Skip to main content
Ralphy has three moving parts and one strict rule about how they talk. The agent in your editor decides what to do. The ralphy CLI binary does the work. The workspace on disk is the state. The agent never reaches past the CLI — no raw curl against a provider, no ad-hoc ffmpeg, no yt-dlp outside a ralphy verb. That rule (AGENTS.md hard invariant #2) is what makes the gen-log, the cost rollup, the asset manifest, and cross-session memory work.

​
The three parts

The agent. Whatever editor you use — Claude Code, Cursor, Codex, or GitHub Copilot — the agent has AGENTS.md auto-loaded into its system prompt. That file is the routing table from “user intent” to “playbook.” When you ask “make a video about my coffee shop’s pastry,” the agent matches the intent to a row, reads the matched playbook in docs/playbooks/<role>.md, then acts. It does not improvise on topics the playbook covers — the playbook has the model picks, the prompt scaffolds, and the failure modes. The CLI. ralphy is a single binary built from TypeScript with bun. The verbs live under cli/commands/. The libraries live under cli/lib/. Every verb has the same shape: parse args, read state from the workspace, call a provider through cli/lib/providers/, write artifacts to the workspace, append to the logs. JSON output by default; -p for pretty tables. The CLI is also the only thing on the machine with the API keys — the agent never sees them. The workspace. workspace/ is a gitignored directory in the repo root. Projects, brand definitions, persona definitions, refs, the asset cache, and the generations log all live there. The workspace is treated as canonical-but-wipeable: canonical because every file under workspace/projects/<id>/ is append-only, and wipeable because everything can be regenerated from a brief plus the registry at ~/.ralphy/. See /concepts/workspace for the directory layout.

​
How they talk

┌─────────────┐                ┌─────────────┐                ┌──────────────┐
│             │                │             │                │              │
│   Agent     │ ──invokes────▶ │   ralphy    │ ──mutates────▶ │  workspace/  │
│  (editor)   │                │  (binary)   │                │              │
│             │ ◀──reads logs──│             │ ◀──reads─state─│              │
└─────────────┘                └─────────────┘                └──────────────┘
       │                                                              │
       │  reads playbooks ◀───────────────────────────────────────────┤
       └──────────────────────────────────────────────────────────────┘
                            (docs/playbooks/, AGENTS.md)
The flow has one shape regardless of the task:
1

Agent reads the playbook

The agent matches your intent against the routing table in AGENTS.md, opens the matched playbook in docs/playbooks/<role>.md via the Read tool, and reads any sub-docs the playbook points at (e.g. docs/playbooks/researcher/yt-dlp.md). The playbook tells the agent which ralphy verb to run with which flags.
2

Agent invokes a ralphy verb

The agent runs ralphy <verb> <args> (or bun run ralph -- <verb> in development). The verb is a TypeScript file in cli/commands/. The CLI parses args, picks the model from MODELS.md defaults (or honors --model), and calls the provider through cli/lib/providers/media.ts or cli/lib/providers/llm.ts.
3

ralphy mutates the workspace and writes logs

The verb writes the produced file under workspace/projects/<id>/assets/, appends an entry to workspace/projects/<id>/logs/generations.jsonl with {provider, endpoint, kind, slot, input, output, status, latency_ms, cost_usd}, and updates workspace/projects/<id>/asset-manifest.json to point at the new slot version. If a previous version of the slot exists, the old file becomes scene-03.v1.png and the new one is scene-03.png (auto-archived since commit 753d2f7).
4

Agent reads the logs back

The agent reads generations.jsonl, asset-manifest.json, and (for evaluation) the produced files. The next playbook step uses that state — e.g. the editor playbook reads the manifest to know which scenes are ready for composition.
This loop runs for every action — researching a reference URL, drafting a scenario, generating an image, regenerating a scene, composing the final cut. The agent does not retain “what files exist on disk” in its head; it asks the workspace.

