The Facilitator's Field Guide

What EGI is

What you are about to run

Emergent Generative Inquiry (EGI) is a structured way to run a room so that everyone in it contributes, the themes come from the contributions rather than from your notes, and the group owns the result. The AI does the sorting and drafting work that would otherwise cost you a wall of sticky notes and a week with a transcript. The group does every act of judgment.

That division of labour is not a detail of the implementation. It is the method.

The Prime Directive: the AI proposes, the group disposes

Every interpretive act in a session has exactly one owner. The AI may draft, accelerate and organize. The group ratifies, corrects and decides.

Two rules follow, and both are visible to your participants:

• No unlabeled AI authorship. Anything the AI wrote that the group has not ratified is marked as an AI draft. Pattern names, narratives, unverified research notes, the synthesis: all of it carries that mark until the room changes it.
• No invisible facilitation power. When you regroup ideas, skip a step or hide a contribution, it is recorded and it appears in the session's exportable record. You cannot quietly steer a session and still produce a clean-looking account of it.

If you remember one thing from this guide, remember that everything the software produces is a proposal. Your job is to put it in front of the room and let the room overrule it.

Where the method comes from

EGI is not invented from nothing. It is three established practices, sequenced and given AI assistance.

Technology of Participation (ToP). The participatory core, developed for facilitated group work over decades: the wisdom is in the room, the group owns the interpretation of its own words, and a workshop is expected to end in something resolved rather than something noted. If you have run a focused conversation or a consensus workshop, you already know this shape and you will recognize most of what follows.

The Question Formulation Technique (QFT). The question is the instrument. Better questions produce better collective thinking, and question design is a discipline rather than a warm-up exercise. This is why one whole chapter of this guide is about the question, and why the question is the part of your preparation you should never hand to anyone else.

Grounded theory. From social research: the literature comes after the data, never before it. Read the field first and it tells you what to look for, and then you obligingly find it. EGI keeps the discipline by running its research step after themes have emerged from the room, so existing theory contextualizes what the group found instead of pre-structuring what the group is able to say.

What EGI adds: the iterated question

EGI's own contribution is the iterated question. Each round's collective output becomes the raw material for a better next question. Across several rounds a session produces a genealogy of questions: what was asked, what the room said, why the next question went in the direction it did, and who chose it.

For a multi-round inquiry that genealogy is the primary artifact, worth more than any single round's themes. It is a record of a community's thinking changing shape, and it is the thing a survey structurally cannot produce, because a survey's questions are fixed before the first person answers.

You are not obliged to run multiple rounds. Most first sessions do not, and this guide will tell you plainly where a single round can honestly stop. But the design assumes the question is alive, and the product keeps the reasoning when it changes.

Why a survey would not do this

Surveys are excellent at measuring something you already know how to ask about. They fail in exactly the rooms EGI is built for.

• The categories are yours, not theirs. A multiple-choice question has already decided what counts as an answer. Anyone whose answer is outside the list picks the nearest thing or stops responding.
• Nobody sees anyone else. A survey collects opinions in parallel. It cannot produce what happens when someone reads a stranger's contribution and reconsiders their own.
• The analysis happens in private. Free-text answers go into a coder's spreadsheet and come back as a summary. The people who wrote them never see how their words were read, and have no way to object.
• The response rate tells you who was already engaged. The people a consultation most needs to hear from are the people least likely to complete a form.

EGI answers all four inside the room. The categories are drafted from the contributions themselves. The room watches them form. The room renames them when they are wrong. And nobody has to write an essay to be counted, because two honest sentences from a phone is a full contribution.

What you get in exchange for that is a different kind of obligation. A survey can be defended by its instrument. A session is defended by what the room did with it, which is why the rest of this guide is mostly about your decisions rather than the software's.

The six steps, and where to stop

The six steps of a round

A round moves through six steps in a fixed order: spark, gather, emerge, validate, enrich, evolve. A session is one or more rounds.

Spark

The question goes live and the whole room sees it at once. Two to five minutes, and the moment your preparation either pays off or does not. The question determines what the rest of the session can possibly produce, which is why it has its own chapter.

Gather

Participants write and submit from their own phones, ten to twenty-five minutes in a live room. They can watch the shared stream of what others are writing, or you can keep gathering blind so nobody anchors on the first confident answer. Blind gathering suits contested subjects and rooms with a status gradient.

Emerge

The AI reads every contribution and drafts a set of patterns, each with a name and a short narrative. In a live room this takes a minute or two, and the room watches it happen rather than waiting out a spinner. Participants can then see which pattern their own contribution landed in. Every pattern arrives as a draft, not a finding.

