The Intelligence Layer

The Copilot is an honest advisor, not an agent that clicks the deck for you. It reads your current prompt, your gallery selection and its metadata, then answers questions about what you've built — analysing a single prompt, comparing two, scoring confidence, or handing a starting seed to a fresh Atelier session. It does not silently mutate the workstation; anything that changes controls you still confirm yourself. In beta, availability is gated per project.

You'll see written analysis, comparisons and confidence scores in the Copilot panel. When you send a seed to the Atelier, a new session opens pre-filled with your prompt — the old session is not disturbed.

Ask for judgement, not settings. 'Which of these two prompts is more coherent?' or 'What's weak about this?' is where the Copilot shines. For setting parameters, use the deck; for deep rework, use the Atelier.

Visual DNA reads a reference image and writes a prompt that captures its style, lighting, framing and material qualities. It's the reverse of normal prompting — you give it the result and it gives you the words.

You get a fully-formed prompt in the current engine's shape, ready to refine. It's rarely perfect on the first pass but usually a strong starting point.

It reads style and structure, not intent. Set the Logic Gate yourself afterwards — the DNA won't know whether you wanted a concept sketch or a finished render, so its result will drift unless you place it in the right design world.

Semantic Heatmap analyses your prompt and colours each phrase by how much it's contributing to the final image. Hot phrases are steering the result. Cold phrases are along for the ride.

Adjectives and specific nouns tend to run hot. Vague filler words tend to run cold. Position matters — words earlier in the prompt usually pull harder.

Use it to find dead words. If a phrase you care about shows up cold, it isn't doing anything — move it earlier or make it more concrete.

Style Interpolation takes two style references and blends them at a ratio you set. Instead of choosing between reference A and reference B, you get a controlled mix of the two.

The blend rarely averages the two references — one usually dominates until you push the weighting past a threshold. Small ratio changes can have big visual effects.

Start at a lopsided blend, not 50/50. An even mix usually produces something muddier than either original.

When you pick an image in the Gallery, the Copilot sees its prompt, parameters, engine target and any locked segments. Ask about 'this' and it means the selection, not a guess. With nothing selected it falls back to your live workstation prompt.

The panel names the source it's reading — the selected render, or the current prompt — so you always know what it's reasoning about.

Select the render you actually care about before asking. A vague question about 'the last one' picks up whatever the workstation happens to be showing, which is rarely what you meant.

Analyse breaks the current prompt down into subject, style, camera, light and composition and calls out what's carrying weight and what's dead weight. Compare takes two prompts — usually the current one and a saved one — and tells you where they diverge and which is tighter.

Written, not visual. The output is a short structured read, not a re-roll. It won't rewrite the prompt for you; it tells you what's actually there so you can decide.

Use Analyse before you Refine and Compare before you commit. The point is to see the prompt clearly, not to have the Copilot fix it invisibly.

The Confidence Check scores your prompt across the five axes that most often decide whether an image lands. Each axis gets its own score and a one-line reason. Low scores are always attributed to specific missing or vague segments, not to overall vibe.

Scores are honest, not flattering. A prompt with no camera language will always score low on camera, even if the subject line is beautiful.

Fix the weakest axis first. Adding more subject detail to a prompt whose real problem is lighting is the classic mistake this check exists to catch.

Send to Atelier opens a new Atelier session pre-filled with your prompt as the opening brief. The Atelier picks a domain from it, seeds the Spine, and starts the Bring stage from there instead of a blank room. Your existing sessions are untouched.

A brand-new session, not an edit of an old one. This is deliberate — the Atelier's memory (the Spine) is per-session, and a seed hand-off is the clean way to move context in.

Run a Confidence Check first. Sending a low-confidence prompt to the Atelier is the fastest way to get a strong second opinion on the axes it flagged.

Learn Mode is the contextual tooltip layer. Every ? next to a control reveals the one-line tip for the matching manual entry, with a link that opens the full chapter. Nothing about the app changes; the reading path just comes to you.

Tips read exactly as the manual does — same tone, same wording. If a tip and the manual disagree, it's a bug in the manual, not in the tip.

Leave Learn Mode on while you're new to a mode. Turn it off once the muscle memory is there — the ? marks are additive UI you don't need forever.

When a meaningful update ships, a small What's New modal surfaces once per user per entry so you don't miss it. The full history is on the /changelog page, newest first, with links back into the manual for anything that gained a new entry.

The modal is quiet — one card, one dismiss, never twice for the same entry. Old entries never re-surface unless a new one ships.

If you're returning after a break, skim /changelog once. The Atelier and the parameter deck evolve steadily and it's the fastest way to see what changed while you were away.