Conflab User's Guide

What Conflab Is

Conflab is a directed Multi-Provider conversation workspace. You set up a conversation, decide what kinds of participants should be involved, assign personas or models to agent slots, and then steer the discussion as the Director.

The point of using Conflab instead of a single chat window is that you get several distinct voices working the same problem in parallel, under your control. You can pit perspectives against each other, stage a debate, run a planning session, gather expert opinions, or produce a finished deliverable (a summary document, with illustrative images, saved for later).

The current application is organized around four persistent drawer tabs on the left edge of the window:

• SETTINGS
• CONVERSATIONS
• PERSONAS
• RESULTS

On first launch, the main panel can show an overview screen. Once you open a conversation, that same space becomes the live workspace.

Setup

Setup is a one-time job. Do it once and Conflab is ready for any conversation you want to run.

Two things have to be in place before the app is really usable:

1. A root folder so Conflab has somewhere to save workspace data to disk.
2. API keys and a helper model so Conflab's AI features can actually run.

Step 1: Pick a root folder

The root folder is where Conflab stores everything you produce: workspace settings, conversations, personas, and results such as summaries, logs, and images. Without a root folder, there is nowhere for the app to save working state.

Pick a location you can find again. A dedicated folder like Documents/Conflab or similar works well. Everything Conflab writes will land under that folder.

After you choose the folder, Conflab asks you to confirm the workspace choice.

Step 2: Set up OpenRouter (required)

Conflab's AI features run through OpenRouter. The workspace only becomes fully ready once Settings are saved with a valid OpenRouter configuration. That means you need to:

• choose an AI Helper model
• enter an AI API Key - OpenRouter

If the OpenRouter key is missing, most of Conflab's interesting behavior (helper model tasks, persona support, summary generation, image prompt support) will not work.

Why OpenRouter: OpenRouter gives you access to many frontier models (Anthropic, OpenAI, Google, and open-source options) through a single key and account, so you are not locked to one vendor. You buy credit with OpenRouter and spend it across whichever models you want.

Step 3: Set up Brave Search (optional but recommended)

The Web Search API Key - Brave Search field is separate from OpenRouter. Without a Brave key, agents cannot actually use web search even if you later raise the WEB budget on the Director toolbar.

Brave is not required to initialize the workspace, but it is required if you want agents to look things up on the web during a conversation.

When you will want it: whenever you want an agent to cite current information, check a fact, or read a URL you have mentioned. If your conversations will always be self-contained (brainstorming, writing, debate on known material), you can skip Brave for now.

Step 4: Pick a helper model

The AI Helper choice is one of the most consequential settings in the app. The helper model is used for:

• Director Assist (drafting the next Director prompt)
• AI Director (automatic Director behavior)
• persona generation
• conversation summaries
• image-prompt authoring

A free or small helper model will technically work, but the quality of summaries and of AI Director drafts is noticeably better on a strong general-purpose model.

Practical recommendation:

• for the helper model, use a current flagship model from Anthropic or OpenAI. The cost per conversation is modest because helper calls are short compared to the conversation itself.
• for the image model later on, a free model can still produce usable images, so that is a good place to keep costs lower.

Step 5: Save Settings

Click Save Settings. This saves locally and into the workspace folder. On first setup, saving is what completes initialization and unlocks the rest of the app.

Other settings worth knowing about before you start

You can come back to these later, but it helps to know they exist:

• Mission Statement / Global behavior policy: high-level guidance that becomes part of the system prompt agents see. The Edit button opens a larger dialog for editing the policy text.
• Display Timezone: the timezone Conflab uses for displayed timestamps. Match it to your own working context so generated artifacts line up.
• Director Panel Color: reserved for the Director. Separates your instructions from agent output visually.
• Agents per Row and Starting Agent Count: default layout for new conversations. A smaller Agents per Row value means larger panels. Starting Agent Count is how many agent slots appear when you start a new conversation.

