Getting started

prompt2md converts unstructured text, messy prompts, and documents (PDF, Office, HTML, CSV, scans) into token-optimized Markdown — and reports exactly what it saved. Everything below works offline with zero configuration; optional sidecars unlock more capability without ever being required.

Install & build

git clone https://github.com/Hotragn/Prompt2MD.git prompt2md && cd prompt2md
pnpm install && pnpm build        # Node >= 20, pnpm >= 9
pnpm test                         # 90+ tests: golden corpus, MCP integration, CLI

CLI in 60 seconds

# messy prompt -> structured spec (works with zero setup)
node packages/cli/dist/index.js convert --text "your messy prompt here"

# document with a token budget (auto-compresses when exceeded)
node packages/cli/dist/index.js convert ./contract.pdf -b 6000

# batch a folder
node packages/cli/dist/index.js batch "docs/**/*.html" -d out/ --report

# compress oversized context, then recover any detail verbatim
node packages/cli/dist/index.js compress big-context.md -b 4000
node packages/cli/dist/index.js retrieve "p2md:src=<id>#<start>-<end>"

# what's wired up on this machine?
node packages/cli/dist/index.js doctor

Web studio

pnpm --filter @prompt2md/web dev   # http://localhost:3100

Paste, pick a budget and cache profile, and watch the before/after token meter.

MCP connector & skill

claude mcp add prompt2md -- node <repo>/packages/hermes-mcp/dist/bin.js
cp -r packages/skill/prompt2md ~/.claude/skills/prompt2md

In any MCP client, the optimize prompt turns a chat-box paste into optimized Markdown before the model sees it. Full client configs (Claude Desktop, Cursor, Windsurf) and bring-your-own-provider setup: Integrations.

Optional sidecars

Sidecar	Unlocks	Setup
MarkItDown	Fast path for Office/HTML/CSV/simple PDFs	pip install "markitdown[all]"
docling-serve	Scans (OCR), complex tables, multi-column PDFs	docker run -p 5001:5001 quay.io/docling-project/docling-serve then P2MD_DOCLING_URL=http://localhost:5001
LiteLLM	LLM optimizer + summarizer (any provider)	litellm --port 4000 then P2MD_LITELLM_BASE_URL=http://localhost:4000/v1

Without them, conversion never hard-fails on textual input — the pipeline degrades to deterministic cleanup with an explicit warning in the report.

The reliability contract

1. Fast — content probes route each input to the cheapest engine that can handle it; misroutes self-heal by evidence, costing ~0.6s, not minutes.
2. Easy — one runtime, configured by env vars, exposed as CLI, MCP, skill, and web app. doctor tells you what's missing and how to add it.
3. Reliable — 90+ tests pin every routing decision to a golden corpus; CI runs Linux + Windows × Node 20/22 plus browser E2E.
4. Nothing is lost — originals are stored content-addressed before any compression; every summarized section carries a p2md:src anchor whose span returns the byte-exact source text.