Validate

The room works on the drafts: rating how well each pattern reflects what people meant, commenting, saying what is missing, suggesting better names. This is where the group takes ownership of the map, and it is the difference between an AI's reading and a community's account of itself. Budget ten to twenty minutes and do not compress it.

Enrich

Research. The platform searches for literature and frameworks that speak to the patterns the room produced. In a live room it runs in the background or after everyone has gone home, and the results attach to the report. In a longer inquiry it runs inside the round, because participants have time to read it.

Evolve

The platform drafts four candidate next questions, one per direction: deepen a validated pattern, expand into adjacent territory, bridge two patterns or communities, or pursue a contrast where the room disagreed. Participants vote among them. You decide, and you may write your own instead.

Resolve is a closing segment, not a seventh step

Once the synthesis report exists, participants endorse or reserve on each of its recommendations, and those rates appear in the report itself. This is the moment the session lands: a recommendation that 70 percent of the room endorsed and one that split it are different objects, and the report says which is which. Resolve is not a step in the round. It is the closing segment of the session.

Why the research comes after emergence

Running the literature review first would tell the group what to look for, and a group told what to look for finds it. Running it afterwards means existing theory contextualizes what this room actually said instead of pre-structuring what it was able to say. If you have ever watched a consultation produce exactly the themes in the briefing note, you have seen the alternative.

Two timescales

Live Room is synchronous: thirty to a hundred and twenty minutes, a facilitator and a room of phones, Enrich pushed out of the live loop. More than two rounds in one sitting is discouraged, because energy decays and a tired room is not better data.

Inquiry Campaign is asynchronous, over days or weeks, and multi-round is the point. All six steps run inside each round, including Enrich, and every participant in round two sees round one's validated patterns and the reasoning behind the new question.

Six steps are a ceiling, not a requirement

This is the part facilitators most often get wrong, usually by assuming a session that stopped early was a session that failed.

Depth is your choice. Stopping early is legitimate session design, it is logged like any other skip, and the report states how far the room went. What changes with the stopping point is not the quality of the session but the claim it entitles you to make afterwards.

A partial loop carries a correspondingly thinner provenance record, and no partial loop is exportable as research. Decide your stopping point while you are planning the session, not while the room is watching you.

What ratification means, mechanically

Renaming happens in Validate, and it is the checkable event that separates the third row of that table from the second.

Until the group renames a pattern, the product keeps showing the AI's name for it and marking it as an AI draft, in the room and in every report. When the group does rename it, the export records "Renamed by group" against that pattern. That column is what an auditor, a board or a journalist can inspect. It does not depend on your recollection of how participatory the session felt.

A rehearsal session makes the point. It ran spark, gather, emerge and validate, generated its report and closed without running enrich or evolve, and produced a complete and useful report that way. (The session was reopened later to exercise the two remaining steps; the report described here is the one produced at close.) It also voted seventy-two times and renamed nothing. Every heading in that report stayed an AI draft, correctly labelled as one, because nobody in the room had ever put a pattern in their own words.

The move to make before you close

Ask the room to rename at least one pattern.

Not all of them, and not as an exercise. Pick the pattern that reads most like the software talking, read its name aloud, and ask what the room would call it. Someone will say something better within thirty seconds, because they wrote the contributions and the AI did not. Then take the new name.

It costs two minutes, it visibly transfers ownership of the map to the people who made it, and it is the difference between a report about a room and a report by one.

What the AI does, and what it does not

One owner per interpretive act

Every act of interpretation in a session belongs to somebody. The design rule is that the AI proposes and the group disposes, and the table below is the specific form that takes. It is worth knowing before you run a room, because participants ask, and a facilitator who can answer this precisely is trusted for the rest of the session.

Read down the middle column and you have a description of a fast, tireless assistant. Read down the right column and you have a description of a group doing exactly the work a good workshop asks of it. Neither column is decoration.

What "AI draft" means on the screen

When the AI names a pattern, the product keeps that name as the AI's draft and shows it marked as such. If the group renames the pattern, the group's name becomes the one used everywhere, and the original draft is retained rather than overwritten.

This has two consequences you should say out loud in the room.

First, an unrenamed pattern is visibly the software's reading, not the room's. Nobody has to take your word for how participatory the session was, because the labels themselves report it.

Second, the record is honest in both directions. The export shows which patterns the group renamed and which it left alone. A session where the room changed six names out of six looks different from one where it changed none, and both look like what they are.

Your powers are real, and they are public

You can regroup a contribution that landed in the wrong pattern. You can skip a step. You can hide a contribution that names a private individual or that no room should have to read.

