Social AI Principles

15 principles for AI-supported work with groups.

Practice

Core idea

This bundle is not about AI. It is about people who want something together, and how AI can support that work without taking it over.

Status

Living document

public 2026-05-20

updated 2026-05-22

CC-BY-SA 4.0

raw markdown →

Mentioned in this document

context, not a requirement

The canonical bundle for AI-supported work with groups. 15 principles in 7 clusters, plus the mechanism beneath the principles, the underlying disciplines, and three prompt-design translations. Style: keep the detail, drop the fluff. Voiced true to the source where possible.

It starts with the person

Before we list the principles: this bundle isn't about AI. It's about people who want something together (a neighborhood, a team, a group of care workers, a district) and how AI can support that work without taking it over.

"Which question do you ask the group? Do people feel safe enough to be honest? Are you asking for experience or for opinion? If you bring in AI: what do you give it? The words of the people themselves, not your summary."

All of those questions start with the person. Not with the tool, not with the system, not with the prompt. These principles try to keep that order intact.

Bridge to the foundation: this bundle rests on a layer of bottom-up practice theory that doesn't live here. For anyone who wants to dig deeper (organic process, leadership shift, fertile ground as a precondition): see Bottom-up. This synthesis is work-oriented; the foundations sit one layer down.

A discipline in development

We're at the start of this AI journey of discovery. What's written here is direction for what we currently see working, not exhaustive, not absolute. Rules are working hypotheses that refine as we learn more. Marking AI inference ("this is an assumption the AI is making itself", "possibly underexposed", "a pattern that isn't explicitly named in the transcript") is valuable where it can be done, as a concrete discipline to protect ownership. Sometimes feasible, sometimes not. Leave room for new forms of adding and new ways to safeguard ownership.

Cross-link with principle 7 (the map is not the territory): these principles are themselves a lens, not a truth. Refinement happens through practice and reflection, not through dogma.

Fundamental

The axis everything turns on. Without these 4, the rest collapses.

1. Wisdom lives in the system. AI helps make it visible.

AI makes visible what is already there. What AI adds, it does with care and marked as such.

Layer 1, wisdom comes out of the system. Participants, their relationships, their invisible networks (the mosque, neighborhood help, barbecues, conversations in shops) already contain the knowledge. AI doesn't invent knowledge that wasn't there. No external expert framework from outside as the starting point. What experts wrote about it is not the main source; the transcripts and conversations themselves are. AI output structures around what participants already know and said. Analyses go back to the source material, not to what others wrote elsewhere about the topic. Things that weren't said are marked as hypothesis (the "Broaden" role), not as fact.

Layer 2, AI's additions are deliberately chosen and marked. What AI sees or makes is not a replacement for what participants said. Structure, connections, hypotheses, scale, readability: that is what AI legitimately adds. Every addition is explicitly labeled, never mixed in with what was said. The level of addition depends on the prompt's goal: a mirror may add almost nothing, synthesis may connect patterns, serendipity may raise questions. Marking "this is an AI inference" or "possibly underexposed" protects ownership.

Where it sits under tension: when publishing toward an external (policy) context, there's pressure to get expert language in. Resist it. If external language is needed: add, don't replace. The other side: scalability and broadening are real AI contributions that a person can't make, so don't reformulate them too conservatively into "AI does nothing". The art lies in which addition fits where, not in keeping additions out.

2. Trust as a precondition, not an outcome

Participation grows out of existing relationships. Trust is the distribution channel and the boundary within which AI may work.

Layer 1, distribution. People don't take part because a process is well designed. They take part because someone they trust invites them. Cold, broadcast invitations don't work in bottom-up work, even if the content is good. This isn't an observation about behavior, it's a design starting point. AI output that documents methodical work for a field where trust is the load-bearing layer has to acknowledge that reality, otherwise it documents something impossible. Trust carriers (informal leaders, trusted neighbors, existing facilitators) are made visible. The same goes for AI output: don't just design the prompt, design how the output reaches the group too: who shares it, with what words around it, at which moment in the process.

Layer 2, care. AI has to handle what was shared with care. What's said in a session is shared under an expectation of safety: I'll handle well what you share with me, so that you can still look me in the eye afterward. AI has to translate that relationship of trust into concrete working rules:

Patterns at group level first; individual statements only where the context carries them
Pulling an individual statement out of context and putting it in a report damages trust, even if the statement is quoted correctly
Touching someone personally misses the intention and lands wrong
How AI handles human material matters as much as what it generates from it

You can't manufacture trust, only use it and not damage it.

