAME]. Produce a voice profile I can use to direct future drafts, covering: (1) sentence length distribution and rhythm; (2) vocabulary register — words they favor and words they’d never use; (3) how they open arguments (anecdote, data, contrarian claim?); (4) how they handle hedging and certainty; (5) signature constructions or verbal tics worth preserving; (6) three things a ghostwriter would most likely get wrong imitating them. Output as a directive style sheet, not a description — phrase each item as an instruction to a writer.
0.3 Standing project instructions
Paste into the Project’s custom instructions. This is the single highest-leverage artifact in the whole pipeline:
You are the drafting partner for bylined articles by [NAME], [TITLE] at [COMPANY]. Always: (1) Write in the voice defined in the attached voice profile. (2) Never invent statistics, studies, quotes, or sources — if you don’t have a verified source in this conversation, write [EVIDENCE NEEDED: description] instead. (3) Flag any sentence that promotes [COMPANY]’s products; bylines must not read as sales collateral. (4) Target [PUBLICATION TYPE] editorial standards: 700–1,200 words, clear nut graf by paragraph three, actionable takeaway for [READER]. (5) When I paste research excerpts, treat only those excerpts as citable — do not blend in your training-data recollection of the same reports.
Rule (5) deserves emphasis. Claude’s training data contains summaries of older consulting reports, and it will happily blend a 2023 half-memory into your 2026 citation if you let it. Force the discipline: only pasted or fetched-and-verified material is citable.
0.4 The evidence log
Create a running document (spreadsheet or markdown table) per byline: claim → source → exact stat/quote → URL → date verified → verified by whom. This is your audit trail when an editor or client asks “where did this number come from?” — and they will.
Stage 1 — Capture the Raw Material #
The most common failure mode in ghostwritten bylines is starting from nothing and asking Claude to “write an article about X.” The output is competent, generic, and unpublishable. Claude’s job at this stage is extraction, not generation.
For agency/in-house operators: the structured interview
Get 20 minutes with the principal. Record it. If you can’t get time, send them five questions and take voice-memo answers. Then:
PROMPT — Interview MiningAttached is a transcript of my interview with [NAME] about [TOPIC]. Extract: (1) every opinion that is arguable — i.e., a smart peer could disagree; (2) every first-hand anecdote or specific example from their operating experience; (3) every claim that needs external evidence to stand up; (4) anything surprising or contrarian relative to the conventional wisdom on this topic. Do not paraphrase into blandness — preserve their exact phrasing where it’s vivid. Rank the arguable opinions by how distinctive they are.
For executives writing as themselves: the brain dump
Talk into the mic for ten minutes, unstructured. Or answer Claude’s questions live:
PROMPT — Socratic ExtractionI want to write a byline about [TOPIC] but my thinking is rough. Interview me one question at a time — no more than one question per turn. Push on: what I believe that most of my industry doesn’t, what I’ve seen firsthand that shaped this view, where I might be wrong, and who specifically needs to hear this. After 8–10 questions, stop and summarize the strongest byline-worthy thesis you heard.
The one-question-per-turn constraint matters. Left alone, Claude will ask six questions at once and you’ll answer none of them well.
The raw-material bar
Before moving to Stage 2, you should have: at least one arguable opinion, at least one first-hand anecdote or proof point, and a sense of the target reader. If you don’t have all three, the byline will be a listicle wearing a suit. Go back and extract more.
Stage 2 — Extract and Pressure-Test the Thesis #
Editors reject bylines for being “fine.” A publishable thesis is specific, arguable, and timely. Run the raw material through a deliberate pressure test:
PROMPT — Thesis ForgeFrom the raw material in this conversation, propose 5 candidate theses for a byline in [TARGET PUBLICATION]. For each: (1) state it in one sentence a busy editor would read; (2) name the conventional wisdom it pushes against; (3) identify what evidence would be needed to make it credible; (4) rate its risk of being generic on a 1–5 scale, where “everyone in this industry has read this take before” is a 5. Then tell me which one you’d pitch and why — and which one is the trap that sounds good but says nothing.
Then invert it:
PROMPT — Steelman the OppositionTake thesis [N]. Write the strongest 200-word rebuttal a credible skeptic — say, a partner at a top consulting firm or a domain academic — would publish in response. Then tell me: does the thesis survive? What would it need to concede or sharpen?
Two operator notes:
The rebuttal often becomes a paragraph in the byline. Pre-empting the smart objection is a hallmark of senior writing, and editors notice.Timeliness is a multiplier. Ask Claude (with web search on) what news hook existsthis monthfor the thesis — a fresh report release, an earnings season pattern, a regulatory move. A byline pegged to something current gets read; the same byline “in general” gets queued and dies.
