Начало работы

Feature Guides

One section per Milly Lab surface: what it is for, when to use it instead of the alternatives, a worked example, the controls, the limits and costs, and what usually goes wrong. For all Milly Lab users.

Each section is written to stand alone, so any of them can be published as a separate page.

#Which surface should I use?

Reach for Chat first. Council and Project cost several times more because they call several models.

#Chat

What it is for

Conversation with one model at a time. The everyday surface — questions, drafting, editing, analysis, code.

When to use it instead

Use Chat unless you specifically need several models. A single good model handles most work at a fraction of a Swarm Mind run's cost.

A worked example

Here is our onboarding email (pasted below). Rewrite it for someone who has never used a developer tool before. Keep it under 150 words, cut the jargon, and end with one clear next step.

Specific about the audience, the length, and the outcome. "Improve this email" would produce something vaguer.

Controls

Markdown and Mermaid diagrams render in responses. Diagram rendering requires a plan that includes it.

Limits and cost

• The whole conversation is resent each turn, so cost grows as it lengthens. Start a new conversation for a new topic — cheaper and clearer.
• Very long conversations exceed the model's context. Use gemini-3.6-flash-200k for genuinely long material.
• Free tier: 10 messages per model per 4-hour window, on two models only.

What usually goes wrong

The answer is confidently wrong. Models fabricate facts, figures and citations. Verify anything that matters. If accuracy is critical, put the same question to a second model — that is what Council Mode is for.

It forgot earlier context. The conversation exceeded the model's window. Start fresh with the key points pasted in, or switch to a large-context model.

It is slow. Check whether web search is on. Turn it off unless you need it.

#Light Studio

What it is for

Generating images from a text description.

When to use it instead

For a picture. If you want a moving image, use Video Studio — generating a still and animating it is not the same thing.

A worked example

A ceramic coffee cup on a windowsill in morning light, shallow depth of field, muted colours, photographed on 35mm film

Subject, setting, lighting, depth, palette, medium. Compare with "a coffee cup", which gives the model no direction at all.

Controls

Switching model clamps your aspect and quality selections to what the new model supports. Check them after switching.

Limits and cost

Each image draws on that model's allowance. Batch multiplies the cost — a batch of four costs four images. The estimated cost is shown before you generate.

What usually goes wrong

It ignored part of the prompt. Image models weight early words more heavily. Put what matters first, and use the negative prompt for exclusions.

Generation was refused. Usually an upstream content-policy refusal. Rephrase.

Nothing like what I imagined. Add specifics — lighting, angle, medium, mood — or supply a reference image.

#Video Studio

What it is for

Generating short video clips from a description.

A worked example

Slow push-in on an empty train platform at dusk, light rain, sodium lamps reflecting on wet concrete, no people

Camera movement, subject, time of day, weather, lighting, plus an explicit exclusion.

Controls

Switching model clamps duration, resolution and audio to what it supports.

Limits and cost

Video is the most expensive thing in Milly Lab. Cost scales with duration and resolution — clips are billed per second on a published ladder, so a 4K clip costs several times a 720p one of the same length.

Generation takes minutes. A queue position is shown. You can leave the page — the job continues and appears in your history.

Plan allowances are small in absolute terms: a plan including video typically covers a handful of clips per month.

Extend produces a new job at additional cost. Remix prefills the composer for a variation.

What usually goes wrong

Stuck at "queued". Video genuinely takes minutes. Leave it.

Controls changed by themselves. You switched model and the selections clamped to what the new one supports.

Audio is greyed out. That model does not generate audio.

#3D Studio

What it is for

Generating 3D assets from a description.

A worked example

A low-poly wooden treasure chest with iron banding, game-ready, neutral pose

State the intended use — game-ready, printable, render-only — because it changes what a usable result looks like.

Controls

The controls are intentionally minimal: describe the object and generate. Milly Lab applies the model's supported settings automatically.

Limits and cost

The scarcest allowance in Milly Lab — a plan including 3D typically covers a very small number of assets per month. Generation is asynchronous; poll or wait for it in your asset list.

Available on Advanced.

What usually goes wrong

The mesh is not production-ready. Generated 3D usually needs cleanup. Treat output as a starting point.