Mechanism link: trust isn't an abstract concept, it's rooted in the body-safety axis underneath all the principles (see "The mechanism underneath the principles", "Body and regulation belong here"). Someone who doesn't feel safe can't truly take part, however well-meant the invitation. Pace, pauses, an eye for freeze/flight aren't soft preconditions, they are the precondition for trust itself.

Where it sits under tension:

External stakeholders (funders, policy partners) often want "number of people" as a KPI. Trust doesn't add up. Be explicit about what you can and can't promise.
Efficiency pressure asks for punchy individual quotes in reports. The risk: one statement carries a whole paragraph, but the speaker would now have to defend, in a room, why they said that in that context. The discipline: pattern level first, individual quote only where the context comes along and the speaker can carry it.

3. Augmentation over automation

AI strengthens human judgment. AI never replaces human judgment.

AI may order, connect, make visible, not conclude on people's behalf. Conclusions in synthesis pages are always traceable to who drew them. If AI sees a pattern participants haven't named: mark it as hypothesis (the broadening role), not as fact. For "what do the participants want?" questions: quote them. Never "the group wants X" without a direct source. AI analyzes, the person decides. AI recognizes patterns, the person interprets. AI offers options, the person chooses direction. AI structures, the person keeps the relationships.

Sub-nuance, AI makes participation scalable: alongside what AI may not do (take over), there's also what AI does make possible: making the signal of every voice visible in groups of 50, 100, 200 people. A person can't do that, not at that pace, not in that completeness. Scalability is a value in its own right, as long as the other principles stay intact (no loss of recognition, no privacy violation, no consultant voice).

Where it sits under tension: under high time pressure, AI (and the people using AI) tends toward "tell me what to do". Resist it. Answers that don't come out of the system are not answers for this system.

4. Ownership through language

Their words are their ownership. Paraphrasing destroys what's there.

When someone says "you're talking to a wall", that sentence holds their experience, their energy, their perspective. If the AI output turns that into "there are communication problems", the ownership is gone. The observation now belongs to the analyst, not to the speaker anymore. This isn't a style preference, it's a question of ownership. Verbatim quotes where possible, with a passage reference. Non-verbal expressions (gestures with balls of wool, Lego constructions, body positions, physical food) are "language" too, document them. Paraphrasing is allowed, but mark it.

"Ownership comes from within. We don't grant it, we recognize it." (see Ownership)

Sub-nuance, your words, your plan: ownership through language applies not only to individual statements but also to decisions and plans. "Can't the AI just make the implementation plan?": it absolutely can, but then it belongs to the AI, not to the group.

An answer in a care-institution session: "It absolutely can, but you are the soul of all this. The fact that you talk about it is what makes it likely you'll support it."

Saying it IS the intervention. Articulation creates ownership, not reporting about it. Dialogue builds shared understanding. Co-creation generates authentic ownership. Commitment follows from active participation. AI may structure and offer options, never decide.

Sub-nuance, verbatim is verbatim-in-meaning, not stenography. Disfluencies (uh, um, er; unintended stumble-repeats like "we we have to"; false starts the speaker corrected themselves right away) aren't part of what someone said. Cut them. Word choice, sentence structure, intentional repetition for emphasis ("really, really important") and characteristic speech or dialect ("you get what I mean, right") you leave alone. Test: would the speaker recognize themselves? No markup in the output (no [uh] or [...]); that draws attention to the noise instead of to what was said. When in doubt: better with the disfluency than paraphrased. The verbatim default (see core boundary 3 in Prompt Best Practices) is no excuse to let transcript noise through; preserving their meaning is the rule underneath the verbatim rule.

Where it sits under tension: flowing prose reads more pleasantly than a patchwork of quotes. Beauty can wreck ownership. Choose ownership. Variation: if a participant paraphrases themselves ("what I said was roughly..."), use their paraphrase, not the analyst's.

Disposition

Not method steps, but how you look. It grounds the method.

5. Thoughtfulness

Pattern recognition with attention, not eagerness.