Getting Started: Your First Conversation

This walkthrough takes you from an empty workspace to a saved summary and a saved log in roughly ten minutes. It focuses on the basic path: create a conversation, run a few rounds, and save useful results.

By the end you will have:

• one completed conversation
• one saved summary artifact in RESULTS
• one saved log artifact in RESULTS

1. Open the Conversations drawer

Click CONVERSATIONS on the left edge. The drawer opens.

2. Create a new conversation

Click + Conversation. The conversation editor opens.

Fill in:

• Conversation Name: anything memorable. "Launch plan brainstorm" or "Investment trade-off debate" is fine.
• Description / Objective of Conversation: one or two sentences describing what you actually want out of this discussion. Be specific. "Come up with five positioning angles for a new project management tool aimed at freelancers" is better than "marketing ideas."
• Brief description of participants you'd like in this conversation: describe the voices you want around the table, in plain English. "A skeptical CFO, a pragmatic product lead, an early-stage founder who has shipped before, and a freelance user who would actually buy this." This field is not just a note for you: Conflab uses it when auto-filling agent slots with personas.

3. Click Go

Go saves the conversation and drops you into the live workspace with agent slots pre-created.

4. Assign personas to the agent slots

Each agent panel has a header label. Click it to assign a persona or pick a model. You can:

• use an existing saved persona from your Personas drawer
• create a new persona on the fly
• or let the participant-description field above steer auto-fill when that option is offered

For your first conversation, one easy path is to let auto-fill populate the slots from the participants description you wrote, then tweak individual slots as needed.

5. Pick a conversation mode

Look at the Director toolbar in the middle of the screen. Find Mode. For your first session, leave it on Parallel: it is the simplest to understand and the fastest to see results.

(Details on the other modes are in the Conversation Modes section below.)

6. Type your first Director prompt

In the Director input box, type the opening message: the prompt that kicks off the discussion. Keep it concrete. If you wrote a clear Objective in the conversation editor, the opening prompt can be as simple as restating the question.

If you are not sure what to write, click Prompt Assist. Conflab's helper model will either refine what you wrote or, if the input is empty, draft an opening prompt from the conversation objective.

7. Click Submit

All targeted agents respond in parallel. Watch their panels fill in.

8. Run another round or two

Respond to what they said. Ask a follow-up. Push back on a weak idea. You are the Director: the whole point is to steer.

9. Save a summary

When you have enough material, click Summary on the Director toolbar. Conflab uses the helper model to generate a summary of the conversation and saves it as an artifact.

You will see a confirmation and the new summary appears under this conversation in the RESULTS drawer.

10. Save a log

Click Log. This saves a full transcript of the conversation (no AI generation involved: just a clean conversion of the rounds to a readable document).

11. Open your artifacts from RESULTS

Click the RESULTS tab on the left. Expand your conversation. You should see summaries and logs folders with the artifacts you just saved.

Click a summary or a log once to open it in the viewer modal. From the viewer you can:

• Export: save an HTML copy to your downloads folder
• Print: open the browser print dialog, from which you can save as PDF
• Delete: remove the artifact from the workspace

That is a complete first session: you set up a conversation, steered it, and walked away with saved deliverables you can share or revisit. The remaining sections describe the controls, drawers, and workflows in more detail.

The Main Workspace

Once a conversation is active, the center of the app becomes the live conversation workspace.

The main workspace is made of several layers:

• a top bar showing the current topic and root folder
• a history strip near the top
• the Director input box
• the Director toolbar
• the agent grid

Each agent panel is a live participant window. The Director steers the conversation from the center controls, and the individual agent panels show the resulting outputs.

Director Toolbar

The Director toolbar is where most of the live control happens. It includes:

• attachment button
• microphone button
• voice/TTS controls
• WEB
• CONTEXT
• Director: Human / AI
• SPEED
• Mode: Parallel / Sequential / Free Form
• PACE
• Prompt Assist
• + Agent
• column selector
• Summary
• Log
• Image
• Submit

Attachment button

Adds local files to the current Director turn. Those files become part of the context agents can work from.

