deck-reader

A Claude Code plugin that reads a pitch deck and writes, next to the PDF, a report of what the deck answers and what it does not, with the grid of the stage the deck announces and the block of its business model.

• Pre-seed: completeness. Per grid question, found / partial / absent with the page and the verbatim quote; six completeness bars; the questions to ask the founder; what the deck does not say.
• Seed: completeness, plus verification. Every figure and fact the deck states is listed, looked up in the annexes the founder sent, then, for what a document cannot settle, checked on the public web with sources on both sides. A figure with no document behind it cannot score higher than partial. What is missing goes into a draft email to the founder. What is contradicted goes through a second independent review before the reading stops.
• Series A: the same engine, another question: does the machine repeat? Six documents are required, always the same (monthly P&L over 24 months, cohorts over 12 months or more, CRM export with weighted pipeline, cap table, three-year financial model, contracts of the top 10 customers); a missing one stops the reading with the request email. Unit economics and net retention are the heaviest blocks. New claim types (NRR, weighted pipeline, sales cycle by segment, gross margin against the P&L, top 10 concentration, burn multiple, deals closed without a founder, key hires) and a wider web check (public reviews, open job posts, press of previous rounds, key hires on LinkedIn). Benchmarks for the model and the stage are shown next to every figure, with source and date, and never enter the score.

It never says whether to invest. It says what to ask, and what did not hold up.

One grid per stage, one block per business model, one engine. The deck's announced stage picks the grid (scripts/grids/preseed.json, seed.json, series_a.json). The detected business model picks the model block (scripts/grids/models/<model>.json: saas, marketplace, consumer, ecommerce, hardware, fintech, biotech; no model detected means saas), applied to the stage grid in three verbs: remove, reweight, add. A deck of a stage with no grid is refused with a one-line explanation. Never one grid for all, never one grid per sector.

Usage

/deck-reader path/to/deck.pdf
/deck-reader path/to/deck.pdf --annex revenue.csv --annex cohorts.xlsx --annex model.pdf
/deck-reader path/to/deck.pdf --annex pnl.xlsx --annex cohorts.csv --annex crm.csv --annex cap-table.xlsx --annex model.xlsx --annex contracts.pdf

Output: path/to/deck.reading.md and the same reading as path/to/deck.reading.pdf, and at seed and series A, when documents are missing, path/to/deck.founder-email.md. Add --keep-work to keep the intermediate files in a temporary folder for inspection. Annexes: .pdf, .csv, .tsv, .xlsx, .txt, .md, .json.

To read a grid in one piece, model block applied and benchmarks listed:

python scripts/render_grid.py series_a saas
python scripts/render_grid.py seed marketplace --lang fr --out seed-marketplace.md

The plugin itself is in English. The report is written in the user's language (labels shipped for English and French; other languages get English labels and prose in the user's language). Quotes from the deck and the annexes are never translated.

How it works

Isolated steps, so that a model never fills a gap with something it saw elsewhere, and code takes every decision that must not go through a model.

Every stage

Step	What happens	Who
1. Read	PDF text layer extracted by code; pages transcribed by a model into raw JSON	pdf_text.py, page-transcriber
2. Profile and route	Sector, model type, B2B or B2C, announced stage, each with a quote. The stage picks the grid, the model type picks the model block; no grid, stop	deck-profiler, grid_lib.py
5. Grid	One checker per block in parallel, each sees only its own questions. One question at a time. Every quote is verified by code against the cited page; rejected quotes are retried twice, then marked absent	block-checker × blocks, verify_quotes.py
5c. Confirmation	If the first-pass completeness is between 65 % and 80 %, two more independent passes; each question takes the median	consolidate.py
7. Reading, report	Completeness computed by code from the weights; the writer sees only the answers, never the deck nor the score	score.py, report-writer, report.py