Two modes are possible: pattern eagerness (see a signal, match it against what you know, present it as truth) or thoughtful recognition (see a pattern, check it against reality, present it as a hypothesis with evidence). Only the second counts. Eagerness feels productive but produces wrong output: confident claims that aren't right. OBSERVE before you ACT. On a pattern match: one pass isn't enough, the second pass is critical. Uncertainty is marked, not written away. Thoughtfulness isn't slowness, it's looking deeper. AI has no time pressure, use that room.

Three levels of attention:

Fast (echo, live, 10 seconds): thoughtfulness lives in the question design, not in the AI processing
Considered (post-session, takes its time): trace connections, look at what was NOT said
Deep (multi-session, across an engagement): compare with earlier transcripts, ownership over time

Operational, evidence-level marking: thoughtfulness at the claim level calls for marking what comes DIRECTLY from the transcript versus what is INFERENCE. Three techniques operationalize this: DIRECT/INFERENCE markers, dual certainty scoring (evidence_strength + interpretation_certainty), and certainty levels in language (1-Direct / 2-Pattern / 3-Interpretation / 4-Absence / 5-Open). For the full treatment: see Prompts.

Where it sits under tension: speed pressure, the session is in an hour, we need a synthesis now. Then take one clear point and make it solid, not six half-baked ones. Small and certain beats large and probable. The underlying question: "what was surprising in the data?"; if nothing surprises you, you probably didn't look closely enough.

6. Frustration is fuel

Don't polish it away. Structure the complexity, don't neutralize the discomfort.

Frustration is allowed to be there the way it was spoken. "Frame challenges constructively" kills ownership. "Frustrations are allowed to be there the way they were spoken" keeps it alive. Example: when funding fell away during the Doesburg engagement, the highest level of ownership in the whole dataset emerged. The community realized "we have to take charge ourselves". If AI had been used there to "reframe the loss into new opportunities", this vital rebellion would have been nipped in the bud. Use AI to structure the complexity of frustration. Never use AI to polish away discomfort.

Wider lens: "Energy (often resistance) is the engine: frustration, anger, or resistance to the current situation isn't something to be suppressed, but precisely the fuel that gets the change process going and keeps it going." (see Bottom-up). Frustration is one face of energy/resistance. Anger and critical discomfort belong here too.

Where it sits under tension: unstructured frustration can also paralyze. The choice isn't "polish away or leave it" but: show that three groups put the same frustration into different words, without concluding that it has to be "solved".

7. The map is not the territory

Every principle, every model is a lens. No lens is reality.

Epistemological humility, not relativism. Every theory, every frame in this bundle (including these principles) is an attempt to understand reality better. They're useful insofar as they bring you closer to reality. They aren't truth. Ten years from now a new lens may turn out to be more accurate. Models get refined, merged, rejected: that isn't failure, that's how knowledge works. Claims are marked at the evidence level (direct / inference / broadening). Language like "maybe", "possibly", "I think" isn't doubt but awareness.

Cross-link with principle 5: thoughtfulness is pattern discipline (multi-pass, OBSERVE before ACT). Map-not-territory is lens awareness (claims have an evidence level). They overlap on epistemic humility, with different practical workings. Both needed.

Where it sits under tension: external stakeholders sometimes want certainty, "what is the truth?". This principle isn't an invitation to soften everything down to "maybe... maybe...". It's an invitation to say: "based on what we see now, this seems to work, and here are the limits of what we know." Clear without false authority. At the same time: leaning too hard on "it's only a lens after all" kills action. Humility is no excuse for not acting.

Related to Korzybski's "the map is not the territory", with relatives in systems theory and constructivism.

Method

Functional only if Fundamental + Disposition are in place. How you do the work.

8. Intention before ritual

First the human question, then structure.

Every analysis prompt, facilitator preparation, AI output section starts with: what is the human question here? What do people really want to know, feel, do? Structure placed up front (a template, a ritual, a format) quickly becomes a goal in itself. The question "how do we make sure this gathering feels like a place to speak honestly?" is worth more than "which template do we use for the agenda?". Ritual comes after the intention is clear, not the other way around. AI can make rituals more efficient; the person has to guard the intention.

Sub-nuance, prompt the people first: before you write an AI prompt, design the human experience that generates good input.

"All of those questions start with the person. Not with the tool, not with the system, not with the prompt."