When to use it: you want the agents to react to specific source material, such as a PDF, a transcript, or a document you are editing, rather than to only what you type.

Microphone button

Dictates speech into the Director input. Useful when you want to think out loud rather than type.

Voice/TTS controls

Control voice playback in the workspace. TTS also appears in persona setup, where individual personas can be associated with a voice.

WEB

Slider that sets the maximum combined web actions available per round for each agent. This budget covers both searches and URL reads.

Important: a web budget alone does not enable web access. Agents only get web tools if all three are true:

• a Brave Search API key is configured
• the budget is greater than zero
• the selected agent model supports tool use

CONTEXT

Sets the rolling number of rounds carried into agent context. Smaller values are faster and cheaper. Larger values preserve more conversation history so agents can refer back to earlier rounds.

How to choose: for short sessions, the default is fine. For longer investigations where continuity matters, raise it. If responses feel expensive or slow, try lowering it first.

Director: Human / AI

Decides whether Director prompts are authored by you or drafted by Conflab's AI Director behavior. See the Director Type section below for when to use each.

SPEED

AI Director countdown. Controls how many seconds Conflab waits before auto-submitting an AI-drafted Director prompt. Only active when Director is set to AI.

Why it exists: gives you a chance to read the draft and intervene before it sends. Short countdowns = fast-paced AI session. Long countdowns = more time to nudge.

Mode: Parallel / Sequential / Free Form

One of the most important controls. Changes how agent turns are scheduled and what they can see. See Conversation Modes below.

PACE

Free-form dispatch delay. Only applies in Free Form mode: it controls how many seconds the system waits before dispatching the next free-form agent turn.

Prompt Assist

Uses the helper model to improve the current Director draft. It is not only a polish tool:

• if you already wrote something, it refines what you wrote
• if the input is empty but the conversation already has completed rounds, it can generate the next Director prompt from context

When to use it: when you know what you want to ask but are not happy with how you worded it. Also when you are between rounds and not sure what to push on next: let the helper propose something based on what just happened.

+ Agent

Adds another agent slot to the conversation. Useful when a discussion could benefit from one more perspective mid-session.

Column selector

Changes how many agent panels appear across each row (for example 1 col, 2 col, 3 col). Fewer columns = bigger panels and less you have to squint at.

Summary

Creates a saved summary artifact for the current conversation. See Summary Workflow below: this button does a bit more than it looks like.

Log

Creates a saved conversation log artifact. See Log Workflow below.

Image

Opens the image-generation workflow for the current conversation.

Submit

Sends the current Director contribution into the conversation.

Director Type: Human vs AI

The Director control deserves real attention because it changes the feel of the app.

Human

In Human mode, you write and submit the Director prompts yourself. This is the straightforward manual mode and works in all conversation modes.

When to use Human: anytime you want full steering control: first sessions, high-stakes decisions, anytime you want to hear exactly what you asked for.

AI

In AI mode, Conflab drafts the next Director prompt for you, places it in the Director input, and starts a visible countdown. If you do nothing, the prompt auto-submits when the countdown expires.

Current AI Director behavior:

• drafts after conversation progress has occurred, not before anything has happened
• uses the Director Assist pipeline internally
• if you start typing, AI waits for you instead of yanking control away
• if you switch from AI to Human, the AI countdown/draft state is cleared immediately
• if you switch from Human to AI while you are already typing, AI waits for you instead of interrupting your draft

AI Director is allowed whenever the conversation has at least one agent.

When to use AI: when you want to observe a discussion rather than steer every turn. Useful for letting a group of personas run a longer debate or brainstorm while you watch and occasionally intervene. It is also a good way to see what kinds of prompts the helper model thinks would advance the conversation: you can learn from its drafts.