Generation failed. Often an over-complex prompt. Describe one object, not a scene.

#Project Mode

What it is for

Handing over a task with several parts. An orchestrator model breaks it into a dependency graph of subtasks, worker models execute them, and the results are synthesized into a deliverable.

When to use it instead

When the work genuinely has parts that depend on each other and would take several rounds of Chat. For anything a single conversation could handle, Chat costs far less.

A worked example

Review the authentication module in the connected repository. Identify security weaknesses, write a summary of each with severity, and propose a fix for the three most serious. Put the fixes on a new branch.

Several dependent stages — review, then rank, then write — and an explicit instruction about where output should go.

Controls

Limits and cost

The most expensive surface after video. One brief becomes many model calls. A connected repository adds file contents as input on top.

There is a cap on concurrent runs per user. Requires Max or above.

If you disconnect, the run continues server-side and reopening the session replays what you missed.

Working with a connected repository

Read the Connected Workspaces Guide before connecting. In short: file contents go to third-party model providers, and nothing is written to your repository until you press Apply.

What usually goes wrong

It cost far more than expected. Many calls, plus file contents. Check the per-model breakdown in Settings → Billing.

A subtask went off track. Reassign it to a different model, or replace its result, rather than restarting the whole run.

It is slow. Dependent subtasks run in sequence. Progress is shown per subtask.

#Council Mode

What it is for

Several models answering the same brief independently, as separate cards you can compare, diff, debate and merge.

When to use it instead

When you want to compare approaches rather than accept one answer. Good for decisions, designs, and anything where a single model's confident answer should not be trusted on its own.

Project Mode divides work between models. Council Mode duplicates it deliberately.

A worked example

We need to add offline support to our mobile app. Propose an architecture, with the main trade-off you are accepting and what it would cost us to change later.

An open question with several defensible answers — exactly where independent opinions are worth paying for.

Controls

Limits and cost

Cost multiplies by the number of cards. Four models means four full answers to the same brief. Debate rounds add more.

Requires Max or above.

What usually goes wrong

The cards disagree. That is the point. Use the diff to see where and why, and merge the parts you trust.

They all said roughly the same thing. The brief was too closed. Council pays off on open questions, not factual lookups.

#File Generation

What it is for

Producing formatted documents — DOCX, PPTX, XLSX, PDF, Markdown — from a prompt.

A worked example

Create a 10-slide deck introducing our Q3 roadmap to the sales team. One slide per initiative, with the customer benefit as the headline and detail in the notes.

Format, length, audience, and per-slide structure.

Controls

Ask for the document in Chat — state the format (DOCX, PPTX, XLSX, PDF or Markdown), the length and the audience, and Milly Lab produces the file with a download link in the conversation.

Limits and cost

Requires Pro or above. Below that you get a feature_required error.

Generation costs against the model's allowance like any other call.

What usually goes wrong

Refused with feature_required. Your plan does not include file generation.

The structure is not what I wanted. Specify it — number of slides or sections, what goes in each, and who it is for.

The download does not start. Check whether the browser blocked it. If the link expired, regenerate from the conversation.

#Spaces

What it is for

Grouping related conversations and work into a collection.

When to use it

Once you have enough conversations that finding one is annoying. A space per project or client.

Controls

Create, rename and delete spaces; move a conversation into one. Reach them at /app/spaces.

Limits and cost

Free — organising work makes no model calls.

What usually goes wrong

A conversation is not where I expected. Conversations belong to a space only if moved there. Check the unfiled list.

#Teams

What it is for

A shared workspace for several people.

Setting one up

1. Go to Settings → Team (/app/settings/team).
2. Create the team.
3. Invite by email address.
4. The invitee receives an email and accepts through it.

Nothing is created until the invitation is accepted — no account, no membership. So an invitation cannot be used to find out whether an address is already registered.

Controls

Ownership transfer exists so a team is never stranded when its owner leaves.

Limits and cost

How team usage is metered against allowances is covered in Plans, Billing & Usage.

What usually goes wrong

The invitation never arrived. Check spam. Invitations expire — ask the owner to resend.

I accepted but see nothing. Sign out and back in to refresh membership.

Anyone in the team can see shared content. By design. Do not share into a team what should stay private.