You should use these. The regrouping power in particular is the most credibility-building thing available to you in a contested room: fixing the software's mistake in front of everyone is proof that the room is not being steered by it.

The constraint is that none of it is silent. Regroupings, skipped steps and hidden contributions are all recorded and appear in the session's exportable record. That is deliberate. A facilitation power that leaves no trace is a power that can be abused invisibly, and this method does not have any of those.

Practically: when you hide something, say that you have hidden an entry and why. When you skip a step, say which one and why. The record will say it anyway, and it is much better coming from you.

What the AI does not do

It does not vote. Every number in a session comes from a person tapping their own phone. There is no AI weighting, no AI-generated participant, no synthetic response filling a gap.

It does not decide. It has no ability to select the next question, approve a pattern name, or determine what the report recommends without the group and you passing through first.

It does not work on individuals. The covenant commits that the group may be studied and the person may not: the AI is put to work on the collective body of contributions, and the design rule is that it is not asked to build a picture of a named participant or to judge who said what. Aggregate displays are built to enforce minimum group sizes rather than showing a cell of one.

It does not know what your room means. This is the important one and it is not a limitation you can configure away. The grouping works on topical similarity: contributions about the same subject sit together whether they agree or violently disagree. The software cannot hear a stance. That property is exactly why group ratification is methodologically required rather than a nicety, and it is the mechanism behind the failure described in the chapter on questions.

What this asks of you

The method spends AI capability on speed and spends your attention on judgment. The steps where you must be most present are the ones where the software has just produced something plausible: the moment the patterns appear, and the moment the report drafts.

Plausible is the risk. A confident, warm, well-written pattern name that quietly merges two opposed positions will pass unchallenged in a room that has been told the machine is clever. Your job at that moment is to slow down, read it to the room, and ask whether it is true. If the answer is no, change it in front of them.

The privacy covenant

Why this is method, not policy

Anonymity is a methodological requirement here, not a preference. People do not say the true thing when the room can trace it back to them, and a session that cannot hold a difficult contribution is a session that produces a comfortable and useless result.

So the privacy rules carry the same force as the Prime Directive. What follows separates three things that are usually blurred: what is true of the product today, what the covenant commits to, and what you personally have to say out loud.

True today

All AI work happens on our servers. The browser your participants use never holds an AI key. Nothing about a session's AI calls is exposed to the device in the room.

Typed contact details are stripped before words travel. Before contribution text is sent to any AI or search vendor, a scrub removes email addresses, phone numbers, social handles and links carrying access tokens. This is a deterministic pattern scrub, and it is worth knowing its edge precisely: it catches those patterns. A first name typed in the middle of a sentence is a harder problem and is not yet caught by it. If that distinction matters to your room, say so rather than rounding it up.

Anonymous means unlinkable to a real identity. An anonymous participant is given a randomly generated name minted by the server. It is not derived from their device, it does not follow them between sessions, and knowing someone's random name grants no ability to write anything as them. Inside a single session, contributions carrying the same random name can be seen as coming from one person, which is what makes voting work at all. That trail is not linked to who they are.

Participants can delete their own contributions without an account. The delete control removes their ideas, votes, comments and likes from the session, working from the browser they used, with no sign-up and no request to you. One exception is worth knowing: if someone asked for the report by email, that email request is not self-service deletable, because there is nothing verifiable linking it back to them.

Where the data physically sits. Contributions are sent to AI providers in the United States, and sessions are stored on US-based infrastructure. Participants are told this. If your organization has a data residency requirement, resolve it before you book a room, not afterwards.

Three caveats you must state out loud

These are true, they are not hidden anywhere, and they are the sentences a hostile questioner will find if you do not say them first.

A raffle links its own entries. If you run a prize draw, the name and email a participant chooses to give go to you so you can reach a winner. That is stored apart from their ideas and is not sent to the AI, but it exists, and it is the one flow where a participant deliberately identifies themselves. In a contested public room, the honest choice is usually to turn the raffle off.

Voice recordings and drawings travel unscrubbed. Audio goes to a transcription vendor as audio, and a drawing goes as an image. Neither can be pattern-scrubbed the way typed text is. If you turn either input on, your room must be told, at the start, that anything they say aloud or draw travels as-is. A session that does not use those inputs does not carry the caveat, and the participant disclosure only mentions them when they are actually switched on.

The hosting layer sees visitor addresses. Serving a web page requires it. It is not joined to contributions, but it is not nothing, and denying it later is worse than mentioning it now.

What the covenant commits to

The rest of the covenant is stated as commitment, in the honest sense that the product is built toward it and can be held to it, not as a description of a finished feature.

