Big 4 advisory
The AI Output Verification Pack
The check Deloitte skipped before it sent a 440,000 dollar report to a government, run on your own deliverable, in your own Claude. 9 prompts plus 4 bonus that verify every citation actually exists, trace every quote back to a real source, tie every number to a document, test each claim against what the source really says, sweep for confident-and-wrong lines, recompute the totals, and run a hard self-check gate that blocks the ship if one item is unresolved. The exact pass that would have caught the fabricated citations Deloitte refunded over and the 16 of 27 sources EY pulled a report for.
Rent it forever, or own it once.
The check Deloitte skipped before it sent a 440,000 dollar report to a government, run on your own deliverable, in your own Claude
Path A
Set it up yourself
The setup prompt and 9 task prompts are below — free. The cost is your time, and the risk of wiring it wrong on live data.
Jump to the prompts ↓Path B
We wire it into your business
We would deliver a private verification workspace: the prompts loaded into your Claude tenant, a connector to the sources your deliverables cite (your VDR, filings, market data), the self-check gate wired as a real pre-ship control, and a retained verification log mapped to PCAOB AS 1215, SOX 404, NIST AI RMF, or ISO 42001 so the pass survives edits and your auditor accepts it. Done with you, then handed over so you own it.
Book a build call →Step 1
Paste this setup prompt — Claude installs it for you
Easy mode · paste this into Claude
Claude installs it for you, step by step.
Never used Claude before? It is free to start. Open it in a new tab, copy the prompt, paste it in. It asks one question, then walks you through everything.
- Step 1Open claude.ai ↗
Sign up free. No card. Takes 30 seconds.
- Step 2
One click. Lands on your clipboard.
- Step 3Paste + send
Claude asks what you need + guides you the rest of the way.
▸Tune the prompt for your level (optional)
▸Preview the prompt (you do not need to read it)
I want to INSTALL the real setup so Claude runs the verification pass on my actual deliverable and its real sources, not just a chat demo. Walk me through it step by step, do not skip the install. Treat me like a CFO, audit committee chair, general counsel, or deal lead who has never built a Claude Project or installed an Office add-in, and define every term once.
This is NOT a Terminal or coding install. There is nothing to compile. It is two paths: a private Claude Project where the draft and its source documents live (recommended) and an optional Word or Excel add-in for checking the file in place.
## Path 1 (RECOMMENDED) — a private Claude Project for the deliverable and its sources
This is the path that makes verification real. The check only works if Claude can see the actual sources the deliverable cites, not just the draft.
Walk me through ONE step at a time, waiting for me to confirm each:
1. **What I need.** A Claude account on a Team or Enterprise plan so the data stays in my tenant. Pin **Opus 4.8**. Nothing to install for this path.
2. **Create the verification desk.** In Claude, create a new **Project** named for the deliverable and the date. A Project is a private workspace with its own knowledge that other chats cannot see. Your draft and your sources stay in your own tenant.
3. **Load the material.** This is the step people skip and then wonder why the pass is weak. Drop in BOTH the draft deliverable AND the documents it cites: the source reports, the filings, the data files, the contracts, the cases. Verification compares the draft against these. No sources loaded means the best Claude can do is mark a line UNVERIFIED.
4. **Run the vault.** Run prompt 01, pick what you are verifying (a report, a model, a legal document, a research piece, a board pack) and your data source (A) upload, or (D) a governed connector if you pull from a VDR or filings database. Then run 02 to 09 in order: citation existence, quote trace, number tie-out, claim against source, the confident-and-wrong sweep, internal consistency, the self-check gate, and the verification report.
5. **Gate it.** Treat prompt 08 (the self-check) as a hard gate. If the unresolved count is above zero, the deliverable is NOT cleared to ship until each item is fixed. A named human signs before anything leaves the firm; a prompt cannot sign.
6. **Shadow run first.** Run the pack on two or three deliverables that already shipped before you trust it on live work. It will catch things in work that already went out. That is the proof. See the 90 day rollout in the bundle.