​
The one strict rule

ralphy is the only entry-point for model calls, ffmpeg recipes, yt-dlp pulls, and project mutations. Reaching for bunx tsx against a TS file, curl against any provider API, or ffmpeg ad-hoc — STOP. Either there is a ralphy verb for it (check the playbook’s ## CLI cookbook section), or the operation is not yet covered — in which case propose adding the verb to cli/commands/ and stop. Never paste raw API code into a project. — AGENTS.md hard invariant #2
The reason this rule exists is consequence, not aesthetic. Four downstream systems all depend on every model call flowing through ralphy:
  • The gen-log at workspace/projects/<id>/logs/generations.jsonl. If the agent runs curl against fal.ai directly, the call is invisible. The next session’s agent will not see it, the cost rollup will under-count, and the postmortem will be wrong.
  • The asset manifest at workspace/projects/<id>/asset-manifest.json. If the agent writes a file under assets/ without going through the verb, the manifest gets out of sync with disk. Compositions then reference a slot that does not exist, or a slot that points at the wrong file.
  • The quality gates. ralphy generate image runs scoreImage on the output. ralphy generate video runs scoreVideo. A direct provider call skips the gate and ships a known-bad asset into the cut.
  • The append-only contract. The verbs know to archive <slot>.<ext> → <slot>.v1.<ext> before writing the new file. A direct write overwrites the previous version, which is exactly the failure mode invariant #13 exists to prevent.
When you find yourself thinking “there is no ralphy verb for this, I will just shell out” — that is the bug. The right move is to read the playbook’s ## CLI cookbook section, look for an existing verb, and if there genuinely is no verb, propose adding one to cli/commands/. Most of the verbs in the CLI started as exactly that: a postmortem flagged a missing one.

​
What the agent sees vs. what the CLI sees

The agent has no API keys and no provider knowledge. It sees:
  • The repo (AGENTS.md, CLAUDE.md, MODELS.md, CLI.md, docs/playbooks/).
  • The workspace state (workspace/projects/<id>/...).
  • The output of ralphy verbs (JSON to stdout).
The CLI has both keys (OPENROUTER_API_KEY, ELEVENLABS_API_KEY) and all provider knowledge. It reads:
  • ~/.ralphy/config.json for keys and registry.
  • The workspace for project state.
  • MODELS.md is a documentation source — the CLI does not parse it. Default model picks live in cli/lib/providers/media.ts and are kept in sync with MODELS.md by hand.
This split is deliberate. The agent should not be able to spend money without the CLI mediating. The CLI should not be able to decide what to do — only how to do the next requested thing.

​
The five mandatory reads at session start

docs/playbooks/meta.md names them:
  1. AGENTS.md — auto-loaded by CLAUDE.md.
  2. MODELS.md — checked before every model call.
  3. CLI.md — verb / flag reference cheatsheet.
  4. The closest sibling postmortem under workspace/projects/<id>/postmortem/02-lessons.md. The postmortems are where the high-density “what went wrong, what to do instead” content lives.
  5. The matched playbook from AGENTS.md routing — read fully, then act.
The discipline is: read first, act second. The single most expensive failure mode across 10 postmortems was “the agent felt confident and skipped the read.” A reread costs ~10 seconds. Skipping costs $1-50 in regen burn and 30-90 minutes of user-flagged cleanup.

​
What the workspace looks like during a run

A project mid-flight has roughly this shape: 
workspace/projects/coffee-shop-001/
├── BRIEF.md                     # the user's brief, captured during intake
├── prompts.json                 # the resolved per-scene prompts
├── asset-manifest.json          # slot → file pointer table
├── STORYBOARD.md                # human-readable beat list
├── assets/
│   ├── scene-01-bg.png
│   ├── scene-01-bg.v1.png       # previous version, archived on regen
│   ├── scene-01-clip.mp4
│   ├── voiceover-en.mp3
│   └── music-bed.mp3
├── render/
│   └── final.mp4
└── logs/
    ├── generations.jsonl        # every model call (append-only)
    ├── user-prompts.jsonl       # every user prompt (append-only)
    └── user-assets.jsonl        # every user-uploaded ref (append-only)
Every file in this tree has a producer in cli/commands/ and a schema. See /concepts/projects for the lifecycle, /concepts/generation-log for the JSONL schemas, and /concepts/workspace for the full directory layout including .ralph/ (brands, personas, refs, templates, asset cache).