• The group may be studied; the person may not. The design rule is that no individual is profiled by the AI without specific, informed and revocable consent, and that aggregate displays enforce minimum group sizes.
• De-anonymization is a facilitation power, never a silent one. The covenant commits that any decoding of scrubbed identifiers is permission-gated and recorded in the session's event log, on the same principle that forbids invisible facilitation power anywhere else.
• Right to leave, and retention limits. Sessions honour their access-after-finish setting, and operational records such as access logs and extracted documents carry retention limits rather than living forever.
• Transparency over trust-me. The covenant commits to publishing the data-flow description and the list of subprocessors, and to every session report stating the models used, the research depth and the cost.

If a participant or a client asks whether one of these is built or committed, tell them which. The distinction costs you nothing and buys you the right to be believed about the first section of this chapter.

What participants already see

You are not the only disclosure. Before anyone contributes, the session itself shows a statement of these terms, written for the identity the participant actually has: different text for someone who joined anonymously, by invitation, or with an account. It states where their words go, that the scrub covers contact details, that nobody is profiled, and how to delete what they contributed. The raffle line appears only if the raffle is on. The voice and drawing lines appear only if those inputs are on.

This changes what you have to do at the top of a session. You do not have to recite a privacy policy. You do have to name the two or three things that are specific to your room: who convened it, what happens to the output, and whichever of the caveats above applies to the configuration you chose.

Asking a question that works

The test that produced this chapter

Two sessions were run in the same product, on the same day, on the same fictional town. Same eight groups of participant personas. Same session preset. Same clustering. Thirty-two contributions each, from thirty-two distinct anonymous participants each.

One thing differed: the shape of the question.

Session A asked what stories the town's forty-seven-year-old arena had held for people. A memory question, warm, obviously well-intentioned, about one object.

Session B asked what matters most about the town's next ten years, and what people would be willing to give up to protect it. Several concerns, and a tradeoff.

Session A produced one pattern. Session B produced four.

Session A's single pattern was called "The Heartbeat of Millbrook", and its themes were listed as shared memories, emotional connections, community identity, financial concerns and practical needs. Into that one warm bucket went the hockey father, the farmer who wanted the money spent on washed-out grid roads instead, the parent whose son cannot reach rink level in a wheelchair, and the curler who said thank it for forty-seven years and let it go.

Four incompatible positions. One theme. Nothing to vote on, because there was only one thing on the ballot, and the report told a town in genuine disagreement that it resonated with a deep shared connection.

Nothing warned the facilitator

Both sessions displayed the same coverage line: "Every idea is in a pattern."

That statement was true in both. It is maximised by exactly the collapse it ought to be catching, because putting everything in one bucket loses nobody. Session A scored full marks on the only quality signal on the screen while producing a useless result.

So the software will not tell you. This is the single most important reason the question is your responsibility and not something to be delegated to a preparation step or an assistant.

Why it happens

The clustering works on topic, not on stance. Contributions about the same subject sit near each other whether they agree or completely disagree, because agreement and disagreement about one thing are, mathematically, the same thing being talked about.

Ask about one object and every contribution is about that object. One topic, one cluster, no matter how divided the room is. The heat is all there in the words and none of it is visible in the grouping.

Ask about several concerns and people sort themselves by which concern they raise. Now the patterns are real distinctions, and voting on them means something.

The rule

Ask a question that invites different people to talk about different things. A question about one object produces one theme and nothing to vote on.

What it looks like in practice

Four pairs. The failing version in each is not a bad question in the ordinary sense. Each one is warm, clear, and would read perfectly well on a comment card.

A workplace session after an engagement survey

• Fails: "What do you think about our new hybrid work policy?" One object. Everyone who loves it and everyone who hates it lands in the same pattern called something like "Perspectives on Hybrid Work".
• Works: "What makes a genuinely good week here, and what would you trade away to get more of them?"

A town hall about a contested facility

• Fails: "What has the community centre meant to you over the years?" One object, and memory-shaped, which is the same trap wearing a friendlier face.
• Works: "What matters most to you about the next ten years here, and what would you be willing to give up to protect it?"

A board retreat on strategy

• Fails: "What do you think of the draft strategic plan?" One object, and it also invites the room to react to your document instead of thinking.
• Works: "Where will this organization be strongest and weakest three years out, and if you could only fix one of those, which?"

A conference breakout

• Fails: "What did you take away from this morning's session?" One object, and it produces a single pattern of polite appreciation.
• Works: "What is actually hard in your practice right now, and what have you quietly stopped trying?"

