The Inner Atelier

Normal prompting is for when you already know what you want and just need to describe it precisely. The Atelier is for the other case — when you're still deciding what the thing actually is, or when the result keeps drifting because two different ideas are fighting inside the same brief. It treats the work as a design conversation rather than a wording exercise.

You'll spend more time per output than a normal prompt, and you'll get a decision trail instead of a lucky roll. The final prompt is a by-product of the reasoning, not the point of it.

Reach for it when you've re-rolled five times and each result solves a different problem. That's a decision problem, and re-rolling won't fix it.

The room has three specialists and the division of labour is strict. **Autocraft ▲** owns form and design logic — proportion, package, surfacing, brand DNA and feasibility; its actions are Diagnose Proportion, Create 3 Architectures, Resolve Brand DNA, Production Reality Check and Remove Clichés, and it seals a Form Covenant. **Lumen ◐** owns light and camera — lens choice, lighting hierarchy, material response and physical plausibility; its actions are Find Hero View, Build Lighting Logic, Create 7-Shot Story, Check Material Readability and Fix Physical Plausibility, and it seals an Image Testament. **Sigma ◼** owns production — prompt purification, translating across models, reference planning and risk auditing; its actions are Purify Prompt, Translate Model, Build Reference Plan, Detect Prompt Conflicts and Open Poster Forge, and it seals a Manifest.

Each master refuses to answer questions outside its layer. Ask Lumen about proportion and it will send you back to Autocraft. That's a feature, not a bug — it's what stops the session collapsing into one blurred opinion.

Go in order — form, then light, then production. Lighting decisions made before the form is settled get thrown away.

The Spine is the session's memory. It has four buckets. **Locked** items are settled and must be preserved. **Active** items are still in play. **Conflict** items are real tensions that need resolving before you move on. **Rejected** items have been killed and must never come back. The whole Spine is sent with every model call in the session, which is why the room doesn't drift and doesn't re-suggest things you already ruled out.

You'll watch it grow as you work. A healthy Spine at the end of a session has a full Locked list, an empty Active list, no unresolved Conflicts, and a Rejected list longer than you thought.

Move things to Locked early and often. An empty Spine gives you a chatbot; a full Spine gives you a collaborator with a memory.

Every step returns Direction cards rather than one answer. Each card carries four actions. **Choose** locks it as the active line and produces sub-directions inside it. **Push Further** keeps the same intent but returns a bolder instance of it. **Make Practical** preserves the intent but adjusts everything else to feasibility. **Show Opposite** returns a legitimate inverted answer to the same brief, argued on its merits — not a strawman.

The alternatives are meant to be defensible, not decorative. If a Direction card feels weak, it usually means the brief upstream is under-specified — go back and tighten it rather than choosing the least bad option.

Run Show Opposite even when you like your current direction. It's the cheapest way to find out whether you actually chose, or just settled.

The masters are chained. Autocraft seals a Form Covenant and hands it to Lumen. Lumen builds on top of it, seals an Image Testament and hands that to Sigma. Sigma compiles both into a Manifest, which is what gets applied back into the workstation. Each handover carries protected fields and any unresolved conflicts, so Lumen cannot undo Autocraft's locked form, and Sigma cannot strip either master's decisions.

The chain is one-way. If Lumen surfaces a problem that needs form to change, you go back to Autocraft, unlock the relevant covenant field, and rerun forward — you don't patch it inside Lumen.

Don't skip straight to Sigma. It compiles what it's given — if nothing upstream was decided, you get a well-formatted version of an undecided idea.

You pick a mode when you open the room. **Consult** — you're starting and want direction. **Critique** — you have something and want it interrogated. **Develop** — you've chosen a line and want it deepened. **Resolve** — you have a specific conflict to settle. **Transform** — you want the same intent taken somewhere else.

Each mode changes the opening question the masters ask and the shape of the first round of Directions. The rest of the session behaves the same.

Pick the mode that matches where you actually are. Opening in consult when you already have a direction wastes the session re-deciding what you'd settled.

FAST runs the masters at their standard reasoning depth — quick enough to iterate, sharp enough to diagnose and to generate the first pass of Directions. DEEP runs them at extended depth, with longer chains of reasoning and more thorough conflict analysis. It's slower and it costs more.

FAST is included from THE_OPERATOR tier upwards. DEEP is available from THE_ARCHITECT with 60 sessions a month, and unlimited on THE_STUDIO.

Diagnose in FAST, resolve in DEEP. Spending a DEEP session on a question FAST could have answered is the most common way people burn their monthly allowance.

The room adapts to the kind of thing you're designing. Seven domains are supported — **automotive**, **industrial and product**, **architecture**, **interior**, **fashion**, **character** and **anime**. The Atelier reads your opening brief, picks a domain with a confidence score, and swaps each master's vocabulary and guardrails to match. If it picks the wrong one, you override it from the domain badge.

You'll see the badge at the top of the room. A low confidence score usually means your brief is ambiguous — either tighten the brief or pick the domain manually.

Check the domain badge before you start working. A fashion brief being read as character will give you sensible answers to the wrong question.

Poster Forge is a sub-workspace inside Sigma for posters and editorial layouts. It handles poster category, information density, layout character, typography zones with explicit hierarchy levels, grid logic, reading sequence and asset planning. The output is a poster system — headline, sub-copy, supporting text, marks and grid — not a single decorative image with words on top.

You give it your real headline and body copy. It plans hierarchy around those actual words, then produces a Manifest that describes the whole layout, not just a background.

Give it your real headline and copy. It plans hierarchy around actual text, and planning around placeholder text produces a layout that breaks when the real words arrive.

Sealing produces the finished artifact for that master — the Form Covenant, the Image Testament or the Manifest — and freezes its protected fields for the rest of the session. Apply takes the final Manifest and writes it back into your main workstation as a prompt, a negatives list, and parameter deck changes. Sessions persist as draft, sealed or applied, so you can leave the room and resume later.

After Apply you're back in the normal workstation with the Atelier's decisions loaded into your prompt. You can keep editing there, or reopen the session and rework it upstream.

Seal each master before moving on, even if you might come back. An unsealed session hands the next master an opinion instead of a decision.

You end with an applied prompt, a complete Spine of locked and rejected decisions, and a session you can reopen and rework any point of without losing the rest.

A first full run takes noticeably longer than a normal prompt — you're doing design work, not typing. Later sessions on similar briefs get faster because you learn which masters to lean on for which problems.

Seal every master before moving on. The single habit that separates a useful session from a wandering one is refusing to hand unsealed work forward.