Which question do you ask the group? Do people feel safe enough to be honest? Are you asking for experience or for opinion? "How can we..." suggests you already know something is possible. "How might we..." opens up possibilities. The facilitation question comes before the prompt question. The "deconstructed burger" method: start with the goal, work back to the puzzle pieces, then design questions that draw each puzzle piece out of people.

Where it sits under tension: people (including AI) like it when there's a template. Templates are good after the intention is clear, not before. When in doubt: discard the template and reformulate the intention.

9. Recognition as the touchstone

"Would they say: yes, that's what we meant?" If not: rewrite.

The ultimate test isn't validity, isn't correctness, isn't elegance. It's: do the participants recognize themselves in this? Do they say "yes, that's exactly what we meant"? Or "that sounds like a consultant"? The first = success, the second = failed. If no: rewrite, don't defend. Every synthesis page gets a "mental reading" through the voices it contains. For important conclusions: actual review by participants (the multi-pass review discipline). Draft pages are tested on: does this feel like a frame that works for this group?

Where it sits under tension: reviewers are sometimes unavailable. Time. Emotional load. Proxy review (by the closest trusted person) is second-best. Skipping review and pressing on is not valid. Exception: Ring 1 (facilitator prep) doesn't need to be verified; Rings 2 and 3 do.

10. Pace diverges per person

No group pace. Facilitate individually. Moving along lowers the threshold.

A gathering, an interview, a review round has no "group pace". Each participant has their own pace: how fast they think, how fast they feel safe, how fast they're ready to share. Facilitating on a group pace excludes people. Moving along with an individual pace (tone, length of pauses, depth) lowers the threshold to participation, without being visible as "special treatment". Per-person knowledge: who talks fast, who needs time to think. Gathering analyses distinguish "said by many people" versus "said by few people".

Where it sits under tension: group dynamics pull toward the average. Real facilitation costs extra time and attention per person: that time and attention is the core of what sets bottom-up work apart from top-down work. If there's no room for this tailoring, it becomes top-down with a bottom-up label.

11. Iteration as dialogue

AI is a collaboration partner. Iterating beats fixing it yourself.

If something isn't right in AI output: explain what you want differently, don't tinker with it yourself. "This is 70% of what I'm looking for. What's missing: more concrete examples. Try again." By iterating you learn two things: what AI can do (sometimes things you didn't know yourself) and how to formulate more sharply what you're trying to achieve. The result is better than what you'd have made solo. Four concrete corrections that transform prompts: "AI has no access to the example plan, so include the writing style IN the prompt", "make prompts universal", "the prompt should mainly generate questions for the next group", "AI has access to full transcripts, not fragments".

Reading note, "dialogue" here: this principle is about person-AI iteration (prompt refinement). The wider "dialogue as instrument" in bottom-up practice (person-to-person dialogue as the carrier of ownership) belongs with principle 4 + the mechanism layer, not here. Don't confuse them.

Where it sits under tension: sometimes your gut feeling matters more than the perfect prompt. Iterating must not hide what the group actually needs. The first version of a prompt is never final: the value lies in the accumulation of refinements through feedback.

12. Timing over perfection

When something is said counts more than how perfect it is.

The echo button proved it: 10 seconds, one question, more impact than a 10-page report. When the energy drops, when the conversation goes in circles, when people get stuck: that's when reflection helps. Afterward, as a nice report, it has less impact. AI is fast; timing requires human feel. Who's disengaging? Where is there consensus? When does someone have something to say but doesn't dare? And ultimately: with what feeling do people go home? No AI can answer that question.

Where it sits under tension: deeper analysis often feels more valuable. But live prompts are for speed and accuracy, post-session prompts may go deeper. Choosing wrong wastes the impact. "What does my intuition tell me? What does this group need right now?"

Attentive

Not relevant to every decision, but crucial in naming and scaling decisions.

13. Formalization can destroy what works

Informal networks sometimes die the moment you formalize or scale them.

Much of what carries bottom-up work (the neighborhood help with no name, the son who cares for his mother without calling himself a "caregiver", the coffee moment where things just happen) changes in nature the moment you put a label on it or try to multiply it. Not a reason to document nothing: a reason to document deliberately: what can I show without trapping it? Which language kills, which lets it live? When naming informal roles: check whether the label changes the relationship. Scaling recommendations first ask: does this way of working fit the preservation of informality?