## Path 2 (OPTIONAL) — Claude for Word or Excel add-in for checking in place
If I want the citation, quote, and number checks to run inside the live document:
1. In Word or Excel (Microsoft 365 desktop or web): **Home** or **Insert** tab, **Add-ins** / **Get Add-ins**, search **"Claude"**, **Add**. Tell me where the button is in my version.
2. Sign in to the docked Claude side panel.
3. In vault prompt 01 choose data source (C) the add-in. Run prompt 04 (number tie-out) in Excel on a model, or prompts 02, 03, and 05 in Word on a report, so the checks run against the live file. Keep the source documents in the Project so the add-in has something to verify against.
4. Pin **Opus 4.8** for the verification work.
## Rules for walking me through this
- One step at a time. Define every term once: Project, knowledge, citation, footnote, primary source, fabricated source, paraphrase drift, misattribution, tie-out, recompute, false precision, the self-check gate, partner sign-off, `{{TOKEN}}`.
- Do NOT tell me a step is "not possible." If the add-in is missing, check Word or Excel for Microsoft 365, try the web version, or AppSource. For the Project path there is nothing to install.
- The single most common reason this underdelivers is loading the draft without its sources. Make me load the sources.
- Never paste confidential client material into a public Claude window outside a Project. Use a Project with your tenant controls and follow your firm's data policy.
- This moves the verification work in-house and makes the sign-off defensible. It does not replace the named professional who signs, the subject-matter expert on the substance, or your auditor. Read the audit and compliance overlay before you reduce any external scope. A named human signs. Not legal, tax, or accounting advice.
First message: ask me "Path 1 (a private Claude Project for the deliverable and its sources, recommended) and do you also want Path 2 (the Word or Excel add-in for checking in place)? What are you verifying first (a report, a model, a legal document, a research piece, or a board pack), and do you have the source documents it cites ready to load?" Then start step 1.Step 2
Step 1 installed it. Now run these 9 prompts on your own data.
the vault
The 9 prompts
Tap a prompt to jump to it. Hit copy. Replace the tokens. Paste into Claude Opus 4.7.
where it breaks
Before you connect live data
- • Run dummy data first. Real client data is not a test bed.
- • API keys never go in a public repo. Use env vars and a secrets manager.
- • Add logging, access control, monitoring, and a rollback path before launch.
- • Read the license. Forking a repo without checking is how lawsuits start.
license note
Credit the original author
Prompt set authored by consultance.ai. Deloitte and EY are referenced as the public, documented failures the pack guards against, no affiliation implied. Your draft and sources stay in your own Claude tenant; we never see them. This produces a verification pass and decision-support, not a sign-off; a named human reviews and signs before any deliverable ships, and the relevant professional standards apply. Not legal, audit, accounting, or investment advice.
the newsletter
AI news worth opening.
The AI tools, launches, and shifts that actually matter, in plain English. New library drops the moment they land.
100% freeNo paywall, everUnsubscribe anytime
Want this running in your business, not just your laptop? We build it and hand you the keys.
More like this
Other builds worth a weekend
Big 4 advisory
Setup guide →KPMG Big 4 Claude Prompt Vault
KPMG just rolled Claude out to 276,000 employees in 138 countries. We packaged the 12 Big 4 workflows you can run yourself in 5 minutes.
Big 4 advisory
Setup guide →The Audit Second Read
Last year the largest audit firms were flagged deficient on roughly 1 in 4 audits, double the 2020 rate, with EY worst at 37 percent. This runs a second read on the financials someone already signed, on your own files, in your own Claude. 8 prompts re read the tax provision, test every estimate and reserve, hunt the fraud flags the standard requires, trace related-party and revenue recognition, tie the footnotes back to the statements, run a hard self-check that independently re-derives the key numbers, and write a red-flag memo with a proceed, dig-deeper, or walk read. The questions a careful reviewer asks before trusting the numbers or closing the deal.
Content automation
Setup guide →AiToEarn content agent
The open-source tool that runs an entire content team by itself. One post lands on 14 platforms in 90 seconds. AI replies in your voice.