Menu
Use for creating, editing, extracting, OCRing, filling, and converting PDF documents.
| name | |
| description | Use for creating, editing, extracting, OCRing, filling, and converting PDF documents. |
from pypdf import PdfReader, PdfWriter
reader = PdfReader("document.pdf")
print(f"Pages: {len(reader.pages)}")
text = "".join(page.extract_text() or "" for page in reader.pages)
Use this for fast inspection. Move to the tool-specific sections below when you need layout-aware extraction, OCR, forms, rendering, or creation.
| Task | Tool | Details |
|---|---|---|
| Create PDF from scratch | ReportLab | libraries/reportlab.md |
| Read / merge / split / rotate / encrypt | pypdf | — |
| Extract text and tables | pdfplumber | libraries/pdfplumber.md |
| Render pages to images | pypdfium2 | libraries/pypdfium2.md |
| Create/manipulate in JavaScript | pdf-lib | — |
| CLI merge/split/encrypt/optimize/repair | qpdf | libraries/cli-tools.md |
| CLI text extraction | pdftotext | libraries/cli-tools.md |
| CLI image extraction | pdfimages | libraries/cli-tools.md |
| CLI page rendering | pdftoppm | libraries/cli-tools.md |
| OCR scanned PDFs | pytesseract + pdf2image | Convert to images, then OCR |
| Fill PDF forms | pypdf or pdf-lib | form-filling.md |
| Convert PDF to Word | pdf2docx (load docx skill) | See docx skill — Converting PDF to Word |
Form filling: You MUST read form-filling.md before attempting to fill any PDF form.
Design defaults: See skills/GENERAL-KNOWLEDGE-WORKER/design-foundations/SKILL.md for palette, fonts + PDF pairings, chart colors, and core principles (1 accent + neutrals, no decorative imagery, accessibility).
Typography: PDFs embed any TTF font — use distinctive, professional fonts, not system defaults. Download from Google Fonts at runtime, register with ReportLab, and it embeds automatically. See libraries/reportlab.md (Custom Fonts section) and skills/GENERAL-KNOWLEDGE-WORKER/design-foundations/SKILL.md (PDF Pairings table + Font Strategy by Format). Default to a clean sans-serif (Inter, DM Sans, Work Sans).
CJK text: Fonts like Inter and DM Sans only cover Latin glyphs. ReportLab has no automatic font fallback — unregistered scripts render as tofu. Register Noto Sans CJK for Chinese, Japanese, or Korean text. See libraries/reportlab.md (CJK Font Support).
Always set metadata when creating PDFs:
"Kortix" unless the user asks for a different organization or author nameCanvas API: c.setTitle(...), c.setAuthor("Kortix") right after creating the canvas.
SimpleDocTemplate: pass title=..., author="Kortix" as constructor kwargs.
pdf-lib (JS): doc.setTitle(...), doc.setAuthor("Kortix").
Every PDF that includes information from web sources MUST have:
<super> tags, never Unicode superscripts)Each footnote entry must include the actual URL wrapped in an <a href> tag — never omit the URL or substitute a plain-text source name. See libraries/reportlab.md (Source Citations) for the implementation pattern.
All URLs in generated PDFs must be clickable. In ReportLab Paragraph objects, use <a href="..." color="blue"> markup. On the canvas, use canvas.linkURL(url, rect). See libraries/reportlab.md (Hyperlinks).
Never use Unicode subscript/superscript characters in ReportLab PDFs. Built-in fonts lack these glyphs, rendering them as black boxes. Use <sub> and <super> XML tags in Paragraph objects. For canvas text, manually adjust font size and y-offset. See libraries/reportlab.md (Subscripts and Superscripts).
Text extraction: pdftotext is the fastest option for plain text. Use pdfplumber when you need tables or coordinate data — don't use pypdf.extract_text() on large documents, it's slow.
