Explore the build library
consultance.ai
Deal Copilot
Build libraryControlsAboutCase studiesBlogBook an audit
consultance.ai

Systems

WorkflowsDeal CopilotPrivate knowledgeFinance automationBuild library

Company

About usCase studiesBlogContact

Start a conversation

Free AI auditfounder@consultance.ai

Teams

Investment bankingCFOs & financePrivate equityPrivate creditReal estateProperty managementProfessional servicesOther industriesAll industries

Legal

Privacy policyTerms of service

Our approach

Audit the workflow.Build in your environment.Hand over a system you own.
Build something useful © 2026 consultance.ai
Build library/Data Room Table Extractor

Finance and data

Data Room Table Extractor

For PE deal teams and corp dev: pull every financial table in a data room into page cited spreadsheets on your own laptop, tie every total, and stop retyping PDFs into Excel.

Technical setup4 steps11 prompts
Open the setup guide Discuss an implementation
Finance and data

Start with the guided setup

Follow the setup below in your own workspace. For a demonstration of the prompt-copying process, open the optional walkthrough.

Watch the setup walkthrough
Step 1 · setup
Three minutes, four steps, nothing to install by hand

Claude sets it up for you. You just paste.

Never used Claude? It is free and takes 30 seconds to open. Copy the instruction below, paste it into Claude, and it reads this page and walks you through everything, one question at a time.

  1. Tell Claude how to talk to you

    One tap. It changes how much Claude explains, and how slowly it goes. You can change it any time.

  2. Copy your setup instruction

    A short instruction plus a link to this page lands on your clipboard. First copy asks for your email once. That unlocks every button across the whole library.

  3. Open Claude in a new tab

    Free account, no card, 30 seconds. This tab stays open so you can come back.

    Open claude.ai
  4. Paste, send, and answer one question

    Claude reads this page, asks one question about your work, then guides you step by step until your first output is right. If anything looks wrong, tell Claude what you see, and it fixes it with you.

▸Prefer the full prompt instead of the link? (optional)
Click to copy
I am comfortable copy-pasting and following instructions, but I am not a developer.
There is nothing to install for this one and no commands to type: it all happens inside Claude. If any instruction below implies a Terminal, translate it into the equivalent click path for me instead.
- Plain English. Define jargon the first time it appears.
- One step at a time, then wait for me to confirm before the next one.
- Tell me what success looks like at each step, and diagnose any error before moving on.

Follow the instructions below with those rules applied.

If you can browse the web, open and read this page in full first, it has the complete guide and every prompt you will run (the vault is under the-vault anchor): https://consultance.ai/library/data-room-table-extractor#the-vault . If you cannot open links, tell me and I will paste the page in, do not guess the prompts.

You are the setup concierge for the Data Room Table Extractor pack from consultance.ai. Your job is to get one deal team member from "I have a downloaded data room" to their first clean, tied P&L table, calmly, one step at a time.

House rules for you:
- Ask ONE question in your first message. Nothing else.
- Define a term the first time you use it. A data room is the folder of documents a seller shares in a deal. Docling is a free, open source program that reads PDFs, decks and spreadsheets on the person's own laptop and pulls out the text and tables. A tie-out means adding the numbers up a second way to check the printed total.
- One step at a time. Wait for them to say it worked before the next step.
- Select Claude Opus 5.5 in the model picker for every unit. Claude Sonnet 5 only for the first bulk read of a very large room.
- Never type a number from memory or from a picture. If a page will not convert, it goes on the list for a person.

First message, exactly this and nothing more:

"Where is the data room right now: already downloaded to a folder on your laptop, still only in the data room website, or a few files you can paste here?"

Then route on the answer:

If A FOLDER ON THE LAPTOP (the normal case): this runs in Claude Code, opened in that folder. It is a Terminal install, one command at a time.
1. Ask them to open Claude Code in the room folder. Success state: Claude Code shows that folder as its working folder.
2. Ask which Python they have: run python3 --version. Docling needs Python 3.10 or newer. Success state: it prints 3.10 or higher. If it is older or missing, stop and say they need Python 3.10 or newer from python.org, or their IT team on a managed laptop.
3. Paste prompt 01 from the vault. Answer its questions: the company, its fiscal year end, the periods, the unit for outputs, and who signs off.
4. Run prompt 02, the fake sample, in this chat first. Match the answer key. If it reports EBITDA of 35,450 or an add back of 4,501, stop and find out why before real data.
5. Run unit 04, room-inventory. Good output is out/inventory.csv and a list of what is missing against the room index.
6. Run unit 05, converter-setup. It installs Docling into a project folder with one pinned version, writes the script, writes a test, and proves it on one real file. Good output is the main P&L table saved with the right page number, and its first three numbers matching the PDF.
7. Then units 06 to 11 in order.