Notice what the working versions share. Each names more than one thing to talk about, and each asks for a cost, a choice or a tension. A tradeoff clause is the cheapest reliable fix available: asking what someone would give up forces them to name a second thing, and second things are what patterns are made of.

Three checks before you commit

Name four answers that would be genuinely different from each other. Not four opinions on the same subject: four different subjects a reasonable participant might raise. If you cannot get to four, the question is too narrow and you already know what the session will produce.

Ask whether two people who disagree would talk about different things or the same thing. If they would raise the same subject with opposite feelings, you are looking at session A. Add a second concern or a tradeoff.

Read it as someone on the losing side of the decision. Can they answer honestly without first conceding your premise? A question containing the proposal produces a fight about the proposal, and half the room will refuse it, correctly.

One honest warning about the assistant

The product's question assistant will help you develop a question, and it works. It also produced session A's question on its own, unprompted, and gave no indication that anything was wrong with it.

There is no neutrality check and no warning connecting the question you choose to the clustering that will fail because of it. The assistant is a drafting partner, not a reviewer. Run the three checks above yourself, every time, on whatever it hands you.

Reading patterns critically

The AI drafts, the group decides

When clustering finishes, the room sees a set of patterns. Each one carries a name, a short narrative, and the contributions it holds. Every one of those names was written by the model, and until somebody changes it the product says so: a pattern whose name still matches the AI's draft is visibly tagged as an AI draft. Rename it and the tag goes away, the model's original wording is kept beside it in the record, and every report from that moment uses the group's words instead.

That badge is the whole method in one detail. The AI reads fast and shallowly. The room reads slowly and knows what it meant.

The coverage line is not a quality score

Above the patterns is a coverage statement. When every contribution has been assigned somewhere, it reads Every idea is in a pattern.

Read that for what it is: a statement that nothing was dropped. It is not a statement that the grouping is any good, and it cannot be, because the worst possible result scores full marks on it. One giant pattern containing all thirty-two contributions leaves nothing outside any pattern, so the line reports perfect coverage on a result that carries no information at all.

This is not hypothetical. Two rehearsal sessions were run on the same fictional town, with the same participant personas, the same preset, and the same clustering. One question produced four patterns. The other produced one pattern holding all thirty-two contributions, merging a father talking about hockey, a farmer who wanted the money spent on washed-out roads, a parent whose son could not reach rink level in a wheelchair, and a resident who wanted the building closed after forty-seven years, into a single warm narrative about community connection. Both sessions reported Every idea is in a pattern.

The metric is maximised by the result it should be catching. Nothing on the screen will flag it for you.

What to look at instead

Coverage answers "did we lose anyone?". You still have to answer "did we learn anything?". Look at the distribution:

• Is one pattern holding most of the room? If the largest pattern has much more than half the contributions, the disagreement worth having is inside that pattern, not between patterns.
• Are two patterns near-duplicates? Two names a stranger could not tell apart usually mean the model split on wording rather than on substance.
• Did a large set of contributions produce a single theme? Above roughly twenty contributions, one pattern is a failure, not a consensus.

Every one of those points back at the question rather than at the software. A question about one object gives every contribution the same subject matter, so the grouping has nothing to separate them with no matter how sharply people disagree. If the patterns collapsed, the fix belongs in the next question, and the chapter on asking a question that works is where to take it.

Even a good result is lopsided. In the rehearsal's good case, 23 of 32 contributions (72 percent) landed in one of the four patterns, and two of the four held two contributions each while being drawn at exactly the same size as the one holding twenty-three. Visual weight on this screen does not track how many people said it. Say the real numbers aloud when you present the patterns.

Your three moves

Re-run the clustering. This is legitimate and it is logged. It is also destructive, and the product warns you before it happens: re-running rebuilds the pattern set from scratch, so votes and comments the group already left on the patterns being replaced are deleted with them, membership you corrected by hand is recomputed, and which contributions fall outside every pattern can change. Re-run before the group votes, never after.

Move contributions between patterns. When you can see that something has landed in the wrong place, move it. This is the correction drill the session guides talk about, and doing it in front of the room is the most credibility-building thing available to you: it shows, rather than asserts, that the group outranks the software. In a contested room it is the single move that answers the accusation that the process is steering people.

Get the room to rename. This matters most and gets skipped most. Ask out loud: "Does that name describe what you wrote? What would you call it?" Take the room's wording even when it is clumsier than the model's. Clumsy and theirs beats elegant and imported.

Before you close

A report full of AI draft headings is the AI's reading of your room, and anyone who looks closely will read it that way. The export makes it checkable rather than a matter of opinion: there is a column named "Renamed by group", and it says no for every pattern nobody touched.