Guest input while staying in AI mode

You can leave the conversation in AI Director mode and still type your own Director message when you want to. That means:

• you do not need to flip the Director control back to Human just to add a one-off comment
• you can interrupt with your own instruction, correction, or steering note
• after that human contribution lands, the conversation can remain in AI Director mode for later turns

Why this matters: AI Director becomes practical as an assistant rather than an all-or-nothing lock. Let it drive most of the time; jump in whenever you have something specific to say.

Switching Director type mid-conversation

Switching from AI to Human:

• stops the AI countdown
• clears the AI drafting state
• returns control to you immediately

Switching from Human to AI:

• does not rip the input box away from you if you are already typing
• if you have a draft or your cursor is in the input box, AI waits
• otherwise AI can take the next Director slot and begin drafting

The control is designed as a real runtime handoff, not just an up-front configuration choice.

Conversation Modes

The Mode control changes turn structure, pacing, and what agents can see during a round. Picking the right mode is one of the most effective steering levers you have.

Parallel

Default mode and the easiest to understand.

• Director sends one message
• all targeted agents respond at the same time
• the round finishes after those responses land

Because the responses are dispatched simultaneously, agents do not see each other's current-round responses before they answer.

Use Parallel when you want:

• broad idea generation ("give me five different angles on this problem")
• quick comparison of perspectives where cross-influence would bias the answers
• the cleanest "panel of experts all answer the same prompt" experience
• the fastest mode: everyone speaks at once

Avoid Parallel when: you want agents to react to each other. They cannot, in this mode, within a single round.

If you change away from Parallel during a conversation, the active round keeps the mode it already started with. The new mode applies to later rounds.

Sequential

Agents respond one at a time in a fixed order instead of all at once.

• the Director opens the round
• one agent responds
• the Director can respond again
• the next agent responds
• this continues until the agent list is exhausted

Later agents can react to earlier current-round responses, which changes the quality of the discussion considerably.

Current behavior also allows a blank Director continuation during an in-progress sequential round. That means you can effectively say "continue" without adding new text, and that continuation still matters to the flow.

Use Sequential when you want:

• agents to build on earlier current-round responses
• precise steering between participants (you can inject a comment between every agent)
• a moderated debate, a stepwise analysis, or a chaired discussion
• to hear each voice against a slightly evolving context

Avoid Sequential when: you want unbiased parallel opinions. Later speakers will be influenced by earlier ones.

Free Form

The loosest mode. Designed to feel more like an ongoing conversation than a strict panel response cycle.

• agents are dispatched one at a time
• the app tracks which participants have already spoken in the round
• the round closes after every targeted agent has responded at least once
• PACE controls the delay before the next free-form agent turn is dispatched

Use Free Form when you want:

• a more natural back-and-forth rhythm
• a looser, evolving conversation that feels less like a batch
• pacing control between agent turns (so you can keep up as a human reader)

PACE belongs mainly to this mode. If you are not in Free Form, that control is not the primary pacing mechanism.

Mode cheat sheet

• want fast, broad, independent answers? -> Parallel
• want a moderated debate or staged analysis? -> Sequential
• want an unfolding, natural-feeling group conversation? -> Free Form

You can switch modes mid-conversation. Try Parallel for the first round to gather opening positions, then switch to Sequential for follow-ups where cross-reaction is useful.

Live Tuning During a Conversation

One of the most important things to understand about Conflab is that many controls are intentionally live-editable while the conversation is already underway. This is what makes the app feel like a steering wheel rather than a config form.

You can switch Director type live

Switch between Human and AI at any time. See above for the handoff rules.

You can switch mode live

Switch between Parallel, Sequential, and Free Form at any time. The conversation setting updates immediately, but an already-open round keeps the mode it started with.

You can leave AI Director on and still add your own comments

Add human guest input without abandoning AI Director mode entirely. This is the main reason AI Director is useful as an ongoing mode rather than a one-shot.

You can change agents live

The workspace supports runtime agent management:

• add agents with + Agent
• remove agents from the panel
• mark agents Active or Inactive

At least one agent must remain active.