If ONLY IN THE DATA ROOM WEBSITE: ask them to download the room the normal way their platform allows, into one folder, keeping the folder structure. Many platforms forbid automated download in their terms, so do not suggest scripts or scrapers. Then go to the folder route.

If A FEW FILES TO PASTE: use the Claude app in a private Project. Say plainly that the conversion step does not run in chat, so paste the table text or a page at a time. Paste prompt 01, run prompt 02, then run clean-tables and tie-out on each pasted table.

Know these facts and say them when they matter:
- Docling runs on the laptop and does not send the files to a remote service. It downloads its conversion models the first time it runs, so the first run needs the internet unless the models were downloaded in advance (unit 05 covers this).
- Old .xls, .doc and .ppt files need LibreOffice installed to convert. Without it they go on the unread list.
- Scanned pages need OCR, which reads text from an image of a page. If a scanned page comes back empty, unit 10 turns OCR on for that file only.
- Check the NDA allows the files to be processed on the laptop with Claude. The materials go only to the person's own Claude account. Use Team or Enterprise, or turn model training off, before adding confidential files.
Step 2 · run it on your data

Step 1 set it up. These 11 prompts do the work.

the vault

The 11 prompts

Grab the whole pack as one file, or tap any prompt below to copy it on its own. Placeholders that look like {{THIS}} get swapped for your own numbers — and if you ran Step 1, Claude fills them in for you.

One .md file · all 11 prompts, numbered, in order · nothing left out.
Click to copy
DATA ROOM TABLE EXTRACTOR: SHARED RULES
Every unit in this pack follows these rules. They are written here once.

1. WHERE TO RUN IT
- One table or one page you can paste: the Claude app, in a private Project. Chat is fine for prompt 02 and for checking one table.
- A downloaded data room folder: Claude Code, opened in that folder. This is the normal case. The conversion runs on your laptop with the open source Docling library (MIT licence). Docling does not send your files to any remote service.
- A room that must stay on a managed machine: Claude Code on that machine, with the conversion models downloaded once and run offline (unit 05).
- The same room refreshed after each data room update: Claude Code, so every output lands as a file you can compare.
Model: select Claude Opus 5.5 in the model picker. Use Claude Sonnet 5 only for the first bulk read of a very large room.
Privacy: the conversion happens on your machine. What Claude reads goes only to your own Claude account, never to us. Use Team or Enterprise, or turn model training off, before you add confidential material. Check your NDA allows the data room files to be processed this way.

2. BEFORE THE FIRST RUN
First ask which numbers they will rely on most (the EBITDA bridge, net debt, working capital, revenue by customer) and record the answer. Those tables get extracted and tied first. Then ask for these, one at a time:
{{ROOM_FOLDER}} the folder that holds the downloaded room.
{{TARGET}} the company name and its fiscal year end.
{{PERIODS}} the periods that matter (for example FY2024, FY2025, LTM Sep 2026).
{{CURRENCY_UNIT}} the unit the outputs should use (for example USD 000s).
{{TOLERANCE}} the tie tolerance. Default: a difference no bigger than the number of lines summed, in the table's own rounding unit, counts as rounding.
{{QUESTION_THRESHOLD}} default: a difference above 2% of the latest full year reported EBITDA, or 1% of its revenue, is worth a question. Either test is enough.
The named person who signs off the extraction.
A missing item gets a plain default, labelled ASSUMED.

3. STOP RATHER THAN GUESS
- A folder path pasted into the Claude app: this is a Claude Code job. Say so and stop.
- A page that came back as an image with no table, a scan with no text, or a password protected file: name it in the unread list. Do not type the numbers from memory or from a picture.
- A figure needs a document that is not in the room: ask once, name the document, mark it OPEN.
- Two documents disagree and nothing in the normal list explains it: show both with their pages. Do not pick one.
A stop blocks only the outputs that need the missing input. Everything else still runs.

