What Fulminare reads before it starts.
Short playbooks copied into every session: which library to reach for, and how to check its own work.
Generating Word docs
Builds real .docx reports and briefs with python-docx, never a raw markdown file pretending to be a document.
Generating PDFs
Builds real PDFs with reportlab or weasyprint, from scratch or from HTML.
Generating slide decks
Builds real .pptx decks with python-pptx: one idea per slide, charts embedded as images.
Generating spreadsheets
Builds real .xlsx workbooks with openpyxl: formulas, formatting, multiple sheets.
Building web apps
Scaffolds a real React/Next.js app instead of loose HTML files, then verifies it actually renders.
Serving static sites
Serves plain HTML with a two-line Flask app, never python -m http.server, which leaks a directory listing.
Running background tasks
Backgrounds dev servers and slow installs with a task ID, then waits on them properly instead of blocking or polling.
Verifying with browser
Opens what it just built in a real browser and checks it actually works before calling the job done.
Visualizing data
Turns numbers into the right chart type with matplotlib, never defaulting to a bar chart for everything.
Debugging systematically
Reproduces, isolates, and fixes the root cause of a bug, not guess-and-check edits.
The format is the same open SKILL.md standard used by Claude, Codex CLI, Cursor, and others. Fork the repo, add your own, or read the docs for how Fulminare picks a skill mid-task.