extract-text
Skill1 starsChanged 2 months agoMITView on GitHub
Get the words out of a picture or a PDF so they can be worked with as text. Use this whenever someone has something they can see but cannot select or copy: a screenshot of a dashboard or an error, a photo of a page in a book or a whiteboard, a scanned contract or invoice a client emailed, a slide deck sent as images, a folder of phone photos of receipts. It triggers on plain phrases as well as technical ones, including "get the text out of this", "my client sent a scan", "I can't copy from this PDF", "read what this screenshot says", "type up this photo for me", "turn these receipts into a spreadsheet I can use", "what does this say", and "OCR this". It also fits when a PDF refuses to give up its text, or when text copied out of one comes back empty or as gibberish. Do NOT use it for a PDF that already selects and copies cleanly, since reading that one directly is faster and exact. Do NOT use it to tidy up prose that is already text, which is make-it-readable, and do NOT use it to check that some other job really ran, which is prove-it.
- Installs packages
In this repository6 other files it gives its agents.
- first-buildSkill
- grill-my-planSkill
- make-it-readableSkill
- prove-itSkill
- remember-my-rulesSkill
- tailor-my-resumeSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 6 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 32
Before the first run on a machine, check what is installed:
- Very specific: line 63
If brew is not a recognised command on a Mac, Homebrew itself is missing.
- Distinctive: line 70
When a user hits a wall here, say plainly which single program is missing and what it is for.
- Distinctive: line 105
A scanned PDF: Every page is a photograph, so every page has to be rendered and read.
- Distinctive: line 128
Look hard at any flag sitting on a number, a date, a name, an account or an invoice total.
- Very specific: line 131
If the flags are heavy, rerun with --dpi 400, which renders PDF pages larger and often clears up small or faint print.
- Distinctive: line 135
Only reach for it once the uncertain words have actually been checked.
- Distinctive: line 144
When almost nothing comes out, do not guess.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
read this
The useful thing is trapped in a picture. Someone has a screenshot, a photo of a page, a scan a client emailed, and the numbers or the paragraphs inside it are the whole point. This skill lifts them out as text, and it is honest about the parts it could not read confidently, because a wrong figure copied out of a scanned invoice does more damage than no figure at all.
Start here
scripts/read_text.py does the work. It ships with this skill so that nobody has to write the same
helper again, and so that the awkward parts, telling a real PDF from a scanned one, rendering pages
at the right size, scoring how sure the reader was of each word, behave the same way every time.
Before the first run on a machine, check what is installed:
python3 scripts/read_text.py --checkIt names each missing piece and gives the one command that installs it. Read the next section before running anything, because this is the step where people give up.
What has to be installed, and what to do when it is not
Two free programs do the actual work. They are not Python packages and they do not come with Claude Code, so they are installed once and then forgotten.
| Program | What it does here |
| tesseract | Turns pixels into letters. This is the reader. |
| poppler | Handles PDFs. It gives us pdftotext, which lifts text a PDF already has, and pdftoppm, which turns a PDF page into a picture that tesseract can read. One install, two tools. |
# macOS (installs Homebrew first if the brew command is not found)
brew install tesseract poppler
# Ubuntu or Debian Linux
sudo apt install tesseract-ocr poppler-utils
# Windows
winget install UB-Mannheim.TesseractOCR
winget install oschwartz10612.PopplerIf brew is not a recognised command on a Mac, Homebrew itself is missing. Install it with the one
line at https://brew.sh, close the terminal, open a new one, then run the brew install above. The
new terminal matters, because the old one does not know about a program installed after it started.
On Windows, winget sometimes finishes and the tools still are not found, which is almost always the
same thing: open a new terminal window and try --check again before assuming the install failed.
When a user hits a wall here, say plainly which single program is missing and what it is for. Being stuck at an install with no idea what is missing is where this audience stops, and one sentence of explanation is usually the whole fix.
Running it
# one screenshot or photo
python3 scripts/read_text.py dashboard.png --out dashboard.txt
# a PDF of any kind
python3 scripts/read_text.py contract.pdf --out contract.txt
# a whole folder, one text file out per file in
python3 scripts/read_text.py ./receipts/ --out-dir ./receipts-text/
# something other than English
python3 scripts/read_text.py page.jpg --lang deu # German
python3 scripts/read_text.py page.jpg --lang hin+eng # Hindi mixed with English
# small print that came out poorly at the default size
python3 scripts/read_text.py fine-print.pdf --out fine.txt --dpi 400A language pack has to be installed before --lang can use it: brew install tesseract-lang on
macOS, or sudo apt install tesseract-ocr-deu and friends on Linux. --check lists which ones are
already there.
Say which kind of file it was
The script reports this and it is worth repeating back, because the same command can finish in 3 seconds or run for 3 minutes and a user who does not know why assumes something has broken.
- A PDF that already has text. Nothing is read, the text is simply lifted out. Nearly instant, and perfectly accurate.
- A scanned PDF. Every page is a photograph, so every page has to be rendered and read. Budget a second or two per page, more on a big book.
- A mixed PDF. Very common when someone signs a page and scans just that one back into a digital document. The script handles each page the right way on its own and tells you how many went each route.
- A picture. One screenshot or photo, read directly, usually under a second.
So when the wait is long, the reason is that there were 200 photographs of pages in there, not that anything is wrong. Say so while it runs.
Being honest about accuracy
Tesseract scores its own confidence in every word it reads. Any word it scored below 60 comes back
marked [?], and the summary line says how many there were.
Those marks are the point of the skill. A blurry 8 read as a 3 is invisible in clean-looking text, and it stays invisible until somebody acts on it. When a run comes back with flags:
- Say how many there were and what share of the total that is. Eleven flagged words out of nine thousand is background noise. Eleven out of forty means the scan itself is poor.
- Open the picture and look at the flagged spots. Claude can read the image directly, and for a few uncertain words that is the fastest correction there is. Fix them in the text file, and say which ones were changed.
- Look hard at any flag sitting on a number, a date, a name, an account or an invoice total. Those are the ones that cost something when they are wrong. Prose usually survives a bad word. A figure does not.
- If the flags are heavy, rerun with
--dpi 400, which renders PDF pages larger and often clears up small or faint print. For a photo, a straighter, better-lit retake beats any setting.
--no-flags removes the marks when the text is headed somewhere the brackets would be a nuisance.
Only reach for it once the uncertain words have actually been checked.
Finish with something in their hands
End on a real file, named after the thing it came from, in a place they can find, with the full path said out loud. Then say in one line what kind of file it turned out to be, how many words came out, how many were uncertain, and which of those were corrected. If the source was a folder, one text file per original, sitting together in one directory.
When almost nothing comes out, do not guess. A PDF can be locked against copying, a photo can be too dark to read, and a picture of a chart may genuinely have almost no words in it. Say which of those it looks like, and what the next attempt would be.
Others in this pack
- make-it-readable: once the words are out and someone has to read them, this restructures a draft so people finish it.
- prove-it: makes Claude show the receipt for work it says it finished, rather than taking the word for it.
- grill-my-plan: stress-tests a plan before any of it gets acted on.
- remember-my-rules: writes the CLAUDE.md that makes Claude remember these preferences next time.
- first-build: empty folder to one working thing, inside an hour.
tailor-my-resume, read one job ad properly, interview them about the work behind their bullets, and rewrite the resume for that one job.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