4. LABELS ON EVERY FIGURE
REPORTED (file, page, table id) · CALCULATED (show the arithmetic) · ASSUMED (who set it) · MARKETING (a claim in a CIM, teaser or management presentation, such as an add back or a run rate: it makes a question, never a number in the tie) · NOT IN SOURCE.
A number is REPORTED only if you can point to the file and page it came from (for a spreadsheet, the file, sheet and cell).

5. THREE FLAG TAGS
- CHANGES THE DECISION: a number the human relies on does not tie, or a sign or unit error would change it. It reaches the final summary.
- WORTH A QUESTION: only if the answer could move a figure by more than {{QUESTION_THRESHOLD}}.
- EXPLAINED BY CONTEXT: one line under "Checked, normal". Once closed, it stays closed.
Test every flag against the list below before you raise it.

6. LOOKS WRONG, IS NORMAL
- CIM EBITDA higher than audited profit plus D&A | the CIM shows adjusted EBITDA after add backs | not normal when the add backs are not itemised or the bridge does not close.
- Management accounts full year do not tie to the audited accounts | year end audit adjustments land after the last management pack | not normal when the gap is large, always one way, or no reconciliation exists.
- A total off by 1 | each line was rounded on its own | not normal when the gap is bigger than the number of lines summed, or repeats in every period.
- One file in 000s, another in millions | different authors | not normal when one table mixes units without saying so.
- A prior year figure differs between two sets of audited accounts | a restatement | not normal when no restatement note explains it.
- Budget and actual columns differ side by side | it is a variance table | not normal when the budget column is used as actual elsewhere.
- LTM revenue in the CIM appears in no single document | LTM is built from FY plus year to date | not normal when the year to date figures are not in the room.
- Segment totals do not equal the group total | central costs or eliminations sit on their own line | not normal when no reconciling line exists.
- The same table appears in two files | the CIM copies the management pack | not normal when the copies hold different numbers for the same period.

7. STANCE
A neutral colleague who has extracted many rooms. Do not hunt for problems in the business: this pack checks that the numbers were read correctly and tie, not whether the deal is good. Do not take a printed total as proof.

8. HANDOFF (every unit ends with this)
Read (and not read) · Outputs, with file paths · Checks: pass, fail or not run · Open items, with the document needed · Next unit · Sign-off needed from the named person.
Click to copy
TRY THIS FIRST: A 5 MINUTE TEST ON FAKE DATA
Run units 07, 08 and 09 on this sample in chat, as if Docling had already extracted it. Compare your result to the answer key. If one figure does not match, stop and find out why first. One document is missing on purpose, so you see the pack stop instead of guess.

SAMPLE (Harbor Freight Controls, fiscal year ends 31 December)

FILE A: 03 CIM.pdf, page 14, table 1, "Summary P&L (USD 000s)"
| Line | FY2024 | FY2025 |
| Revenue | 48,200 | 53,900 |
| Cost of sales | (29,100) | (32,400) |
| Gross profit | 19,100 | 21,500 |
| Staff costs | (8,000) | (8,333) |
| Rent | (2,100) | (2,167) |
| Other overheads | (2,600) | (3,451) |
| Total operating expenses | (12,700) | 13,950 |
| Reported EBITDA | 6,400 | 7,550 |
Note under the table (scan of a printed page; the OCR text reads exactly as shown above).

FILE A: 03 CIM.pdf, page 15, table 2, "EBITDA bridge FY2025 (USD 000s)"
| Reported EBITDA | 7,550 |
| Owner salary normalisation | 450(1) |
| One off legal fees | 300 |
| Adjusted EBITDA | 8,300 |
Footnote (1) on page 15: "Owner salary above market rate, removed on completion."

FILE B: 07 Management accounts Dec 2025.xlsx, sheet "P&L", values in USD millions
| Line | FY2025 |
| Revenue | 53.9 |
| EBITDA | 7.55 |

FILE C: 00 Index.pdf, page 1, lists "05 Audited financial statements FY2025.pdf". That file is not in the folder.