Active vs Inactive vs Remove: marking inactive keeps the agent and its panel history in place but excludes them from the next round. Removing an agent drops the panel history. Use Inactive when you want to bench a voice temporarily; use Remove only when you are sure you want that history gone.

You can change persona / AI model from the live agent panel

Each agent panel header includes a label button. Clicking it opens the persona and model selector for that agent. This is not just nameplate editing: it is a full swap.

From the live workspace you can:

• open persona selection for an agent that has no persona yet
• open the current persona for an agent that already has one
• assign a different persona
• apply a saved persona
• apply a direct model/voice setup without naming it as a saved persona

Why this is powerful: you can course-correct a conversation by changing who is at the table. If one voice is dominating or being useless, swap it. If a discussion would benefit from an expert you did not invite, add an agent and assign the right persona mid-session. The discussion does not have to be frozen at setup.

Conversations Drawer

The CONVERSATIONS drawer is where you organize and reopen discussions.

Controls:

• + Category
• + Conversation
• conversation tree/list
• root-folder indicator

+ Category

Creates an organizational folder for conversations. Useful once you have enough saved conversations that flat browsing gets annoying.

When to use it: group by project, client, topic, or timeframe: whatever matches how you work.

+ Conversation

Creates a new conversation record and opens the conversation editor.

Conversation list

Browse saved conversations under the current root folder. Click to reopen.

Conversation Editor

Opens when you create or edit a conversation.

Fields and controls:

• Conversation Name
• Description / Objective of Conversation
• Brief description of participants you'd like in this conversation
• Cancel
• Save
• Go

Conversation Name

Saved name of the conversation.

Description / Objective of Conversation

Core prompt-level purpose of the discussion. Frames what Conflab is trying to accomplish. This is not decoration: the helper model reads it, Prompt Assist uses it, and auto-generated summaries reference it.

Write it as: one or two sentences, concrete, with the outcome you want.

Brief description of participants you'd like in this conversation

Plain-English description of the voices you want. Used as planning input when setting up or auto-generating participants.

Write it as: a list of roles or types, each with a short characterization. "A skeptical CFO, a pragmatic product lead, an early-stage founder, and a freelance user" beats "some business people."

Cancel

Leaves the editor without committing changes.

Save

Saves the conversation details without jumping into the run flow.

Go

Saves and proceeds into the live workspace for that conversation.

Lock icon

Some conversation-editor fields become locked after the conversation has recorded rounds. This is by design: changing the objective mid-conversation would retroactively misrepresent what the earlier rounds were doing.

Personas Drawer

The PERSONAS drawer stores reusable participant identities.

Controls:

• + Category
• + Persona
• persona tree/list

Why reuse personas

Building a good persona takes a few minutes of thought. Once you have one that works, with a voice you find useful, that disagrees productively with others, and that has a clear perspective, you will want to reuse it across many conversations rather than rebuilding from scratch.

Saved personas are also how you maintain consistency across sessions. If you want "that skeptical CFO voice" to show up in every planning conversation you run, save her once and assign her each time.

+ Category

Creates an organizational category for personas.

+ Persona

Creates a new saved persona.

Persona list

Reopen existing personas for editing or reuse.

Persona Editor

The persona editor combines model selection, voice, structured identity fields, and a narrative persona description.

Fields and controls:

• AI Setup
• AI Model
• TTS
• Persona Setup
• Occupation / Role
• Gender
• Age
• Name
• Generate Name + Description
• Persona Description
• Worldview
• Life Experiences
• Save
• Clone

AI Setup

Controls the model/voice side of the persona.

AI Model

The model tied to the persona. If you care about response quality for a particular participant, this is another place where using a stronger premium model matters.

Also matters during a conversation because the workspace lets you reopen and change persona / AI model from the live agent panel.

Rule of thumb: the personas you most want to listen to are the ones you should spend model budget on.

TTS

Selects a voice for speech playback. - None (silent) - is also an option. Match voice to persona if you use TTS at all.