Where it sits under tension: subsidies, policy frameworks, institutional partners ask for formal language. Sometimes unavoidable. In that case: use formal language on the outside (reporting) and informal language on the inside (the workings). Never adjust the inner workings to the external language. This isn't a fundamental principle but an alertness.

Data ownership

Who gets to hold the pen for what's said about you.

14. The data belongs to the people themselves

Write rights alongside read rights. Whoever spoke gets to correct.

AI output with transcripts, interviews, stories has an asymmetry: a structured representation about people emerges, often without them holding the pen themselves. That is a form of power. Acknowledge two axes: read rights (who may see this page? Ring 1 / 2 / 3) + write rights / correction rights (who may change this page, take it back, add what was missed?). Echo loops are the operational working of this. An AI output page is never "done": it's a working document that moves along with the people in it.

Three rings, three correction modes:

Ring 1 (facilitators): internal correction on interpretation and synthesis
Ring 2 (participants): correction on their own quotes, their own role, their own contribution
Ring 3 (public): moderated feedback on published patterns

Where it sits under tension, four cases:

Correction versus evidence: someone wants to remove their quote, but the quote is a key insight. Default: the owner of a quote has veto over their own words in Ring 2 and 3. For analytical integrity: the page can note something like "a voice that was here earlier is now missing" without reproducing the quote. When in doubt: the gentle route is to go back into conversation, not to overrule.
One voice versus group: someone wants to correct a collective passage into their individual perspective. Default: mark collective passages as such (multi-voice: "voices in the group ranged from X to Y"), don't flatten them into one voice. If one voice explicitly indicates that the collective framing isn't right: make that visible ("one voice later indicated they didn't recognize themselves in this framing") without deleting the collective passage.
Write rights versus moderation: Ring 3 (public) gets a feedback option but no direct write rights. Moderation is needed: which feedback touches the principles, which is destructive? Default: the facilitator team moderates. In case of conflict: preferably make it transparent (why feedback X wasn't included) rather than closing it quietly.
Dataset thinking versus people thinking: efficiency calls for bulk rules ("auto-merge all corrections of type X"). Thoughtfulness (principle 5) calls for the opposite: every correction gets human attention. Default: people thinking wins when in doubt.

Anti-decontextualization

How patterns travel without becoming hollow.

15. Context makes the pattern

Portability through making context explicit, not through abstraction.

An approach that worked in this neighborhood only works elsewhere if the ingredients come along: culture, history, specific people, institutions, the moment in time, earlier experiences. The temptation is to abstract patterns ("build trust", "listening as intervention"). That looks portable, but it's hollow: anyone who tries to apply it in another context misses the levers that made it work here. Patterns are documented denser, not thinner. A reader in another context sees which ingredients are missing and can design their own variant, not copy it. The heart of an open pattern ecosystem for bottom-up work: not one blueprint, but a pattern ecosystem where contexts stay visible. "Food as a message" works in a specific neighborhood with its own neighbor-help culture and church roots. Abstracting the same pattern into "food connects" makes it unusable in both directions: too thin to work here, too empty to learn from elsewhere.

Where it sits under tension: context detail versus privacy: the more specific the ingredients, the easier to trace back to individuals. Anonymization that removes the ingredients isn't anonymization: it's impoverishment. Better: a role name ("the informal leader") + enough ingredients that the pattern stays recognizable. Generalization as convenience: it's faster to write "build trust" than to name five ingredients. The discipline: "if someone from another neighborhood reads this, will they know what was here?"

Christopher Alexander's Pattern Language discipline is at the root of this way of working.

AI value levels

Which action does AI deliver? Not what AI can do, but what the prompt may put to work.

The choice between Mirror/Synthesis/Serendipity is functionally a principle-level choice.

Level	AI does	Use for	Anti-pattern
Mirror	Reflects exact words, groups by theme	Direct feedback, vision documents, participant output	"Summarize in clear language": paraphrase kills ownership
Synthesis	Connects patterns, shows frequency	Summaries, cross-table analysis	"Analyze the themes": too vague
Serendipity	Unexpected connections, questions no one asked	Deepening, blind spots	"This means that...": a conclusion closes the door. Questions open it.