ANSWER KEY
| Check | Expected | Working |
| FY2025 total operating expenses | (13,950), a negative | The OCR lost the brackets. 21,500 - 13,950 = 7,550 ties to Reported EBITDA. The trap answer is 21,500 + 13,950 = 35,450. |
| FY2025 opex lines vs total | EXPLAINED BY CONTEXT, rounding | 8,333 + 2,167 + 3,451 = 13,951 vs 13,950. Off by 1 across 3 lines. |
| FY2024 checks | All tie | 48,200 - 29,100 = 19,100. 8,000 + 2,100 + 2,600 = 12,700. 19,100 - 12,700 = 6,400. |
| Owner salary add back | 450, not 4,501 | "(1)" is the footnote marker on page 15, not part of the number. |
| Adjusted EBITDA | 8,300 ties, labelled MARKETING | 7,550 + 450 + 300 = 8,300. Add backs are a CIM claim: they make a question, not a tied number. |
| Add backs | WORTH A QUESTION: ask for support for both add backs | 750 / 7,550 = 9.9%, above the 2% threshold. |
| CIM vs management accounts revenue | Ties, EXPLAINED BY CONTEXT (units) | 53.9 m x 1,000 = 53,900. |
| CIM vs management accounts EBITDA | Ties, EXPLAINED BY CONTEXT (units) | 7.55 m x 1,000 = 7,550. |
| Tie to audited FY2025 | OPEN: ask for "05 Audited financial statements FY2025.pdf" | It is on the index, not in the folder. This blocks only the audited tie. Every other check above still runs. |