Rename at least one pattern before you close. It takes thirty seconds, and it is the difference between "here is how software grouped what we wrote" and "here is what we found".

Facilitating the room

Read what your participants are told

Before anyone contributes, the product shows each participant a short statement of how the session handles what they share. It tells them whether they are anonymous and what that means inside the session, that their words travel to AI providers in the United States and that email addresses, phone numbers, handles and links are stripped out of the text first, that the AI works on the group's ideas rather than building a picture of any individual, and that they can remove their own contributions from the browser they used without needing an account. When a raffle, voice input or drawings are switched on for that session, it says what happens to those as well.

Read it yourself before the session, for two reasons. You must never say anything in the room that contradicts it. And most of the questions you get in the first five minutes are already answered there.

What moderation actually does

You can hide a contribution from the room, and you can put it back. Hiding removes the text from the participant view and drops it out of the visible count.

Three things follow, and all three are worth knowing before you need them:

• It is reversible. Hiding is not deletion.
• It is not silent. The action is written to the session record and appears in the exportable provenance, so anyone reviewing the session afterwards can see that something was hidden and when.
• It does not un-read what was already read. If clustering has already run, the pattern was drafted from a set that included the hidden contribution. The product tells you this at the moment you hide it.

There is no way to remove a contribution without leaving a trace, and that is deliberate. A facilitator who can quietly delete what they dislike is a facilitator whose session record proves nothing.

In practice: hide abuse and anything that names a private individual. Then say out loud that you have hidden an entry and why. In a contested room, silently removing something is how you lose the room permanently.

Why gathering and voting run blind

Every session preset gathers blind: participants write without seeing what anyone else has written yet. The reason is anchoring. The first three contributions on a shared screen set the vocabulary for everyone who writes after them, and the room converges on an early framing instead of producing what it actually knows.

Every preset pins this setting, so under a preset it is settled before anyone joins. In a session created without a preset, you can open the shared stream live when the room needs the energy. A good moment for it is a gather that has gone quiet early. A bad moment for it is the first two minutes, when the silence is people thinking rather than people stuck.

Voting runs blind then reveals. A participant sees no counts on an item until they have voted on it themselves, and the order of patterns is shuffled per participant so position does not decide what gets read first. Both are on by default under every preset, for the same reason as blind gathering: a running tally is an instruction, not information.

Words to say aloud

Adapt the wording, keep the content.

Opening.

We are going to do this in a particular order. You will each write on your own, without seeing anyone else's answers, so nobody's phrasing sets the frame for everyone else. Then software will group what we wrote into themes, and then we will correct it. The software drafts. This room decides what the themes are called and whether they are right.

Before gather.

Write in your own words. One idea per entry, so it can land in the right place. Specific beats general every time. Nobody in this room sees your name on anything you write.

Before validate.

Vote your own view first. You will not see how anyone else voted on an item until you have voted on it, and the order is different on every phone, so nothing here is telling you what to think. Once you have voted, the results open up.

If voice input or drawings are switched on.

If you record a voice note or draw something, that travels as audio or as an image to an AI provider, and the text protections do not apply to it. Type instead if that matters to you.

Closing.

Before we finish, look at these names. Does each one describe what you wrote? Tell me what you would call it instead, and I will change it, because the names in the report should be yours and not the software's.

Two practical notes

Decide who is in the room before the date is booked. In an organizational session, whether leadership attends changes what gets written more than any setting does. Anonymity does not help when people can see who is standing at the back.

Bring a second person if you can. One of you watches the screen, the other watches the room. The software does not support shared facilitation accounts, so this is an arrangement between two humans at the venue, and in a difficult room it is worth more than any feature.

Choosing a guide and a preset

The eight session guides

A session guide is a document you read before you run a room. It names the job, gives you question seeds and the question that fails, a run of show with timings, and the way that particular session goes wrong. They live in the guide library (ideajar.ai/templates).

Match the ending to the promise you are making

Every guide declares where it ends, and that is a design decision rather than an accident of timing. Six steps are a ceiling, not a requirement.

A conference breakout that finishes at Validate and a multi-round participatory study are different promises to the room. The breakout promises delegates their own themes on the screen and a record afterwards. The study promises that their framing shapes the next round's question. Announce the one you are actually making, at the start, and stop where you said you would stop.

Stopping early is legitimate session design. It is logged like any other skip, and the report states how far the room went. Stopping early while implying you went further is the thing to avoid.