Seed, between the profile and the grid

Step	What happens	Who
3S. Claims	Every verifiable statement of the deck: figures, named customers, competitors, founders, past funding, "why now". Page and quote each, verified by code	annex_text.py, claim-extractor, check_claims.py
4S. Proof in the annexes	Each claim looked up in the documents; the gap between the deck figure and the annex figure is classified by code (minor ≤ 25 %, to probe ≤ ×2, blatant above). First gate: no annex, stop with a generic email; key figures less than half covered, stop with a precise email; otherwise continue	annex-matcher, verify_matches.py, seed_gate.py, founder_email.py
5S. Web	Only for what a document cannot settle. Searched for and against; at least two independent domains to conclude; otherwise unverifiable. The only agent on the web	web-verifier, verify_web.py
6S. Double check	Every blatant contradiction goes to a second reviewer whose only job is to find an honest explanation (date, definition, scope, unit, stale source, homonym). With one, the claim becomes a question for the call. Without one, the second gate stops the reading and shows the sources on both sides	contradiction-reviewer, seed_gate.py
5. Grid, with proof	Checkers cite the claims they rely on. A figure without a proven or confirmed claim is capped at partial; an answer citing a gap to probe is lowered one step. Code, not the checker	apply_proof_cap.py

Series A, between the profile and the grid

Step	What happens	Who
3A. Documents	Each annex sorted into one of the six required documents with a verbatim quote and the months covered; the quote is checked by code. First gate, fixed list: a document missing, or covering fewer months than required, stops the reading with the request email. No coverage threshold	annex_text.py, annex-classifier, series_a_gate.py, founder_email.py
4A. Claims and proof	As at seed, with the series A claim types. Claims the documents do not cover are listed, never a stop. No leftovers email: the document list replaces it	claim-extractor, check_claims.py, annex-matcher, verify_matches.py
5S, 6S	As at seed, with public reviews, open job posts, the press of previous rounds and the LinkedIn profile of each announced key hire added to the web scope	web-verifier, verify_web.py, contradiction-reviewer, series_a_gate.py
7. Report	Benchmarks for the model and the stage shown next to each figure, with source and date. Displayed, never scored	report.py, scripts/grids/benchmarks/

The gate decisions, the gap thresholds, the document list, the source rule, the model blocks and the benchmarks live in scripts/grids/, readable by all, corrected at each post-mortem. They are not hidden in a prompt.

Model blocks

scripts/grids/models/<model>.json, readable copy in skills/deck-reader/grids/models/. Per stage, three verbs applied in this order by grid_lib.effective_grid:

Verb	Effect
remove	question ids of the stage grid taken out
reweight	new weight for a block or a question; a block at weight 0 is for information, never counts, never red
add	questions appended to a block of the stage grid, with its weight, or to a block the model brings

Examples: biotech puts traction at weight 0 and adds milestones, IP and regulatory path; hardware adds margin at 1,000 / 10,000 / 100,000 units, bill of materials and MOQs, and doubles economics; consumer adds DAU/MAU, flattening cohorts and organic share, and lowers revenue; ecommerce adds contribution margin per order after shipping and returns, CAC by channel and 60-day repeat; fintech adds licence, cost of compliance, credit or fraud risk. Marketplace at seed is exactly the previous grid (a test checks it). The pre-seed grid has no model section.

Benchmarks are per model and per stage in scripts/grids/benchmarks/<model>.json, each with value, source and date. An entry with no dated source is left empty, on purpose.