A run that reports EBITDA of 35,450, reads the add back as 4,501, flags the 1 rounding gap or the unit difference as a problem, or fills in audited figures, has failed.
Click to copy
<unit name="tools-setup">
Follow the shared rules in prompt 01.
What it does: suggests tools for the jobs in this pack, then installs only the ones the human approves, on the human's own accounts.
Use when: after prompt 02 passes, before the first real room. Skip it if the room is already downloaded to a folder, which is the normal case.
Needs: the "Needs" lines of every unit in this pack, and the systems the human already uses: the data room platform their firm uses and the spreadsheet tool the outputs go to.
Method:
1. See what is already connected. In Claude Code, run `claude mcp list`. In the Claude app, look up the current steps in Anthropic's own help pages with your web search, then ask the human what their Claude already shows as connected.
2. For each need, search the web yourself for an MCP server or connector published by the vendor the human already uses, or listed by Anthropic. Prefer: official publisher, read-only access, sign-in by OAuth, updated in the last 6 months.
3. Show one table: need · tool · publisher · official page URL · read only or read and write · how far it reaches (one folder, one account, or everything the human can open) · what it costs on the human's own plan (write "not stated" if the page does not say) · last update (write "no date shown" if none).
4. For each tool the human approves, open its official page and follow its own steps. A server with a command or URL: copy the command word for word, never from memory. In Claude Code the shape is `claude mcp add --transport http <name> <url>` for a remote server. A connector added in the Claude app: give the steps from the official help page. A page with no command or steps: mark it ASK VENDOR and write the one question for their admin.
5. Install one at a time. Make one small read-only test call and show the result.
Rules: use only the human's accounts and credits. Never ask for a key in chat. Start every tool read-only. Many data room platforms forbid bulk download or automated access in their terms: if the page does not say it is allowed, tell the human to download the room the normal way instead.
<trap>A "PDF to Excel" service that uploads the files to its own servers. It breaks the point of this pack and often the NDA. Leave it out, even when it is free.</trap>
Stop: no official page means do not install it. Mark that need DOWNLOAD and carry on. Every other unit still runs.
Output: the table, what was installed, each test result, and the handoff. Next: room-inventory.
</unit>
Click to copy
<unit name="room-inventory">
Follow the shared rules in prompt 01.
What it does: lists every file in the room, marks which ones hold financial tables, and finds what is missing against the room index.
Use when: first, every time the room is downloaded or refreshed.
Needs: {{ROOM_FOLDER}} opened in Claude Code. The room index, if there is one.
Method:
1. Run ls -R on {{ROOM_FOLDER}} and write the list to out/inventory.csv: path, file type, size, and page count for PDFs.
2. Mark each file: FINANCIAL (P&L, balance sheet, cash flow, EBITDA bridge, budget, KPI pack, debt schedule, customer list), SUPPORTING (contracts, policies), or OTHER.
3. Mark the format risk: digital PDF, scanned PDF (no text layer), legacy .xls, .doc or .ppt (these need LibreOffice), Numbers or Pages (these need an extra install), password protected, or image only.
4. Compare against the index. Every indexed file not in the folder is MISSING. Every financial period in {{PERIODS}} with no source file is a gap.
5. Rank the FINANCIAL files by the numbers the human said they rely on most.
<trap>A file name says "Audited FS FY2025" but the first page says "Draft" or "Management accounts". Read page 1 before you trust the name.</trap>
Stop: a room with no index: say so, and treat the folder as complete for this run. A file that cannot be opened goes on the unread list. Everything else still runs.
Output: out/inventory.csv, the MISSING list, the ranked FINANCIAL list, then the handoff. Next: converter-setup.
</unit>
Click to copy
<unit name="converter-setup">
Follow the shared rules in prompt 01.
What it does: installs Docling in a project folder, writes the extraction script, and proves it on one file before the room is touched.
Use when: once per laptop, then again only if a run fails in unit 10.
Needs: Python 3.10 or newer (Docling requires 3.10 up to below 4.0). Disk space for the conversion models, which download on first use.
Method:
1. Create the tree and write it to disk: mkdir -p extractor/out, then inside extractor create extract.py, tests/test_extract.py and README.txt. Print ls -R to show it.
2. In extractor, create a virtual environment and install the pinned version: python3 -m venv .venv, then .venv/bin/pip install "docling==2.150.1" pandas openpyxl. After it installs, run .venv/bin/pip freeze > requirements.lock so the same versions can be rebuilt.
3. For an offline or managed machine: run .venv/bin/docling-tools models download once while online, note the folder it prints, and pass that folder as artifacts_path from then on. After that the run needs no network.
4. Write extract.py with the documented Docling calls only: DocumentConverter() from docling.document_converter, converter.convert(path) for each file, conv_res.document.export_to_markdown() for the text, and for each table in conv_res.document.tables, table.export_to_dataframe(doc=conv_res.document). Save each table as out/tables/<file stem>-p<page>-t<n>.csv. Take the page from table.prov[0].page_no; if prov is empty (spreadsheets and Word files have no pages), write p0 and record the sheet or section name instead. Write one row per table to out/tables_index.csv: source file, page, table number, rows, columns, first header cell.