Stage 3 — The Evidence Layer #
This is the stage that separates operator-grade output from AI slop, and it’s where the academic + consulting grey-lit strategy lives. Work it in three passes: map, retrieve, vet.
3.1 Map the source landscape
Know where the institutional research actually lives before you search:
Big Consulting grey literature — primary hubs:
| Firm | Hub | Flagship recurring research worth watching |
|---|---|---|
| McKinsey | ||
bcg.com/publications+BCG Henderson Institutebain.com/insightsDeloitte InsightsBeyond the firms: World Economic Forum reports (often co-authored with these firms), Harvard Business Review (the bridge between academic and practitioner literature), and Brookings/Peterson Institute for policy-adjacent arguments.
Academic layer — free access routes:
— broadest index; use the “cited by” chain to find the seminal paper behind a claim.Google Scholar— 214M+ papers, AI-generated TL;DRs, citation-intent classification. The fastest triage tool.Semantic Scholar— working papers in business, economics, law; 12–24 months ahead of journals.SSRN— economics working papers; the source behind half the stats you see secondhand in business press.NBERandOpenAlex— open citation graphs for tracing influence.Lens.org— browser extension that finds legal open-access copies of paywalled papers.Unpaywall— curated open-access journal directory.DOAJ
3.2 Retrieve — with Claude doing the labor
Use Research mode (or web search) in claude.ai for the sweep, then bring the artifacts into the Project. PROMPT — Evidence SweepResearch the evidence base for this thesis: “[THESIS].” I need: (1) the 2–4 most recent, most authoritative statistics from major consulting firm research (McKinsey/MGI, BCG Henderson Institute, Bain, Deloitte Insights, PwC, Accenture) — for each, give me the exact figure, the report title, publication date, sample/methodology if stated, and the direct URL; (2) 1–2 academic or working-paper findings (SSRN, NBER, or peer-reviewed) that either support or complicate the thesis; (3) any credible evidence AGAINST the thesis — I need to know what I’m arguing past. Do not include any statistic you cannot tie to a live URL. If the evidence base is thin, say so plainly.
Then, for each promising source, go deeper with the actual document:
PROMPT — Source Deep-Read(after up the PDF or pasting the URL)From this report only — do not supplement from memory: (1) extract the 3 findings most relevant to “[THESIS]” with page/section references; (2) state the methodology and sample in one sentence; (3) note any caveat the authors themselves flag that would embarrass me if an editor knew it and I didn’t; (4) give me the exact sentence I could quote, verbatim, with quotation marks.
Up the actual PDF is materially better than letting Claude summarize from search snippets. Consulting firms’ headline stats often come with methodology caveats buried on page 40, and the caveat is sometimes the more interesting byline material.
3.3 Vet — the AACODS pass
Grey literature is unrefereed by definition, so borrow the standard librarians use: the ** AACODS checklist** (Authority, Accuracy, Coverage, Objectivity, Date, Significance). Operationalized for consulting research:
Authority— Is this from the firm’s research institute (MGI, BHI) or a marketing team’s blog post? The institutes have named researchers and stated methods; the blogs often don’t. Cite the former.Accuracy— Does the report disclose sample size, respondent profile, and field dates? “A survey of executives” without an N is a red flag.Coverage— Does the sample match your claim? A survey of Fortune 500 CTOs does not support a claim about mid-market adoption.** Objectivity**— Every consulting report has a commercial thesis (usually: this problem is big, hire us). Use theirdata, be skeptical of theirframing. If the firm sells services in the exact area the report covers, say so to yourself and consider whether the stat survives that discount.Date— In fast-moving domains, a 2024 adoption stat cited in 2026 is worse than no stat; it signals you’re not current. Prefer the latest wave of recurring surveys.Significance— Is this stat load-bearing for your argument or decorative? Cut decorative stats; each citation is a place your byline can be attacked.
Run it as a prompt:
PROMPT — AACODS VetEvaluate each source in my evidence log against AACODS (Authority, Accuracy, Coverage, Objectivity, Date, Significance). For each: score high/medium/low per dimension with one-line justification, flag any source where the sample doesn’t actually support how I plan to use it, and identify the single weakest citation in the set — the one a hostile fact-checker would attack first.
3.4 The anti-hallucination protocol (non-negotiable)
Language models fabricate citations with complete fluency — plausible report titles, realistic-sounding percentages, correct-format URLs that 404. For bylines published under a real person’s name, one fabricated stat is a career event. The protocol:
Never cite anything Claude asserts from memory. Only material retrieved with a live URL in-session, or from a document you uploaded, enters the evidence log.Click every URL yourself. Confirm the page exists, the stat appears on it, and the number matches to the decimal.Verify quotes verbatim. Ctrl-F the exact quoted string in the source document. Paraphrases presented as quotes are how corrections columns get written.Two-source load-bearing claims. If a single stat carries the whole thesis, find independent corroboration or hedge the language.Log everything. Date-stamped evidence log, per byline. When the client’s GC or the publication’s fact-checker calls, you answer in five minutes, not five hours.
