pdf

Files (4)

• SKILL.md
• reference.md
• fill_form.py · scripts/
• inspect_pdf.py · scripts/

SKILL.md (first lines)

Run the scripts in `scripts/` rather than rewriting them inline. `reference.md`
covers overlaying text on a flat PDF, locating anchors, tables without ruling
lines, watermarks, rendering pages to images, metadata and encryption.

## Always inspect first

A PDF hides the two things that decide your whole approach: whether there is a
text layer at all, and whether it has form fields.

```
python scripts/inspect_pdf.py file.pdf
```

Reports pages, encryption, metadata, extractable text with a sample, form field
names and types, and page geometry. It distinguishes a **scan** (image with no
text) from a merely sparse page, because that difference decides everything.

> If the text layer is empty, the pages are images. There is nothing to
> extract. Say so and offer OCR — never report a scanned document as blank.

## Filling a form

```
python scripts/fill_form.py form.pdf --list                 # field names first
python scripts/fill_form.py form.pdf out.pdf --set name="Ivan" --set city="Moscow"
python scripts/fill_form.py form.pdf out.pdf --json values.json
```

Why a script and not three lines: a filled form that opens **blank** is the
classic outcome. Field values and the drawn appearance are separate things in
the format, so the file needs `NeedAppearances` set or several viewers (Chrome
included) render nothing. The script sets it, writes both to the pages and the
AcroForm root, reads the file back to verify the values stuck, and refuses
unknown field names instead of silently doing nothing.

It also warns when a value contains non-Latin-1 characters: form fonts are
usually Helvetica, so Cyrillic is stored correctly but **drawn as garbage**.
Software sees the right value, a human sees boxes.

## Extracting text and tables

```python
import pypdf
r = pypdf.PdfReader('in.pdf')
text = '\n'.join(p.extract_text() or '' for p in r.pages)
```

`extract_text()` flattens tables into ragged lines. For real tables use
`pdfplumber` where it is available:

```python
import pdfplumber
with pdfplumber.open('in.pdf') as pdf:
    for t in pdf.pages[0].extract_tables():
        for row in t:
            print(row)
```

Always show the user a sample of what you parsed before computing on it —
silently misaligned columns is the classic failure here.

## Merging, splitting, rotating

```python
w = pypdf.PdfWriter()
for f in ['a.pdf', 'b.pdf']:
    w.append(f)
w.write('merged.pdf')

r = pypdf.PdfReader('in.pdf')
w = pypdf.PdfWriter()
for p in r.pages[2:5]:
    w.add_page(p)
w.write('pages3-5.pdf')

page.rotate(90)          # multiples of 90 only
```

## Creating a PDF

Engine decides the tool — read the first line of the RunPython description:

- **PYODIDE** — `fpdf2`. Its core fonts are Latin-1: for Cyrillic you MUST
  embed a Unicode TTF or the text becomes garbage.

```python
from fpdf import FPDF
pdf = FPDF()
pdf.add_font('DejaVu', '', 'DejaVuSans.ttf')
pdf.set_font('DejaVu', size=12)
pdf.add_page()
pdf.multi_cell(0, 8, 'Отчёт за квартал')
pdf.output('report.pdf')
```

- **LOCAL SUBPROCESS** — check for LaTeX first (`shutil.which('pdflatex')`,
  `'xelatex'`, `'tectonic'`). If present, write `.tex` and run it TWICE — the
  second pass resolves references and the table of contents. If absent, fall
  back to fpdf2 and say LaTeX was not found.
- **PODMAN** — `subprocess.run(['tectonic', 'report.tex'])`; it fetches missing
  TeX packages itself.

When the user wants something editable, deliver the `.tex` as well — the app
previews it with highlighting.

## Encryption

```python
w.encrypt(user_password='...')
```

Never invent a password, never add protection that was not asked for, and never
write a password into the chat.

## Verify before claiming success

```python
import pypdf
r = pypdf.PdfReader('out.pdf')

From iaa2005/monet-directory, MIT.

More from iaa2005/monet-directory