5. Write tests/test_extract.py before the first real run: it converts one known PDF from the room and fails if zero tables come back, if any table has no source file in tables_index.csv, or if a page number is missing for a PDF table. Then run .venv/bin/python -m pytest -q tests (install pytest the same way first).
6. First run: .venv/bin/python extract.py on the single most important FINANCIAL file from unit 04. Pass condition: the main P&L table appears in out/tables with the right page number, and its first three numbers match the PDF when you open it.
<trap>Testing on a clean digital PDF and calling it done. Pick the worst file in the room for the first run too: a scanned page. If it returns no table, OCR is off or the page is an image. Fix that here, not after 200 files.</trap>
Stop: Python older than 3.10, or no rights to install software on a managed machine: stop, name it, and ask IT. Nothing else in the pack can run without the converter, except prompt 02.
Output: the tree, requirements.lock, the passing test, the first run result, then the handoff. Next: extract-all.
</unit>
Click to copy
<unit name="extract-all">
Follow the shared rules in prompt 01.
What it does: runs the converter over every FINANCIAL and SUPPORTING file and builds one index of every table, with its file and page.
Use when: after converter-setup passes.
Needs: out/inventory.csv from unit 04, the working extract.py from unit 05.
Method:
1. Run .venv/bin/python extract.py over the FINANCIAL files first, in the ranked order, then the SUPPORTING files. Convert one file at a time and release each result after saving, so one bad file does not stop the rest.
2. Save the text of every file as out/text/<file stem>.md, so later units can read notes and footnotes.
3. Write out/unread.csv: every file or page that failed, returned no text, or returned a page with text but no table where the index or the text says a table should be. Give the reason.
4. Count: files converted, tables found, pages unread. Compare the tables found per file with what a person would expect (a CIM with a financial section and zero tables is a failure, not a clean result).
<trap>A table that runs across two pages comes back as two tables, and the second one has no header row. Look for a table at the top of a page whose first row is numbers, right after a table at the bottom of the previous page. Join them in unit 07; do not treat the second half as a new table.</trap>
Stop: a password protected file: list it as unread and ask for the password or an open copy. A legacy .xls, .doc or .ppt with no LibreOffice installed: list it as unread with the reason. Everything else still runs.
Output: out/tables, out/tables_index.csv, out/text, out/unread.csv and the counts, then the handoff. Next: clean-tables.
</unit>
Click to copy
<unit name="clean-tables">
Follow the shared rules in prompt 01.
What it does: turns raw extracted tables into clean numbers with the right sign, unit, period and source on every cell.
Use when: after extract-all, on the tables that hold the numbers the human relies on, then the rest.
Needs: out/tables, out/tables_index.csv, out/text for the notes and headers.
Method:
1. Headers: rebuild merged header cells so every column has one period label. Write the fiscal year end next to it. Mark LTM, budget, forecast and pro forma columns as such, never as actual.
2. Signs: a number in brackets is negative. Where a bracket may have been lost (a cost line, or a total printed without brackets next to bracketed lines), test the sign against the subtotal it feeds: keep the sign that makes the table add up, and label it CALCULATED with that arithmetic. A sign the arithmetic settles is a correction, not a flag. If the arithmetic cannot settle it, tag it CHANGES THE DECISION.
3. Footnote markers: strip markers stuck to numbers or labels, such as "(1)", "(a)" or "*", and keep the footnote text in a notes column. "450(1)" is 450 with note 1.
4. Units: read the unit from the table title or header ("USD 000s", "in millions"). Convert every figure to {{CURRENCY_UNIT}} in a new column, keeping the original value and unit beside it.
5. Join split tables from unit 06, and mark totals and subtotal rows as TOTAL so they are never summed as data.
<trap>A totals row read as a data row. If "Total operating expenses" is summed with the lines above it, every check after this one doubles. Mark it before any sum runs.</trap>
Stop: a table whose unit is stated nowhere on the page or in the file: label every figure in it NOT IN SOURCE for the unit and ask the human. The other tables still run.
Output: out/clean/<file stem>-p<page>-t<n>.csv with columns: line, period, value, unit, value in {{CURRENCY_UNIT}}, label, source file, page, table, note. Then the handoff. Next: tie-out.
</unit>
Click to copy
<unit name="tie-out">
Follow the shared rules in prompt 01.
What it does: re-adds every subtotal and total a second way instead of trusting what the page printed.
Use when: after clean-tables.
Needs: out/clean.
Method:
1. Down: for every TOTAL row, add the lines above it that feed it. Difference = printed total minus your sum.
2. Across: where a table has a total column or a year to date column, add the periods that feed it.
3. Between statements in the same file: gross profit minus operating expenses equals the EBITDA printed, the EBITDA bridge starts from the reported EBITDA, the balance sheet balances.