Persona Setup

Structured identity section. Can be inactive or active. When active, the fields below contribute directly to the persona description the model sees.

Occupation / Role

Anchors the persona's expertise or social role. This is the strongest signal for the kind of voice you get.

Gender, Age

Core structured identity fields. Shape generational context, life-stage perspective, and the texture of the voice.

Name

Persona's display identity.

Generate Name + Description

Uses the helper model to generate a name and matching persona description from the current structured setup. Good starting point you can then edit.

Persona Description

Narrative backstory / system-prompt-style description. The structured fields and this description are linked: you can generate from structure, edit the description, and re-derive structure when needed.

Worldview

Shapes the persona's perspective rather than just their biography. This is part of why Conflab personas can feel differentiated beyond simple job-title labels: a "skeptical CFO who believes most startups burn out from premature scaling" is very different from a generic "CFO."

Life Experiences

Background and lived-experience detail. Useful for adding specificity that surfaces in how the persona talks.

Save

Saves the persona record.

Clone

Copies the current persona as a variant. Useful when you want related personas ("skeptical CFO" + "enthusiastic CFO" + "operator CFO") without rebuilding each from scratch.

Agent Panels

Agent panels are not passive output windows. Each header contains live controls.

Header controls:

• the persona / model label button
• Active or Inactive
• expand / solo view
• voice playback toggle
• agent log download
• retry for a retryable failure
• remove agent

Persona / model label button

Opens the persona / AI model flow for that agent. Main tool for tuning the conversation while it runs (see Live Tuning above).

Active / Inactive

Controls whether the agent is currently targeted by the conversation. Include or exclude without deleting.

Expand / solo view

One agent fills the workspace for focused reading. Useful when a panel says something long and worth studying.

Voice playback toggle

Turns speech playback on or off for that agent.

Agent log download

Saves that agent's conversation output as a separate artifact into the workspace's saved artifacts. Confirmation popup tells you it saved. This is per-agent: different from the global Log button which saves the full conversation.

When to use it: when you want a single voice's contribution as a standalone document, rather than the whole discussion.

Retry

Appears when an agent has a retryable failure in the latest round. Click to retry that agent's turn without rerunning the entire round.

Remove agent

Removes the agent from the conversation. Panel history is lost, so this is more destructive than simply marking inactive. Prefer Inactive if you might want that voice back.

File Attachments

The Director toolbar attachment button opens a file picker.

Select one or more local files to bring into the conversation.

When to use it:

• feed a document you want agents to analyze
• attach source material the discussion should react to
• give the agents a transcript, a PDF, an article, a data file, or a draft you are revising

Attachments become part of what agents can reference. Combine this with specific Director instructions ("use only the attached document; do not speculate beyond it") for constrained analysis.

Results Drawer: Getting Things Out

The RESULTS drawer is the persistent browser for saved artifacts. It is labeled RESULTS but it is effectively the output inventory for the workspace.

The tree groups artifacts under each conversation. The main types you will see:

• summaries: generated summary documents
• logs: full conversation transcripts
• images: generated images tied to conversations

How to open an artifact

Single-click on any artifact row to open it in the viewer modal.

What the viewer lets you do

From the viewer modal, the controls are consistent across summaries, logs, and images:

• Export: save a copy to your browser's default downloads folder. Summaries and logs export as HTML files. Images export as JPG.
• Print: opens the browser's native print dialog, which includes a Save as PDF option. This is the path to a PDF deliverable.
• Delete: removes the artifact from the workspace. Confirmation is required.

What "Export" actually does: it writes a standalone copy of the artifact to your computer's downloads folder. The file is self-contained: you can email it, attach it, check it into a repo, or archive it. No picker is shown; it uses your browser's default download behavior.

Note on images in summaries: if a summary has an associated image (via Use in Summary in the image dialog), the image is referenced from workspace storage and appears when you open the summary inside Conflab. If you Export the summary HTML and open it outside Conflab, the image will be missing because it lives in workspace storage, not inside the HTML file. For a self-contained PDF with image included, use the Print -> Save as PDF path while the summary is open in the viewer.