Claude can help with its own audit:
PROMPT — Citation AuditHere is my draft and my evidence log. Cross-check every factual claim, statistic, and quote in the draft against the log. Output a table: claim → log entry it maps to → EXACT MATCH / PARAPHRASE / NO SOURCE FOUND. Flag any number in the draft that appears in no log entry. Do not be charitable.
Stage 4 — Outline and Draft #
4.1 Architecture before prose
PROMPT — Byline ArchitectureBuild a paragraph-by-paragraph outline for a [WORD COUNT]-word byline in [PUBLICATION]: (1) Lede — open with [the anecdote / the sharpest stat / the contrarian claim] from our material; no throat-clearing, no “In today’s rapidly evolving landscape.” (2) Nut graf by paragraph 2–3 — the thesis, stated plainly, and why now. (3) 2–3 argument blocks, each pairing one verified piece of evidence from the log with first-hand experience or interpretation the evidence alone can’t supply. (4) The concession paragraph — the strongest objection, acknowledged and answered. (5) Close with a specific, doable call to action for [READER] — not “leaders must embrace change.” For each paragraph, note which evidence-log entry it draws on.
The pairing rule in (3) is the craft secret of the whole genre: evidence proves the problem is real; experience proves the author has standing to talk about it. A byline that is all stats reads like a report summary. All anecdote reads like a LinkedIn post. The alternation is the form.
4.2 The draft
PROMPT — First DraftDraft the full byline from the approved outline, in the voice profile, at [WORD COUNT] words. Constraints: only cite from the evidence log, using [EVIDENCE: log-ID] markers I’ll convert to prose citations; no sentence may mention [COMPANY]’s products or services; vary sentence length aggressively — this author uses short declaratives after long setup sentences; and give me three headline options plus a one-sentence dek, written for [PUBLICATION]’s style.
4.3 Killing the AI-tells
Editors in 2026 pattern-match AI prose instantly, and duplicate-content and AI-detection screening is now routine at trade publications. The tells and the fix:
PROMPT — De-Tell PassEdit this draft to remove AI-characteristic patterns: (1) triadic lists everywhere (“faster, cheaper, and more scalable”); (2) hedged both-sidesism (“while X, it’s also important to note Y”); (3) empty transition scaffolding (“Moreover,” “Furthermore,” “In conclusion”); (4) the words “landscape,” “leverage,” “delve,” “crucial,” “robust,” “navigate,” and “ever-evolving”; (5) any paragraph that summarizes the previous paragraph. Where you cut, replace with something only this author could say, drawing on the interview material. Show changes as a before/after list, then the clean draft.
Then the human pass that no prompt replaces: read it aloud in the author’s voice. Anywhere you stumble, or can’t imagine them saying the sentence to a peer over coffee, rewrite by hand. Ten minutes of this is worth more than three more Claude passes — and for ghostwritten pieces, this is also the moment to send it to the principal with the note “mark anything that doesn’t sound like you.” Their redlines feed the voice profile for next time.
Stage 5 — Editorial QA #
Three independent passes before anything leaves your desk:
Pass 1 — Red team.
PROMPT — Editorial Red TeamAct as a skeptical senior editor at [PUBLICATION] with 40 competing submissions this week. Read this byline and give me: the reason you’d reject it in one sentence; the weakest paragraph and why; any claim you’d send to a fact-checker; anywhere it drifts into vendor self-promotion; and whether the headline overpromises what the piece delivers. Then — only after the critique — what would make you accept it.
Pass 2 — Fact-check. Run the Citation Audit prompt from Stage 3.4 against the final draft, not the first one. Numbers drift during editing — a “nearly 40%” becomes “40%” becomes “more than 40%” across revisions, and only the first was true.
Pass 3 — Compliance and self-promotion scrub. Trade publications uniformly reject bylines that pitch the author’s company; readers are turned off the moment the piece points them toward the author’s product. Also confirm: no recycled passages from the author’s previous published bylines (editors screen for duplicate content), no client-confidential examples without clearance, and — for regulated industries — whatever legal/comms review your context requires.
Stage 6 — Submission Prep #
6.1 Targeting
PROMPT — Target and Angle Map(web search on)Given this byline on [TOPIC] by [NAME, TITLE], identify 6–8 publications with active contributor or op-ed programs reaching [AUDIENCE]. For each: what they’ve published on this topic in the past 90 days (so I don’t pitch a duplicate), the specific angle-adjustment that would fit their readership, typical word count, and the submission route. Rank by fit, and flag any that require exclusivity.