4. Judge each difference: zero is TIES. A difference no bigger than {{TOLERANCE}} is rounding, EXPLAINED BY CONTEXT. Anything bigger is a BREAK: check the sign and unit from unit 07 first, then tag it.
<trap>Calling a break "rounding" because it is small. Rounding is bounded by the number of lines summed. A 40 gap in a 10 line table in 000s is not rounding.</trap>
Stop: a total whose lines sit on a page that came back unread: mark the tie NOT RUN, name the page. The other totals still tie.
Output: out/tieout.csv: table, row, printed, recomputed, difference, verdict, working. Then the handoff. Next: cross-document.
</unit>
Click to copy
<unit name="cross-document">
Follow the shared rules in prompt 01.
What it does: lines up the same figure across the CIM, the management accounts, the audited accounts and the model, and explains every gap or names it.
Use when: after tie-out, for the lines the human relies on first, then revenue, gross profit, EBITDA, net debt and working capital for every period in {{PERIODS}}.
Needs: out/clean, out/tieout.csv, out/inventory.csv for what is missing.
Method:
1. Build one grid per key line: rows are periods, columns are sources (CIM, management accounts, audited, model, KPI pack). Every cell carries its file and page.
2. Convert every cell to {{CURRENCY_UNIT}} before you compare. A figure given in millions to 2 decimals is only precise to 5 in 000s: a gap inside that precision ties.
3. For each gap, test it against the normal list in prompt 01: add backs, year end audit adjustments, a restatement, pro forma, LTM build. If one explains it and its own note supports it, tag EXPLAINED BY CONTEXT with that line.
4. A CIM add back, run rate or pro forma figure is MARKETING: show it, bridge it, and make it a question asking for the support. It never becomes the tied number.
5. A gap no normal item explains, above {{QUESTION_THRESHOLD}}: tag CHANGES THE DECISION if it hits a number the human relies on, otherwise WORTH A QUESTION.
<trap>Tying the CIM to itself. The CIM often copies one management table onto two pages. Two CIM pages agreeing is one source, not two. Tie the CIM to a document the seller did not write for marketing.</trap>
Stop: the audited accounts for a period are not in the room: mark that column OPEN and name the file. The CIM to management accounts tie still runs.
Output: out/crossdoc.csv and one grid per key line, then the handoff. Next: gaps-and-debug.
</unit>
Click to copy
<unit name="gaps-and-debug">
Follow the shared rules in prompt 01.
What it does: works through out/unread.csv by the real failure causes, retries what can be fixed, and writes the final list of what still needs a human.
Use when: after cross-document, and any time a run fails.
Needs: out/unread.csv, the run logs, out/inventory.csv.
Method:
1. Scanned page with no text: turn OCR on for that file. Docling's CLI options include --ocr-mode full_page to force OCR on every page (the older --force-ocr flag still works but is deprecated). Re-run that file only.
2. Table missing on a page that has one: re-run that file with --table-mode accurate if fast mode was used. If it still fails, list the page for a human to key in, with the page number.
3. First run fails while downloading models: the machine is offline or blocked. Use docling-tools models download on a connected machine and the artifacts path from unit 05.
4. Legacy .xls, .doc, .ppt: install LibreOffice if policy allows, or ask the seller for the modern format.
5. Out of memory on a very large PDF: convert it in page ranges (the converter accepts a page_range) and join the tables in unit 07.
6. Write out/gaps.md: every page a human must key in, every MISSING document from unit 04, every OPEN item, and every question from unit 09, each with the file and page.
<trap>Re-running the whole room after one fix. Re-run only the files that failed, or the outputs that already tie will change under you.</trap>
Stop: a failure that is not one of the causes above: show the exact error and stop that file. The other files keep their outputs.
Output: out/gaps.md and the updated out/unread.csv, then the handoff. Next: final-check.
</unit>
Click to copy
<unit name="final-check">
Follow the shared rules in prompt 01.
What it does: re-derives the key figures from the clean tables before it reads any summary, then writes the workbook and a one page view.
Use when: last, before anyone uses the numbers.
Needs: out/clean, out/tieout.csv, out/crossdoc.csv, out/gaps.md.
Method:
1. Before reading tieout.csv or crossdoc.csv, recompute the numbers the human relies on most (for example revenue, gross profit and reported EBITDA for each period) straight from out/clean. Compare with the earlier units. If any do not match, stop the release and name the line.
2. Write out/extraction.xlsx: one sheet per clean table, a Tie-out sheet, a Cross-document sheet, a Gaps sheet, and a Read-me sheet with the source of every sheet and the date.
3. Write out/summary.html, one file, from the human's own numbers: counts (files, tables, pages unread), every BREAK and CHANGES THE DECISION item first, then WORTH A QUESTION, then OPEN items, then "Checked, normal".
4. Leave the release unsigned: the named person signs off after reading the summary.
<trap>Reading the earlier summaries first and then "confirming" them. Recompute first, then compare.</trap>
Stop: any key figure that does not recompute: block the workbook's Summary sheet and say which line. The sheets that tie still write.
Output: out/extraction.xlsx, out/summary.html, then the handoff with the sign-off line for the named person.
</unit>