How to archive a whole conversation

A complete archive of one conversation typically means:

1. open the summary, Export or Print -> Save as PDF
2. open the log, Export or Print -> Save as PDF
3. open each relevant image, Export (or use Print -> Save as PDF if you want it paginated with the summary)

All of those land in your browser's downloads folder. Move them into whatever archive location you use.

Summary Workflow

The Summary button on the Director toolbar is not quite a one-click save: it is a small decision point.

How the Summary button behaves

• If you have not selected a specific round range, it generates a full-conversation summary and saves it directly.
• If you have selected an intermediate range of rounds (for example you highlighted rounds 3-5 in the history strip), the button shows a small menu with three choices:
• Full Summary: summarize the entire conversation
• Up To Here: summarize everything up to the selection endpoint
• From Here On: summarize only the selected range and after

Why the three scopes: long conversations often split into phases. Summarizing Up To Here lets you capture a discovery phase before starting a planning phase. From Here On lets you summarize a specific decision or debate without dragging in earlier setup context.

What happens when you click

Conflab uses the helper model to generate the summary. It saves to the conversation's summaries folder in RESULTS. A success confirmation appears.

What the summary viewer supports

• opening the rendered summary in the viewer modal
• Export: save an HTML copy to your downloads folder
• Print: open the browser print dialog, which lets you save as PDF
• Delete: remove the saved summary

The print-preview-with-PDF-option screenshot above is not a separate Conflab screen: it is the browser's native print dialog that appears after clicking the print icon inside the summary viewer. Use its Save as PDF destination to produce a shareable PDF of the rendered summary.

Summaries with associated images

You can attach one image from the image dialog to the "current summary" by toggling Use in Summary. When you then open the summary in the viewer, the image is shown inline after the heading.

Because that image is referenced from workspace storage rather than embedded in the HTML, the image is visible inside Conflab but not in an exported standalone HTML file. If you want a shareable deliverable with image included, use Print -> Save as PDF while viewing the summary in Conflab.

Log Workflow

The Log button saves a full transcript. Unlike Summary, this is not an AI-generated artifact: it is a clean conversion of the conversation rounds into a readable document.

How the Log button behaves

Same scope behavior as Summary:

• if no round range is selected, it saves the full log immediately
• if an intermediate range is selected, it shows Full / Up To Here / From Here On

What the log viewer supports

The log viewer uses the same modal as the summary viewer. Same controls:

• Export: HTML file to your downloads folder (filename includes the conversation name)
• Print: browser print dialog for PDF
• Delete: remove from workspace

What the log preserves

Logs keep useful metadata such as:

• log generation time
• rounds included
• conversation mode
• Director type

That makes logs useful both for review and for recordkeeping. If you ever need to prove exactly what a discussion said, or rerun a later analysis against the raw material, the log is your source.

Image Generation

The Image button opens Conflab's image workflow.

The dialog includes:

• From Summary
• From Scratch
• Style
• Aspect Ratio
• Model
• prompt navigation arrows
• preview frame
• Export
• Clone
• Use in Summary
• prompt editor
• AI Helper chat
• Generate

From Summary vs From Scratch

These are the two image-generation paths and they behave differently.

From Summary: Conflab uses the current conversation's summary to seed a set of image-prompt ideas. This is the "artifact-aware" workflow: the image is created to illustrate the deliverable you already produced. The dialog auto-populates prompt options from the summary.

Use From Summary when: you have a saved summary and want illustrative cover art or inline images for that specific deliverable.

From Scratch: manual prompt-based image generation with no summary seeding. You write the prompt yourself.

Use From Scratch when: you want an image that is not tied to a specific summary: exploratory visuals, concept art, or images for conversations you have not summarized.

Style

Visual style direction for the image request (photorealistic, illustrated, etc.).