Exclusivity is the operational trap: most op-ed desks require it while under consideration. Pitch sequentially or in explicitly disclosed parallel — never silently simultaneous.
6.2 The pitch email
PROMPT — Pitch NoteWrite a pitch email to the editor of [PUBLICATION] for this byline. Under 150 words. Structure: one-sentence hook tying the piece to [current news peg]; one sentence on the argument; one sentence on why [NAME] has standing to make it; logistics (word count, exclusive, bio attached). No superlatives about the piece. Subject line options: three, under 60 characters.
6.3 AI disclosure
Norms have hardened: publishers expect authors to remain fully responsible for content and increasingly expect disclosure of AI assistance — and a specific disclosure (what tool, what task, what verification) is materially stronger than a vague one. For trade-press bylines, current practice for most operators: AI-assisted drafting and research with full human verification and rewriting generally doesn’t require unsolicited disclosure, but check each publication’s stated policy before submitting, disclose when asked, and never sign an attestation that the piece involved no AI assistance if it did. For anything academic-adjacent, follow ICMJE-style disclosure: name the tool, the task, and the human verification performed. The evidence log you built in Stage 0.4 is exactly what makes a strong disclosure — or a fact-check response — trivial to produce.
6.4 The package
Final submission bundle: the draft (publication’s format), 3 headline options + dek, 2-sentence author bio (credential-forward, not sales-forward), headshot, disclosure statement if required, and your suggested art or data-viz note if the piece is stat-driven.
Failure Modes — What Actually Goes Wrong #
The generic-thesis trap. Operator skips Stage 2 pressure-testing, ships “AI is transforming [industry] — leaders must adapt.” Editor deletes on the subject line. Fix: if Claude rates your thesis 4+ on the generic scale, you don’t have a byline yet.
The confident fabrication. A stat with a plausible source that doesn’t exist, discovered by the publication’s fact-checker. This is the reputational kill shot of the genre. Fix: the Stage 3.4 protocol, executed without exception, every time, including under deadline.
The stale flagship. Citing last year’s wave of a recurring survey when this year’s dropped last month. Signals the author isn’t actually current in their own field. Fix: recurring reports have predictable release rhythms — check for the newest wave before citing any recurring survey.
The voice collapse. Piece is accurate and well-argued but reads like every other AI-assisted byline this quarter. The principal won’t share it because it doesn’t sound like them. Fix: voice corpus + de-tell pass + read-aloud. If the principal says “I’d never say that,” that sentence — and its siblings — go.
The consulting-framing capture. The byline unconsciously adopts the source report’s commercial framing (“the $X trillion opportunity”) and reads like a McKinsey summary rather than an independent argument. Fix: AACODS Objectivity check — use their numbers, argue your own conclusion, and cite at least one source that complicates the consulting narrative. That’s what the academic layer is for.
The sales-pitch relapse. Product mention creeps in during the principal’s review round. Fix: re-run the red-team prompt on the final-final version, not just yours.
The Operator’s One-Page Checklist #
Setup (once per author): Project created · voice corpus uploaded · voice profile generated · standing instructions set · evidence-log template ready
Per byline:
- [ ] Raw material captured (interview/brain dump) — arguable opinion + first-hand anecdote + defined reader
- [ ] Thesis forged, steelmanned, and rated non-generic
- [ ] News peg identified
- [ ] Evidence sweep run (consulting grey lit + academic layer), source PDFs deep-read
- [ ] Every source AACODS-vetted; every URL clicked; every quote Ctrl-F verified
- [ ] Evidence log complete and dated
- [ ] Outline approved before drafting; evidence/experience pairing in every argument block
- [ ] Draft in voice; de-tell pass; read-aloud pass; principal sign-off on voice
- [ ] Red team, citation audit on final draft, self-promotion scrub
- [ ] Target list built; exclusivity handled; pitch sent; disclosure per publication policy
Sources #
Consulting research hubs: McKinsey Featured Insights · McKinsey Global Institute · BCG Publications · BCG Henderson Institute · Bain Insights · Deloitte Insights
Grey-literature evaluation: AACODS checklist (Flinders University) · UNC — Evaluating Grey Literature · UBC — Evaluating Grey Literature
Academic access: Semantic Scholar · SSRN · NBER · OpenAlex · Lens.org · Unpaywall · DOAJ
Editorial and disclosure norms: PR Daily — How to write the byline editors want · The OpEd Project — Submission Information · ICMJE Recommendations · Purdue Libraries — Publisher AI Policies