Layout

.claude-plugin/plugin.json
skills/deck-reader/SKILL.md            the orchestrator, routes by stage
skills/deck-reader/grids/preseed.md    the pre-seed grid, readable
skills/deck-reader/grids/seed.md       the seed grid, readable, with gates and thresholds
skills/deck-reader/grids/series_a.md   the series A grid, readable, with the document list
skills/deck-reader/grids/models/       one readable model block per business model
agents/page-transcriber.md             sonnet, every stage
agents/deck-profiler.md                sonnet, every stage
agents/block-checker.md                sonnet, one per block, every stage
agents/report-writer.md                opus, every stage
agents/claim-extractor.md              sonnet, seed and series A
agents/annex-matcher.md                sonnet, seed and series A
agents/annex-classifier.md             sonnet, series A: which required document is each annex
agents/web-verifier.md                 sonnet, seed and series A, the only agent on the web
agents/contradiction-reviewer.md       opus, seed and series A
scripts/grids/preseed.json             grid as data
scripts/grids/seed.json                grid as data, plus claim types, gates, thresholds
scripts/grids/series_a.json            grid as data, plus the fixed document list
scripts/grids/models/<model>.json      model blocks: remove, reweight, add, per stage
scripts/grids/benchmarks/<model>.json  benchmarks per model and stage, value, source, date
scripts/grid_lib.py                    grid by stage, model block assembly, benchmarks
scripts/render_grid.py                 the effective grid of a stage and a model as one markdown
scripts/pdf_text.py                    PDF text per page, stdlib only
scripts/merge_pages.py                 page reference for checkers and verifier
scripts/annex_text.py                  annexes to pages (pdf, csv, xlsx, txt, md, json)
scripts/claim_types.py                 claim types for the extractor
scripts/check_claims.py                claim quotes exist on the cited page
scripts/verify_matches.py              annex quotes exist; gaps classified; annex status
scripts/verify_web.py                  source rule enforced; final status per claim
scripts/seed_gate.py                   the two seed stop decisions
scripts/series_a_gate.py               the series A stop decisions: fixed document list, contradictions
scripts/founder_email.py               the email draft (never sent by the tool)
scripts/grid_block.py                  one block's questions
scripts/verify_quotes.py               quote exists verbatim on the cited page
scripts/apply_proof_cap.py             cap and lowering by proof status
scripts/consolidate.py                 median of independent passes, per question
scripts/score.py                       completeness from values and weights
scripts/report.py                      final markdown, every stage, both languages, benchmarks
scripts/report_pdf.py                  the same reading as a PDF, stdlib only, no font embedded
scripts/claims_lib.py                  shared claim helpers
fixtures/preseed-deck.pdf              12-page fictional pre-seed deck with deliberate gaps
fixtures/preseed-deck.reading.md       its reading, markdown; .reading.pdf = the same, rendered
fixtures/seed-deck.pdf                 14-page fictional seed deck
fixtures/seed-annex-revenue.csv        its revenue export (one minor gap on purpose)
fixtures/seed-annex-cohorts.csv        its cohort table
fixtures/make_preseed_fixture.py       regenerates the pre-seed deck
fixtures/make_seed_fixture.py          regenerates the seed deck and annexes
tests/                                 unittest, stdlib only
tests/fixtures/seed-marketplace-effective.json   the seed + marketplace grid before the model blocks, for the non-regression test

Python 3.9+, standard library only. Nothing leaves the local disk except the web verifier's searches, which carry the claim text and never the deck or the annexes.

Tests

python -m unittest discover -s tests

Guardrails

• "Score" means deck completeness. Never quality, potential or investment score.
• No inference: a figure that is not on a page is absent.
• A value without a page and a verbatim quote is invalid and rejected by code. So is a claim, an annex match, an annex classification or a web source without one.
• A red block produces a question, not a judgement. A contradiction produces the sources on both sides, a review, and a question or a stop, never a verdict.
• The tool can only stop on something false, never on something imprecise. At series A it also stops on a missing document, which is a fact about the file set, not about the company.
• A benchmark is displayed next to the figure with its source and date. It never enters the score, and one without a dated source stays empty.
• The tool never sends the founder email.
• Nothing persists outside the report (markdown and PDF) and the email draft unless --keep-work is given.

Known limits

• Quotes are verified against the PDF text layer. Text that exists only inside images (charts, screenshots) cannot be cited and is treated as absent. When the text layer is unusable (scanned deck), the model transcription becomes the reference and the report header says so. The same applies to annex PDFs.
• Checker values are not perfectly stable between runs on borderline questions; the confirmation passes exist for this reason, inside the 65-80 % band.
• The web check depends on what is public. A customer with no public trace is unverifiable, not contradicted; the report says so and the call clears it up.
• The gap thresholds, the document minimums (24, 12, 36 months, 10 contracts) and the model weights are a first setting. They will be corrected at the first post-mortems, in the grid files, in the open.
• Several benchmarks have no dated source yet (sales cycle by segment, pipeline coverage, share of deals closed without a founder, BOM and MOQ, fintech loss rates, biotech IP and phase durations). They are listed empty rather than invented.
• The grid cannot tell a founder who presents poorly from a founder who did nothing. That is the call's job.