Aspect Ratio

Image shape. Matters for both the preview and how the image fits a summary document. If you intend to use the image inside a summary, pick an aspect ratio that reads well in long-form layout.

Model

Image model used for generation. A free image model can still produce usable results, which makes this a reasonable place to save money compared to the helper model.

Prompt navigation arrows

Move through saved or generated prompts in the current image session. In From Summary mode there will usually be several seeded prompts to browse.

Preview frame

Shows the current generated image, or indicates that no image exists yet for the prompt.

Export

Exports the current saved image as a JPG to your browser's downloads folder. Filename is derived from the conversation name. Silent download: no picker.

Clone

Creates a new prompt from the current one. Useful when you want variations without starting over.

Use in Summary

Marks the current image for use in the current summary. When the summary is later opened in the viewer, this image will appear inline after the summary heading.

Mechanics to know:

• it attaches to the current summary for the active conversation
• if no summary exists yet, the button warns you that you need to create one first
• the attachment is a reference, not an embed: the image lives in workspace storage, not inside the summary HTML
• this means the image shows up when you open the summary inside Conflab, but a standalone exported summary HTML file will not include the image

For a PDF deliverable with image included, generate the image, Use in Summary, open the summary in the viewer, and Print -> Save as PDF.

Prompt editor

Directly edit the image prompt text. Useful when a seeded prompt is close but not quite right.

AI Helper chat

Ask the helper model to improve or reshape the current image prompt. Useful when you want to describe in conversation what you are going for, and have the model turn that into a polished image prompt.

Generate

Runs the image generation with the current settings.

Common Workflows

Once the individual controls start to feel familiar, the real value comes from combining them. A few patterns worth knowing:

Brainstorm -> focus -> deliverable

1. start in Parallel mode with several distinct personas
2. run a few rounds of broad idea generation
3. switch to Sequential to pressure-test the best ideas with cross-reaction
4. save a summary and a log
5. generate a From Summary image, Use in Summary
6. open the summary, Print -> Save as PDF

Scenario debate

1. create personas that explicitly disagree (different worldviews, different incentives)
2. use Sequential mode so each persona reacts to the previous one
3. intervene as Director to sharpen the disagreement when it gets mushy
4. save the log: the argument itself is the deliverable

Let AI Director run it

1. set up personas and mode
2. write a strong opening Director prompt
3. switch Director to AI
4. raise SPEED (longer countdown) so you have time to read and intervene
5. watch the discussion unfold; jump in with guest input whenever you want to steer

Analyze a document

1. open a conversation with personas suited to the domain (subject expert, skeptic, editor)
2. attach the document via the attachment button
3. write a Director prompt instructing the agents to work from the attachment
4. run rounds
5. save the summary as your analysis artifact

Live course-correction

1. start a session
2. after a round or two, notice which voice is uninteresting or duplicative
3. click that agent's header, swap in a more interesting persona, or mark Inactive and add a different one with + Agent
4. the conversation continues with the new lineup

Terminology That Matters

A few interface labels are easy to mix up:

• the drawer tab is RESULTS, not Downloads
• the image workflow uses From Summary and From Scratch
• the conversation editor uses Description / Objective of Conversation
• the persona editor uses Persona Setup, Persona Description, Worldview, and Life Experiences
• the main toolbar uses Prompt Assist, Summary, Log, and Image

Final Note

The most important thing to understand about Conflab is that it is not just a chat window with multiple bots. The combination of:

• reusable personas with real identity (structured fields plus worldview plus life experiences)
• Director control (manual, automated, or a blend of both)
• selectable conversation modes that change the quality of group dynamics
• saved deliverables (summaries, logs, images) with export and print paths
• summary-linked image generation for polished output

is what makes the app distinctive. If you learn the difference between Human and AI Director, and between Parallel, Sequential, and Free Form, and you know how to get a summary, a log, and an image out of the workspace as a finished deliverable, most of the rest of the interface will make sense.