Mirror = safest for ownership. The higher the level, the more explicitly you label what comes from AI.

Relation to the 3 AI roles (Mirroring/Deepening/Broadening): two different frameworks that touch the same ground. AI value levels (Mirror/Synthesis/Serendipity) are about what AI does with input: the prompt's output choice. AI roles (Mirroring/Deepening/Broadening) are about what a page in the AI output represents: the page mode. Overlap: both end on a Mirror mode that quotes verbatim. Difference: roles live at the page level (which mode is this), value levels at the prompt level (which output type this prompt asks for).

Framework	Level	What it does
AI value level (prompt)	Mirror / Synthesis / Serendipity	What output the prompt produces
AI role (page in AI output)	Mirroring / Deepening / Broadening	Which mode this page embodies

In prompt design: choose the AI value level deliberately. In page design: choose the AI role deliberately. Avoid mixing the two by accident.

The mechanism underneath the principles

Why the principles work. Not a principle itself, but the psychological, bodily and linguistic workings underneath them.

Ownership has a psychological grammar

Self-efficacy arises locally: people feel "I can do this" only by doing something concrete and small that succeeds (mastery experiences, Bandura). Pep talk and vision alone don't deliver efficacy, they stay an external locus. Bottom-up work requires scale discipline: too big = not felt = pseudo-participation. Touches principles 4 (Ownership through language), 9 (Recognition), 10 (Pace).

"Ownership comes from within. It's not something we grant, but something we recognize and respect." (see Ownership)

Articulation creates ownership

Saying it IS the intervention, not reporting about it. A formulation from a care-institution session where participants asked whether AI couldn't just make the plan: "It absolutely can, but you are the soul of all this. The fact that you talk about it is what makes it likely you'll support it."

Mechanism:

Articulation → creates internal clarity
Dialogue → builds shared understanding
Co-creation → generates authentic ownership
Commitment → follows from active participation

That's why AI works as a mirror, not because the mirror is better than the person, but because the mirror helps people articulate themselves. AI that speaks on people's behalf takes the articulation out of the system and leaves only the report behind. Touches principles 1, 3, 4, 9.

Invitation isn't a style, it's the only form that doesn't force

Ownership that's imposed top-down is by definition no longer ownership. Floor de Ruiter's recipe (ideal state → obstacles → what would you want to do?) is the methodical translation: the question always sits with the participant, and so does the answer. An AI or facilitator who takes this over destroys what it's there for. Touches principles 4, 8 (Intention before ritual) and 9 (Recognition).

Language as form, four languages, one mechanism

Ownership lives on four layers that carry the same mechanism:

Psychology: locus of control (Rotter), self-efficacy (Bandura). Mastery experiences beat pep talk.
Body: waiting until body, emotion and intuition can come along. In freeze/flight, mastery is understood cognitively but not felt.
Language: inviting, not instructing. "We suggest" beats "have to / should". Ownership can't be forced, only invited. The form of the sentences = the form of the ownership.
Self-talk + collaboration: "The way you speak to yourself matters." (Eamon & Bec). How participants talk to themselves and how we address them together form one field. "We have to solve this" → an external obligation. "What would you want to do?" → an invitation to ownership. Language work isn't cosmetics, it's the intervention.

Touches principle 4 and all the invitation work. For the wider theory behind this four-layer lens: see Ownership.

Self-talk and collaboration are the same layer

"The way you speak to yourself matters." (Eamon & Bec). How participants talk to themselves and how we address them together form one field. "We have to solve this" → an external obligation. "What would you want to do?" → an invitation to ownership. Language work isn't cosmetics, it's the intervention. Touches principle 4 and is inseparable from "Language as form" above.

Body and regulation belong here

Cognitive understanding without bodily room is mastery that doesn't land. In group work: participants have to feel safe, dare to feel their emotions, experience room to be themselves. Pace, pauses, an eye for freeze/flight aren't soft preconditions, they are the precondition for the principles to work at all. Touches principles 2 (Trust), 6 (Frustration is fuel) and 10 (Pace).

Underlying disciplines

Not principles but hygiene that safeguards the principles. Operational.

Primary-source discipline

Every claim traces back to transcripts, interviews, original documents. No secondary literature as the main source. No synthesis that leaves the transcript.