From a working build to an operating system.

For PE deal teams and corp dev: pull every financial table in a data room into page cited spreadsheets on your own laptop, tie every total, and stop retyping PDFs into Excel.

Path A · free

Run the build independently

The setup rail and every prompt above are free and stay free. The cost is your time, and the risk of wiring it wrong on live data.

Back to the prompts
Path B · done with you

Connect it to your operation

The pack extracts and ties one room on your laptop. We wire it into your deal process: every new room converted on arrival, the tie-out and cross-document checks run on each refresh, and the clean tables fed straight into your model template. Reply "wire it" for a 30 minute slot.

Book a build call
data safety

Before you use live numbers

  • • Run last quarter's numbers first. Live data is not a test bed.
  • • Nothing here uploads to us. It runs in your own Claude account, on your own machine.
  • • A named human reviews and signs every output before it reaches a board, lender, or client.
  • • Mask account numbers and names to the minimum the task needs.
the fine print

Straight answers on ownership

Prompt set authored by consultance.ai. Uses Docling (MIT licence, docling-project/docling), installed and run on your own machine; Docling does not send files to a remote service. Your data stays in your own Claude tenant; we never see it. An extraction and tie-out aid, not a quality of earnings review, an audit or investment advice. The worked example is synthetic. A named human signs off before any number is used.

Bring this workflow into your operation.

Book a build callBack to the library

Want this wired into your stack instead of running it yourself? That is our AI deal desk and finance automation service.

Newsletter

New builds, as they ship

One short note when a new build ships: what it does, who it is for, and how to run it in your own Claude.

Free. Three short welcome emails over your first three days, then at most one email a week. Unsubscribe in one click.

More like this

Related builds

All builds
Finance and data
Finance and data

Free Personal Finance App With Claude

For anyone paying for a budgeting app: install a free open source finance app a startup spent $1 million on, connect Claude, and tie every account, transfer and subscription.

Setup guide
Finance and data
Finance and data

AI Deal Diligence Prompt Pack

For PE deal teams and corp dev: read a data room and draft the IC memo in Claude with 11 prompts. Reads every contract, ties the financials to source, flags concentration and change of control, and reconciles before the memo. Replaces the associate grind and a chunk of the third party diligence engagement.

Setup guide
Finance and data
Finance and data

Portfolio Company Reporting Kit

For PE firms to hand to every portfolio company: the monthly package for the sponsor, with management accounts, KPIs, covenants from the credit agreement and 13 week cash, tied before it is sent.

Setup guide
Get the free kitBook a call

Share with your workflow owner.

Bring the relevant team into the conversation with the build and setup guide.

Forward by email
in one line

What is Data Room Table Extractor?

Data Room Table Extractor is a finance and data build in the consultance.ai AI Build Library. For PE deal teams and corp dev: pull every financial table in a data room into page cited spreadsheets on your own laptop, tie every total, and stop retyping PDFs into Excel. It fits PE associates, transaction services analysts and corp dev teams who download a data room and rebuild the P&L, EBITDA bridge, net debt and working capital tables by hand before any model or memo can start. Setup difficulty is Technical, with 4 plain-English steps.

What does Data Room Table Extractor do?

For PE deal teams and corp dev: pull every financial table in a data room into page cited spreadsheets on your own laptop, tie every total, and stop retyping PDFs into Excel.

Who is Data Room Table Extractor for?

It fits PE associates, transaction services analysts and corp dev teams who download a data room and rebuild the P&L, EBITDA bridge, net debt and working capital tables by hand before any model or memo can start.

How hard is Data Room Table Extractor to set up?

Technical to set up — one guided setup instruction covering 4 plain-English steps, plus 11 ready-to-run prompts on the resource page.

How would consultance.ai build this out?

The pack extracts and ties one room on your laptop. We wire it into your deal process: every new room converted on arrival, the tie-out and cross-document checks run on each refresh, and the clean tables fed straight into your model template. Reply "wire it" for a 30 minute slot.

What are the licensing terms?

Prompt set authored by consultance.ai. Uses Docling (MIT licence, docling-project/docling), installed and run on your own machine; Docling does not send files to a remote service. Your data stays in your own Claude tenant; we never see it. An extraction and tie-out aid, not a quality of earnings review, an audit or investment advice. The worked example is synthetic. A named human signs off before any number is used.

Connect this build to your workflow.

Discuss how Data Room Table Extractor fits your systems, controls, and team. We define the integration, review process, and deployment requirements with you.

This build comes from our AI consulting and AI implementation practice — see the full AI in finance guide and how we work with CFO teams.

Book an audit