#
prompt2mdA Markdown Magic

Open source, Apache-2.0

Fold any text into token‑optimized Markdown

Other tools cut and hope. prompt2md folds: smaller Markdown, honest token numbers, and the byte-exact original back on demand.

Open the studioInstall in your tools
Your prompt
Token-optimized Markdown
Press Fold to run this through the real pipeline.
Runs live against this deployment. Nothing is stored on our side.

↓ Watch it fold, chapter one

332tests, green on Linux & Windows
0telemetry, accounts, or uploads
6tools wired by one command
100%of savings figures reproducible

[ 01 / 06 ]The Fold

Everything else cuts. This folds.

The distinction is not stylistic. It is the difference between a transformation you can undo and one you cannot.

retrieve_original: byte-exact

Scroll ↓ — this is your pasted text.

Fold. Structure and duplication come out. Nothing is discarded.

Unfold. retrieve_original hands back the exact source bytes.

The story is the architecture: the original is stored, content‑addressed, before anything is transformed.

Cutting

  • Truncates to fit a window
  • Drops the middle and hopes
  • Summaries replace the source
  • Savings are estimated, or unstated
  • The detail is gone for good

Folding

  • Structures, dedupes, then summarizes only what is safe
  • Head and tail stay verbatim, because models attend to them most
  • Tables, code, and headings are never summarized
  • Every figure comes from a real run you can repeat
  • retrieve_original returns the exact source bytes

[ 02 / 06 ]The Mechanism

How a document becomes cheap context

Five stages. The interesting one is the third: the pipeline checks its own work and escalates when the cheap engine got it wrong.

  1. Sniff

    Cheap byte-level probes read the content itself, never the file extension.

  2. Route

    Fast path for text-layer PDFs, HTML, Office, CSV. High-fidelity path for scans and complex tables.

  3. Escalate

    The fast path's output is inspected for damage. Degraded tables or low yield trigger a re-run on the heavy engine.

  4. Optimize

    Boilerplate, navigation chrome, signatures, and duplicated passages come out. Structure stays.

  5. Layout

    Stable content first, volatile last, provider-specific cache breakpoints in between.

[ 03 / 06 ]The Proof

Built for context budgets, not demos

Token cost is an output, not a footnote

Every conversion returns a report: tokens in, tokens out, compression ratio, per-section costs, and the effective cost of each repeat call under your provider’s cache pricing. Set a budget and it is enforced, not suggested.

Lossless by construction

Originals are stored content-addressed before anything is transformed. Summarized sections carry p2md:src anchors that resolve to exact bytes.

Dual-engine routing

~0.6 s fast path for most inputs; TableFormer and OCR only when the content proves it needs them.

Cache-aware layout

Sections are ordered so prompt caches hit. Repeat calls can cost a fraction of the first one.

Works with nothing installed

No API key, no sidecar, no account. With zero configuration you get deterministic cleanup and honest numbers; add an LLM gateway and the same pipeline restructures far more aggressively. Text input never hard-fails.

[ 04 / 06 ]The Surfaces

Four surfaces, one pipeline

The same engine behind a command line, an MCP server, an agent skill, and this studio.

Command line
prompt2md convert ./contract.pdf -b 6000
prompt2md compress big-context.md -b 4000
prompt2md batch "docs/**/*.html" -d out/ --watch
prompt2md retrieve "p2md:src=<id>#<start>-<end>"
MCP server
{
  "mcpServers": {
    "prompt2md": {
      "command": "node",
      "args": ["<repo>/packages/hermes-mcp/dist/bin.js"]
    }
  }
}
Agent skill
cp -r packages/skill/prompt2md ~/.claude/skills/

# then, in any conversation:
/prompt2md convert this thread to markdown

[ 05 / 06 ]The Access

One command wires every tool you use

Detects what is installed, backs up every config before touching it, and is safe to re-run. Preview it first with --dry-run.

setup
git clone https://github.com/Hotragn/Prompt2MD.git prompt2md
cd prompt2md && pnpm install && pnpm build
pnpm setup
  • Claude Code
  • Claude Desktop
  • Cursor
  • Windsurf
  • Gemini CLI
  • Codex CLI
  • any MCP client

Model providers are independent of the tool: point P2MD_LITELLM_BASE_URL at any OpenAI-compatible endpoint. Claude, GPT, Gemini, Grok, Kimi, or a local model.

[ 06 / 06 ]The Invitation

See it fold something of yours

Paste a rambling prompt, a contract, or an email thread. You will get clean Markdown, a number you can check, and a way back to the original.

Open the studioRead the source