Presets apply a guide's settings for you

A session preset is the configuration side of the same pairing. Each guide ships a matching preset, so you apply the guide's settings when you create the session instead of transcribing seven switches out of a document at the wrong end of a busy week. The town hall guide's preset is called Civic neutral; the rest carry their guide's name.

All eight presets share the same floor, because all eight guides asked for it independently:

• Anonymous contribution on
• Raffle off
• Blind gathering
• Blind then reveal voting
• Shuffled pattern order
• Results open to participants after the session closes
• Voice input and drawings off, because both leave the text protections behind

What differs between presets is what they pin. Five of them (after the engagement survey, town hall, needs assessment, participatory research, position development) pin the whole identity bundle, because in those rooms a participant can be harmed by being identified and the promise has to hold for the entire session. Three of them (strategy offsite, conference breakout, board retreat) pin only the anti-anchoring settings, because the named failure in those rooms is the senior voice arriving first rather than participant harm, and whether the room is named or anonymous stays your judgment call.

What "pinned" honestly means

Say this accurately if a client asks. The difference is real, and it is not in your favour to blur it.

• The raffle lock is enforced by the server. A session created with the raffle off cannot collect a raffle entry, regardless of what any browser tries to do.
• Every other lock is enforced in the interface. The controls render disabled and the session is created with the preset's values, which is genuinely what protects you from a mistake on a busy day. It is not a technical guarantee that the configuration held.

Treat a preset as a strong default and a shared agreement about how this room will be run. Do not describe it to a convener as something they could prove afterwards. If one needs to prove the configuration held, tell them that is not available today.

Sizing the room

The participant cap counts every device that joins the session. Someone who arrives late, votes on the themes, and writes nothing consumes a seat exactly as a contributor does.

This has caught a real session out: 32 contributors plus 19 voters is 51 seats, and a cap set for "50 people" stopped accepting participants partway through voting. Size the cap for everyone who will hold a phone, not for everyone you expect to write something, and leave headroom.

Cultural responsiveness

What the question assistant draws on

When you cultivate a question, the assistant works from a base of more than thirty questioning and participatory traditions across ten disciplines. Alongside the familiar organizational ones (appreciative inquiry, the Question Formulation Technique, solution-focused questioning, systems thinking) it holds traditions that come from specific communities and carry their own assumptions about what a question is for:

• Sikolohiyang Pilipino, and its central idea of kapwa, the shared self
• Ubuntu, and the relational framing that a person exists through other people
• Indigenous Ways of Knowing, including Seven Generations thinking and circle processes
• Community-Based Participatory Research, where the community is a partner in the inquiry rather than its subject
• Critical pedagogy and decolonizing methodologies, which treat who gets to ask the question as part of the question

The assistant reads the title and description you give the session to work out which of these might apply, and it tells you which ones it drew on and why. It can ground a question in the context you described. It does not know your community.

Give it something to work with

The single thing that most improves what comes back is the session description. A description that says only "community consultation" produces a generic question. A description that names who is in the room, what has already happened to them, what the convener has and has not decided, and what the room is likely to be sceptical of, produces something you can actually work from.

It is the cheapest quality lever you have, and it costs two minutes at your desk.

The part that stays yours

Every suggestion is a draft. Cultural fit is a facilitator's judgment, and it is not one the assistant is in a position to make.

Three things worth holding onto:

Communities are not monoliths. A question grounded in a tradition is not automatically welcome to everyone who might be described by it. Generational, class, religious and diaspora differences run through every community the assistant can name, and a framing that lands with one part of a room can read as presumptuous to another.

A named concept used loosely is worse than no concept at all. If a suggested question leans on a term from a tradition you do not know well, either learn enough to use it precisely or take the plain-language version. People notice borrowed vocabulary immediately, and it costs you the room's trust at the exact moment you need it.

Ask someone from the community before the session, not after. If you are working with a population you are not part of, run the question past somebody who is. That conversation catches things no methodology base can, and it also builds the relationship the session depends on. The guides that deal with community rooms treat this as recruitment work, not as a courtesy.

What this does not do

It does not certify that a question is culturally appropriate, and nothing in the product could. It suggests, names its reasoning, and hands the decision back to you. Same division of labour as everywhere else in the method: the software proposes, and the people responsible to the room decide.

After the session

What you leave with

In order of what matters:

1. The ratified pattern map. The themes the room named, with the AI's draft name preserved beside each one, the contributions in each, and the contributions that fitted no theme kept visible as open items rather than discarded.
2. The validation record. How the room weighed each theme, the comments about what is missing, and any reservations recorded rather than smoothed into agreement.
3. The synthesis report. An AI-drafted narrative of the session that quotes real participants verbatim, along with the models used and the session's cost.
4. The provenance record. Every skip, every re-run of clustering, every hidden contribution, every rename, logged and exportable.

