Skip to main content

The assistant

The Assistant builds and edits a blueprint for you. Describe what you want ("a 30-second faceless reel: script, voice-over, one image per scene, then render") and it proposes the stages, wired together and checked, for you to apply.

It runs on Codex or Claude, so you need to connect Codex or connect Claude first. If the one you picked isn't connected, the Assistant tab says so and links to Settings.

Open it​

On the blueprint's canvas, open the Assistant tab in the panel next to the canvas. The first time, you'll see a few example requests. Pick one or type your own and press Enter (Shift + Enter starts a new line).

Messages can contain up to 64,000 characters. This is Reelcraft's per-message limit; the selected model's context capacity also includes conversation history and tool results.

When both are connected, pick Codex or Claude in the provider menu at the top. A chat keeps the provider it started with: switching starts a new chat.

Pick the Model and its Effort next to it. On Codex they start as the model set in your Codex configuration; on Claude, as Claude Sonnet. A higher effort is slower and uses more of your plan.

What it does, and what it can't​

The assistant doesn't guess what Reelcraft can do. Before it uses a stage type, a model, a check, a style, an asset or a character, it looks it up in your Reelcraft, so it only uses things that exist on your computer, in the version you run. You see these steps as a collapsed line such as 13 steps · Proposed a draft. Select it to see each one.

Every proposal is checked with the same rules as Save before you see it. If the assistant gets something wrong, it is told what and fixes it. It never shows you a blueprint that isn't runnable.

It can also read this blueprint's saved versions, so you can ask what changed between two of them or to bring back something from an older one, and it can tell you which model, retries and quality control a stage will really use once your channel and blueprint defaults are merged. It sees your channel's assets, characters and other blueprints, so it avoids name clashes and reuses what exists.

It can read this blueprint's runs too: which stage failed and why, what each cost, which check or quality control rejected an output, and what a stage actually produced, including the pictures. Ask "why did the script stage fail?" or "that run was too expensive, tighten it" and it will look at the run first, then change the blueprint. What it reads from a run is treated as information, never as instructions. It can't start, retry, approve or cancel a run. Pictures and video frames it looks at are sent to the same provider that is running the assistant. The text it reads from a dry run is placeholder content from the fake provider, so it judges those runs by structure only.

When you ask for something Reelcraft can't do, it says so, instead of faking it. For example, Reelcraft can't publish a video to a platform, run stages in parallel, or create characters and assets for you.

It also can't save versions, start runs or approve anything. It only changes the blueprint on the canvas (and its name, description and tags). Save is still yours: see Versions.

Proposals​

A proposal appears as a card with a one-line summary, such as 2 stages added, 1 changed. Select the summary to see which stages, and which of their settings. Any warnings are listed on the card.

  • Apply puts the change on the canvas. It's a normal edit: it shows as Unsaved changes, and you can keep editing or press Save.
  • After applying, a message offers Undo.
  • If you changed the canvas after the assistant started, the card says The canvas changed since this was proposed, and the button becomes Apply anyway. Applying then replaces your edits.

A proposal for the blueprint's name, description or tags works the same way: apply it, or undo.

Auto-apply​

Switch on Auto-apply at the top of the panel and the assistant's valid changes go onto the canvas as soon as they're ready, with an Undo each time. It never applies over edits you made in the meantime: those proposals wait for you, as in manual mode.

Auto-apply is remembered in this browser. It works while the blueprint is open on the canvas.

Questions​

When your request leaves a choice to you, the assistant asks. A question card lists the options, and the last one is always Other…: choose it to type your own answer. Select Send answers to continue. If you type a new message instead, the question is skipped.

Chats​

Each blueprint keeps its chats, so you can come back to one later. Open the ⋯ menu for:

  • New chat: start fresh. Do this when a chat gets long or the assistant loses track.
  • Earlier chats: pick an old one.
  • Delete chat.

Select Stop (the square) to interrupt the assistant while it works.

After you update Reelcraft, a chat started before the update shows Reelcraft was updated since this chat started. Select Start new chat so the assistant uses the new version's tools.

Good to know​

  • The assistant sees this blueprint, its channel's defaults, assets and characters, and your conversation. That text goes to Codex or Claude, under your ChatGPT or Claude plan's terms. Your provider keys and settings never do.
  • It works with the blueprint on the canvas, including edits you haven't saved. While you view an older version, the assistant is switched off.
  • Usage counts against your ChatGPT or Claude plan, and a long chat uses more of it than a short one.
  • If a request fails with a message from Codex or Claude, such as a model that isn't available to your account, pick another Model and send it again.