Image extraction: pdfimages extracts embedded images directly and is much faster than rendering whole pages. Only render with pypdfium2 or pdftoppm when you need a visual snapshot of the page layout.
Large PDFs: Process pages individually or in chunks rather than loading the entire document. Use qpdf --split-pages to break up very large files before processing.
Encrypted PDFs: Use pypdf to detect and decrypt (reader.is_encrypted / reader.decrypt(pw)). If you don't have the password, try qpdf --password=X --decrypt. Run qpdf --show-encryption to inspect what protection is applied.
Corrupted PDFs: Run qpdf --check to diagnose structural problems, then qpdf --replace-input to attempt repair.
Text extraction fails: If pdfplumber or pdftotext return empty/garbled text, the PDF is likely scanned images. Fall back to OCR (see below).
import pytesseract
from pdf2image import convert_from_path
pages = convert_from_path("scan_output.pdf", dpi=300)
ocr_text = "\n\n".join(
f"--- Page {n} ---\n{pytesseract.image_to_string(pg)}"
for n, pg in enumerate(pages, 1)
)
How to cut a Kortix production release — the versioning philosophy (when patch vs minor vs major) and the exact flow: derive the release title + notes from the FULL git log since the last release, run the Promote workflow, then deploy + verify prod. Load WHENEVER the user wants to release, promote, cut/ship a version, publish a release, or bump the version. The release notes ARE the public /changelog, so they must be 100% accurate to what shipped.
Drive Anthropic's Claude Code CLI (`claude -p`) as a non-interactive coding sub-agent from inside Codex. Use when you want to delegate a coding/analysis/refactor task to Claude, get a second opinion / adversarial review from another model, hand off long-context planning, or fan out parallel agents. Covers the exact `claude -p` flags, permission modes, output capture, JSON/stream-json modes, session resume, parallel fan-out, and the mandatory "delegate → capture → independently verify, never trust the self-report" supervision loop.
Drive OpenAI's Codex CLI (`codex exec`) as a non-interactive coding sub-agent from inside Claude Code. Load WHENEVER you want to delegate a coding/analysis/refactor task to Codex, get a second opinion / adversarial review from another model, fan out parallel agents across files or worktrees, or run a long mechanical job while you stay the planner. Covers the exact `codex exec` flags, sandbox tiers, output capture, JSON/schema modes, session resume, parallel fan-out, and the mandatory "delegate → capture → independently verify, never trust the self-report" supervision loop.
Stage and commit ONLY the work done in this session, on the CURRENT branch — no new branch, no PR, no push. Load whenever the user says 'commit this', 'commit-dis', 'git add commit what we did', 'commit dis', 'save my work', or otherwise asks to commit the changes from the current thread. Enforces: stay on the already-checked-out branch, never overcommit unrelated files.
How to read, write, and curate project memory in `.kortix/memory/` — the project brain. Load this skill whenever you (or the memory-reflector agent) need to add, update, or reorganize what this project knows about itself. Defines the rubric for what belongs in memory, the file structure, and the change-request flow for landing memory edits on `main`.
Canonical reference for a Kortix project. Covers (1) the platform overview — repo-native projects, sessions backed by ephemeral branches, the strict boundary between Kortix config (`kortix.toml`) and OpenCode config (`.kortix/opencode/`) — (2) the in-depth `kortix.toml` manifest with every key, every trigger field, the secrets contract, the `[[apps]]` deployment surface, (3) the full `kortix` CLI reference (every command, every flag, the project-scoped token model, what works inside a session sandbox with the pre-injected `KORTIX_TOKEN`), (4) the Kortix change-request (CR) system — the mandatory path for landing any session-branch work on `main`, including the rule that an agent MUST open a CR if it wants its changes merged, (5) the session sandbox runtime environment — what runs inside a session, including the fact that the sandbox **does support Docker and Docker-in-Docker** (`dockerd`, building images, and running containers/services all work in-session, so agents should never assume a sandbox can't run Do