That fourth item is what makes the first three defensible. A room that took the session somewhere unexpected, or a facilitator who intervened, leaves a trail either way, and the trail is available to anybody who wants to check the process rather than take your word for it.

A session that stopped early produces a correspondingly thinner record. That is a legitimate choice and the report says how far the room went. It also means no partial loop is exportable as research, and you should not describe one that way.

Across several rounds, the method is designed so that each round's question is chosen from the previous round's findings and the chain of questions is itself part of the record. That is the participatory research guide's territory and it needs the full loop, every round.

Who can reach the report

The report can be read on screen. On plans that include those features, it can also be printed to a PDF through the browser and emailed to participants by the facilitator. A participant who never made an account can ask for it by email themselves regardless of plan, which matters more than it sounds: in most community rooms, the people you most want to keep informed are the ones who will not sign up for anything.

Whether participants can reach the results after the session closes is a session setting. Every preset leaves it open. Say at the close what you have set it to and when the report will exist.

What the export contains

The data export gives you the contributions with the pattern and question each belongs to, the patterns with the AI's draft name beside the group's name and a column literally named "Renamed by group", per-pattern vote and comment totals, and the comments themselves.

It carries no identity of any kind, and no way to link one contribution to another by the same person. Votes and comments are per-pattern totals rather than per voter. Contributions hidden during the session are absent from the file and are not counted in the totals, which is worth knowing when an export number differs from one you wrote down in the room.

What it costs

The AI steps are metered per session, and the report states the cost. That now includes the literature enrichment step (verified in production on 2026-08-03). One sliver remains outside the totals, stated plainly: when Enrich runs in a live-search mode (Extensive or Research), the small AI call that writes the search queries is counted in the round's enrichment cost but not yet in the session analytics total. It is a few cents at most. Standard and Basic-mode sessions are counted in full.

If a client is asking you what a session costs to run, the metered figure is the right one to quote; mention the search-query sliver only if the session used live search and the invoice needs to be exact.

Participants can still remove what they wrote

A participant can delete their contributions without an account, before or after the session closes. It removes their ideas, votes, comments and likes.

One honest caveat you should give them, especially in a research context: the credential that proves those contributions were theirs lives on the device they used. If they clear that browser's storage or switch to another device, that route is gone. Tell them at the close, while they still have the phone in their hand, rather than in a consent form they will read once.

The commitments behind the room

Two of them outlive the session and are worth being able to state accurately.

The group may be studied; the person may not. The AI is put to work on the group's ideas, and the covenant commits that it is never used to build a picture of any individual participant, and that aggregate displays of demographic data enforce minimum group sizes so nobody can be singled out of them. That is the standing commitment the product is built to, and it is what the participant disclosure tells your room.

Facilitation power is visible. Anything you do that shapes what the room sees, hiding a contribution, re-running the clustering, skipping a step, appears in the record. It is a constraint on you, and it is also the reason a sceptical reader has any grounds to believe the result.

Pre-session checklist

Print this page and take it with you.

The question

• ☐ Does it span several concerns rather than one object? A question about one thing produces one theme and nothing to vote on.
• ☐ Would two people who disagree both find something to write, without either conceding the other's premise?
• ☐ Does it invite different people to talk about different things?
• ☐ Is it answerable from experience within two minutes of hearing it?
• ☐ Does it avoid containing the answer the convener already has?

The commitment

• ☐ Which guide are you running?
• ☐ Where does this session stop: Emerge, Validate, Evolve, or the full loop?
• ☐ How long is the block, and does the run of show fit it?
• ☐ Can you say in one sentence what happens to the output and who responds to it?

The settings

• ☐ Preset applied at creation, so the settings match the guide.
• ☐ Raffle decision made deliberately. Off wherever people describe hardship, criticize their employer, or take a public position.
• ☐ Voice input and drawings decision made. If either is on, you say the vendor sentence aloud: those travel as audio or images and the text protections do not apply.

The room

• ☐ Participant cap sized for every device that will join. A person who only votes takes a seat exactly like a person who contributes. Add headroom.
• ☐ Share code on a slide, not only as a code to scan. Scanning fails in low light.
• ☐ Network tested from the back row, the day before.
• ☐ You have read the participant disclosure and will not contradict it.
• ☐ Decided who is in the room, including whether leadership attends.

The close

• ☐ Ask the room to rename at least one pattern, in their own words, before you finish.
• ☐ Say what happens to the report, who reads it, and when.
• ☐ Remind participants they can delete their own contributions, from the browser they used, with no account.