A programming language for the Intelligent VM.

Last generation, you were a conversant with the model. This generation, you are a tool call to it.

The anatomy. An Intelligent VM is a sufficiently intelligent model plus a harness. The harness needs three things: a loop (the clock), a REPL-like execution environment (the CPU), and a filesystem (memory and disk). Give a capable model those three and it simulates a computer you can program — in English.

You’re probably already programming this computer. Every SKILLS.md, every CLAUDE.md, every well-crafted system prompt is a program in English. Markdown is already the source. Coding agents are already the runtime. The discipline just doesn’t have a name yet.

And those programs are loosely typed. No dependency resolution. No inversion of control. No composition primitives. One file grows to a thousand lines and it starts to fray — the model loses the plot, agents collapse into each other, the same work gets done twice. OpenProse is a subset of Markdown that supplies what’s missing: contracts, shapes, semantic auto-wiring, and a catalog of composition patterns. The same Markdown you already write — with structure where it helps.

---
name: researcher
kind: service
shape:
  self: [evaluate sources, score confidence]
  delegates:
    summarizer: [compression]
  prohibited: [direct web scraping]
---

requires:
- topic: a research question to investigate

ensures:
- findings: sourced claims from 3+ distinct sources, each with confidence 0-1
- sources: all URLs consulted with relevance ratings
- if sources are unavailable: partial findings from cached data, flagged as stale

environment:
- SEARCH_API_KEY: provided by the runtime for web search access

errors:
- no-results: no relevant sources found for this topic

invariants:
- audit log is appended with this invocation's outcome

strategies:
- when few sources found: broaden search terms
- when many low-quality sources: prioritize academic and primary sources

That file is the whole component. Multi-service programs use ## headings to delimit services within a file, or split one service per file in the same directory. Programs can import others from the registry at p.prose.md.

### Execution

let { findings, sources } = call researcher
  topic: question

let evaluation = call critic
  findings: findings
  sources: sources

let report = call synthesizer
  findings: findings
  evaluation: evaluation

return report

Within execution blocks the full imperative grammar is available: parallel:, loop until, for each, try/catch, if/elif/else, choice.

Read the Forme specification →

The manifest is a Markdown file. You can read it, diff it, and check it in. Inspect it to debug; pin it for determinism; regenerate it if the underlying programs change. Single-component programs skip Phase 1 entirely — the .md file is the program.

Token tracking and cost attribution are built in — events capture input/output token counts per LLM call, enabling cost breakdown by phase, service, and depth.

prose status --graph reconstructs the DAG of runs from upstream: headers in state.md, showing which runs consumed which prior runs. Lineage is visible across program boundaries.