document-reader

---
name: document-reader
description: Answers your real question about a long document or folder of them, using only quotes from the text, with the gaps left visible. Use for contracts, reports, policies and research packs.
---

# Ask a long document the question you actually have

An answer built from numbered quotes in your own documents, so every line is checkable and any gap shows as a hole.

## Claude does

1. Ask for two inputs, refuse to start without both: the documents and the real question. "Summarise this" is not a question; push for the decision behind it.
2. Size the input. Below roughly 20,000 tokens (about 15,000 words), read straight through. At or above it, Anthropic recommends the quote-first method below.
3. Wrap each document in this exact shape so answers can cite by number:
   `<documents><document index="1"><source>filename</source><document_content>...</document_content></document></documents>`
4. Documents at the TOP of the prompt, question at the BOTTOM. Anthropic measures end-placed queries as worth up to 30 percent better quality on complex multi-document inputs. Formatting instructions go near the end too.
5. Pass one, quotes only: verbatim quotes into `<quotes>` tags, each tagged with its document index. If none are relevant, write "No relevant quotes found" and stop.
6. Pass two: answer from those quotes only, citing each by number. Outside knowledge is banned in this pass.
7. Retraction pass: claim by claim, find the supporting quote; delete any claim without one and mark the hole with empty `[]` brackets.
8. Say it plainly where it applies: "I do not have enough information to confidently assess this."
9. Refuse to guess dates, figures, names, obligations, or anything a document implies but never states.

## Then Codex does

Bulk file work: walking a folder of PDFs, Word files and emails, extracting the plain text, splitting oversized files into ordered chunks, and packaging it all into the indexed XML bundle with a source table. Mechanical, and slow by hand.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills in first: {DOC_FOLDER}, the folder of files; {QUESTION}, the question word for word; {OUTPUT_FILE}, the bundle's path.

## Claude checks

1. Every file appears once, with a unique index and its real filename in `<source>`.
2. Spot-check three quotes character by character against the original. A near-quote is a fabricated quote.
3. No chunk boundary cuts a sentence, clause or table row in half.
4. Nothing summarised, tidied, reordered or de-duplicated. Raw text only.
5. No answer or analysis in the bundle. Codex packages; it does not read.

Any failure: reject, name the file and line, rerun. Never hand-patch the bundle.

## Rules

- Documents top, question bottom, always.
- Quotes first, answer second. Never one pass.
- No supporting quote means the claim is deleted, not softened.
- A visible gap beats a confident sentence.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude bundles the documents itself, in batches, and says so.

## Built from

- Anthropic prompting best practices, long context: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices (30 percent gain, 20k threshold, XML wrapper).
- Anthropic Reduce hallucinations: https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations (quote grounding, `[]` retraction).
- anthropics/skills: https://github.com/anthropics/skills (174,340 stars via api.github.com).
- anthropics/prompt-eng-interactive-tutorial: https://github.com/anthropics/prompt-eng-interactive-tutorial (38,070 stars via api.github.com).

# Ask a long document the question you actually have - Codex task

## You are given

- {DOC_FOLDER} : a folder path holding the source documents. Formats may include .pdf, .docx, .txt, .md, .csv, .eml, .html.
- {QUESTION} : the user's real question, copied word for word. You do NOT answer it. It is here only so you can order the documents sensibly and name the output.
- {OUTPUT_FILE} : the path to write the finished bundle to.

## Produce

One UTF-8 file at {OUTPUT_FILE} containing, in this order:

1. A source table, as a plain Markdown table, one row per document: index, filename, file type, word count, character count, chunk count.
2. The bundle, in exactly this structure, one `<document>` block per document, indexes starting at 1 and never repeating:

```
<documents>
<document index="1">
<source>exact-filename.pdf</source>
<document_content>
...raw extracted text...
</document_content>
</document>
</documents>
```

3. If a document exceeds 20,000 tokens (estimate at 4 characters per token), split it into ordered chunks and give each chunk its own `<document>` block, with `<source>` written as `exact-filename.pdf (chunk 2 of 5)`. Break only at a blank line, a heading or a paragraph end. Never mid-sentence, mid-clause or mid-table-row.
4. A line at the end reading `TOTAL DOCUMENTS: n` and `TOTAL WORDS: n`.

## Rules

- Extract text verbatim. Do not summarise, paraphrase, rewrite, tidy, translate, reorder or de-duplicate anything.
- Keep original spelling, casing, numbers, currency symbols and line breaks. Keep headings and page markers where the extractor emits them.
- Preserve table content as text; do not reformat tables into prose.
- Do not answer {QUESTION}, do not comment on the documents, do not add analysis, opinions or notes anywhere in the file.
- Escape nothing and strip nothing from the source text apart from control characters that break the file.
- If a file cannot be read (encrypted, scanned with no text layer, corrupt), still give it an index and write `EXTRACTION FAILED: <reason>` inside its `<document_content>`. Never skip it silently and never substitute guessed content.
- Do not fetch anything from the internet. Work only on files inside {DOC_FOLDER}.
- Do not modify, move, rename or delete anything in {DOC_FOLDER}. Read only.
- Never invent a fact, a number or a quote.

## Return

Print to stdout, and nothing else:

1. The absolute path of {OUTPUT_FILE}.
2. The source table exactly as written into the file.
3. `TOTAL DOCUMENTS: n` and `TOTAL WORDS: n`.
4. A list of any files that hit `EXTRACTION FAILED`, with the reason.
5. The first 200 characters of the first `<document_content>` block, so the caller can confirm the text is raw and not summarised.