Workflows an LLM can write and you can read.

Sly is a tiny JSON-shaped language for orchestrating AI work. It's designed to be read, written, and rewritten by an LLM in the middle of a chat — and to render to a one-glance micro-diagram that a non-engineer can confirm or correct without learning anything new.

{
  "slai":   "1",
  "id":     "morning-triage",
  "intent": "Triage overnight email so I land on a clean inbox with drafts ready for the urgent ones.",
  "start":  "fetch",
  "steps": {
    "fetch": { "do":   "mail.list_unread", "out": "inbox", "next": "scan" },
    "scan":  { "loop": { "over": "inbox", "as": "msg" }, "body": "judge", "next": "done" },
    "judge": { "ask":  "Is {{msg.subject}} urgent?", "out": "verdict", "next": "route" },
    "route": { "if":   "verdict == 'yes'", "then": "draft", "else": "file" },
    "draft": { "ask":  "Draft a reply to {{msg.body}}", "out": "reply" },
    "file":  { "do":   "mail.archive", "with": { "id": "{{msg.id}}" } },
    "done":  { "end":  true }
  }
}

do · ask · if · loop · end

Everything else is glue. Anything imperative lives behind a registered tool that do invokes. The diagram is the spec — every primitive maps to one shape, every reference maps to one edge.

Pronounced like Sly Stone. The schema tag inside every flow is slai ("Simple Loops for AI") — that's the engineer-facing acronym and MCP namespace. Sly is the name; SLAI is the namespace.

The voice is verb-adjective:

• "Sly drafts the flow."
• "A sly recovery proposal."
• "Run it sly."
• "Let Sly take the loop."

The spec is built to be pointed at. Open a Claude Code session and:

Read https://github.com/noslzzp/sly/blob/main/spec.md — it's the Sly v1
spec. Help me build a flow that posts a daily standup to Slack.

You'll get back a valid JSON flow, a mermaid diagram, and an offer to walk a mock run. No SDK, no install, no engine.

For a project that authors Sly flows often, drop this line into your CLAUDE.md:

This project authors Sly workflows. Before drafting one, read
https://github.com/noslzzp/sly/blob/main/spec.md (or pin to a tag for
stability: https://github.com/noslzzp/sly/blob/v1.5/spec.md).

Sly is a spec, not a magic prompt. Any model that actually loads and reads the spec produces compliant flows. Tested May 2026:

• Claude (Sonnet/Opus): point at the URL. First-shot compliant — JSON, mermaid, and walked mock-run.
• GPT (Codex / 4-class): point at the URL. First-shot compliant.
• Gemini: paste the spec content directly into the prompt. URL-fetch in some Gemini surfaces returns hallucinated output (a state-machine DSL Gemini invents instead of Sly). Pasted in-context, Gemini is fully compliant including the §5 mock-run protocol.

The spec is ~17KB of markdown — small enough to drop into any frontier model's context. When in doubt, paste over fetch.

If a model emits anything other than JSON with "slai": "1", it didn't read the spec. Re-prompt with the spec content inline.

Sly's intent field is required and feeds failure-recovery prompts verbatim. When a step fails, the LLM proposes skip / retry / reroute / abort while reasoning about whether a partial run still serves the user's stated goal. Recovery is intent-preserving, not mechanical.

A scan of LLM-workflow systems shipping in 2026 suggests this is the unoccupied seat: most DSLs are JSON-shaped or LLM-authorable or diagrammable, but none elevate intent to a declared field that recovery consults.

v1.5 specification, validated across three frontier vendors (see above). No public reference engine yet — the spec is complete and self-contained for authoring + mock-run via any LLM that supports markdown context. A reference renderer (mermaid → SVG with iconography) and a runtime engine are the next milestones.

Specification text licensed under CC-BY-4.0 — adopters may freely use, adapt, and redistribute the spec provided they attribute the original.

Created by Keith Basil — @noslzzp on GitHub.