Worked example: the legibility check, or Claude asks instead of guessing

← All guides·Worked example

When Claude cannot read a word, it should say so and ask. The legibility-review skill makes it do that, one question at a time, before anything goes into the transcript. This page shows three edited extracts from a real intake review.

This is for the documents that PDF intake could not settle: handwriting, faint stamps, signatures, bad photocopies. It is the answer to the referral form with a second applicant written in the margin, the kind of detail that vanishes from every later document if nobody reads it.

What the skill does

The legibility-review skill produces a full, reliable transcript of one document or a whole bundle. The difference from ordinary OCR is that doubtful readings are resolved before they are written down, not after. It works in four steps.

Step 1 — Read everything, and build an index

Claude pulls the text out of every file in the bundle first (typed PDFs directly, scans through PDF intake). For a bundle it builds a cross-reference index: every name, date, case number and amount, and where each one appears. That index is what lets a clean copy in one document settle a doubtful reading in another.

Step 2 — Find the doubtful spots, and settle what it can

Claude goes through each document and lists every spot it is not sure of. Many are settled without troubling you:

  • a smudged case number on a stamped cover page, typed clearly in the heading of the founding affidavit;
  • a faint signature, where the same person’s typed name is in the signature block of three other letters;
  • an unclear handwritten amount that is itemised in a typed schedule elsewhere.

Each of these is logged, not hidden. And if the two copies disagree, Claude does not pick one. A disagreement between two sources is a finding, and it comes to you.

Step 3 — Ask, one question at a time

Three real review questions

These edited extracts retain the relevant pixels from supplied screenshots, with readable editorial summaries alongside them. Names, institutions, document dates, page references and surrounding subject matter are omitted. The full screenshots are not embedded. No answer or completed transcript is shown.

REAL REVIEW · CROPPED IMAGE AND EDITORIAL SUMMARY

Handwriting: a proposed reading

Uncertain handwritten day

The review reports an unusable OCR reading of “Le” and proposes 14, with medium confidence.

Cross-check: The screenshot reports that a stamp on another page supports 14. That page is not shown, so the cross-check cannot be verified from this image.

Do you read the day as 14?

The useful step is putting the image before the reader. A suggested answer is not proof that the reader confirmed it.
REAL REVIEW · CROPPED IMAGE AND EDITORIAL SUMMARY

A number: checking a plausible OCR error

Printed number 13 000

The review reports that OCR read 43 000. The displayed crop reads 13 000; the review assigns high confidence.

Cross-check: The screenshot also shows an arithmetic check: 17 000 + 13 000 = 30 000. That supports the visual reading, but does not establish the truth of the underlying figures.

Approve the correction to 13 000?

This is the clearest example of why extracted text needs checking against the page. The screenshot shows a proposed correction, not completed approval.
REAL REVIEW · CROPPED IMAGE AND EDITORIAL SUMMARY

Degraded print: when context is not enough

Degraded white-on-black reference

The review reports that OCR recovered nothing. It offers 7 of 201[?], with low confidence: the image appears to suggest 2012, while the surrounding table reportedly suggests 2015.

Cross-check: The other rows are not reproduced here. Their reported consistency does not make the doubtful final digit legible.

How should the uncertainty be recorded?

The original interface recommends “7 of 2015 [?]”. That is a recommendation to assess, not a verified reading. Keep the visible reading and the contextual inference distinct, and seek a better copy where the distinction matters.

What is left comes to you, one question at a time. Each question gives the page and place, Claude’s best reading, how confident it is, why it matters, and a cropped picture of the spot so that you are looking at what Claude is looking at. Nothing goes into the transcript until you answer. Every question and answer is logged.

The rules the skill follows:

  • no silent guessing;
  • no silent correction of typing errors: they are copied as they are and marked [sic];
  • no made-up text: an unreadable passage is marked [unreadable];
  • where it is still unsure, it says so, and writes [unclear: best guess “X”].

Step 4 — The transcript and the legibility note

The transcript keeps the original’s paragraph numbers and page markers, and marks handwriting as [handwritten]. At the end, a short note for each document says what you confirmed, what was settled from other documents, what is still unclear, and a reliability rating (high, medium or low) with one sentence of reasoning. The rating is meant to be honest, not “high” by default.

A prompt you can copy

Use it in Claude Cowork with the matter folder connected. Replace the words in square brackets.

Run a legibility review of [folder or file]. First extract the text from every file and build a cross-reference index of names, dates, case numbers and amounts. Then, document by document, list every reading you are not sure of. Settle what you can from a clearer copy elsewhere in the bundle and log it; if two copies disagree, ask me. Ask me the rest one at a time: page, place, your best reading, your confidence, why it matters, and a crop of the spot. Do not write any doubtful text into the transcript until I answer. Mark typos [sic] and unreadable passages [unreadable]. Finish with a legibility note per document with a reliability rating.

Limits

  • It is slow on a bad bundle. One question at a time is the point, but thirty questions is thirty questions. Answer the ones about names, dates, numbers and amounts first.
  • It only sees what it is given. A page missing from the scan is not a legibility problem; check the page count against the index.
  • It depends on you. A careless answer goes straight into the transcript. For a single handwritten form, a guided walk through the form section by section works better than a queue of questions.

Building it yourself

The skill is a written brief of the kind built in Guide 4, with the four steps above as its instructions and the rules above stated as rules. It uses the PDF intake routine for scanned pages.

Or copy this page and use the prompt in Make your own skills: Claude builds a version fitted to your practice.