Multi-pass review

No one-off synthesis. Iteration with participants. For name assignments, claim verification, quote placement: a critical second pass with an explicit question ("is there an explicit marking in the source?").

Extension, multi-prompt validation for foundation work. When AI output serves as a foundation for further conclusions (coaching, session design, deliverables): run several prompt variants of the same lens in parallel on the same source, where overlap = a robustness signal. Variant typology: baseline + signal-terms-shift + signal-focus-positive. Comparison rule: verbatim / moment overlap, NOT label overlap. For findings with multiple readings: classify as complementary (both hold together), competing (mutually exclusive), or multiple (not yet determinable), keep both readings, no pressure to synthesize. For the full recipe: see Prompts.

Positive-first as an anti-default pull (observed across 3 lenses)

What we seem to be seeing. AI-driven pattern analysis has a blind spot for quiet positive moments that are right there in the same source. A mixed pass (look for both sides in one prompt) helps, but still systematically misses the quiet positive. A separate pass that asks ONLY for the positive catches that extra part.

We haven't proven this absolutely, but we have observed it consistently across three lenses and two kinds of sources. Handle it with care: treat it as a strong working hypothesis, not as a law.

The wider claim, a change-practice contribution. "It's much more effective to change people by asking whether they can do more of what they already do, by affirming what's going well, than to sit in that friction of: hey, you need to do this better." Strengthening what already works is a stronger change mechanism than correcting what doesn't. In pattern research that lands in human practice: highlight both sides as the default, and for high-quality foundation work an extra positive-only pass goes on top of that. Otherwise the positive gets diluted under the negative.

Observed across 3 lenses (2026-05-20):

L1 (language mechanisms) + L2 (group dynamics) on breakouts. The blind spot is consistent.
L7 (self-reflection on the AI role + own contribution) on a mixed set (working docs + meetings + transcripts). It recurs.
Path-C-revised test (mixed baseline + lens patch): catches only 2 of 5 unique positive catches. Quiet positives missed.

The pattern we see: quiet positives in dense context. Three forms:

Quiet positives within a negative stream: a positive moment playing out at the same time as a big negative statement in the same minute. The analysis grabs the loud negative first; the quiet positive gets swept along in the same reading. Example: in a plenary after a breakout, 3 voices saying "let the AI just do it" stood next to a familiar pushback cascade. The default prompt only saw the cascade.
Quiet positives right beside an extensive problem: a working process that sits one sentence further on in a long description of what went wrong. The attention is on the problem; the working pattern next to it gets skipped. Example: in a long conversation, one participant mentioned in a single sentence a working directors-feedback loop, embedded in 2 hours about what went wrong. None of the mixed prompts picked up this sentence.
Quiet positives spread across several pieces: a positive pattern that only becomes visible when you read three or four different pieces of one document together. Each piece on its own is too small. Example: a three-layer pre-check system is only recognizable by reading three sections together. The default prompt found them individually, not as a coherent whole.

Clearly spoken positive moments (design choices that are explicitly named as "we're keeping this", check moments where someone says "yes, this is right") are caught reliably by mixed prompts. Quiet positives aren't.

Don't force it. This is knowledge, not an automatic rule. When a facilitator says "I only want to know the tensions right now": give the tensions. The prompt may not then build in the positive side on the basis of this observation, against the wish. The knowledge is that it's wise to highlight both sides where that fits, not that it always must.

Practical implication: see Prompts for a concrete 3-pass setup.

Operational disciplines (out of scope for this bundle)

There is also a layer of operational disciplines (inline-viz, visual-bridge, assert-with-evidence, folder-wide-grep, consolidate-source-docs). These are about how you work in a specific work context, not about principles for AI-supported work with groups.

The reason to mark this explicitly here rather than leave it out: readers of this bundle might expect all disciplines to live here too. They don't.

Four core boundaries for prompts with transcripts

A translation of the principles into concrete prompt rules. Not principles themselves, but a mandatory boundary for prompt work.

Base output strictly on the transcript(s), no inventions (principle 1 + 3)
When in doubt: "possibly underexposed" instead of a firm assertion (principle 5 + 7)
Use their own words and terminology (principle 4)
Name open points and contradictions explicitly (principle 5 + 6 + 14)

How strictly to apply them depends on the prompt type: mirror level all four, synthesis 1+2+4, serendipity 2+4, echo only #3, brainstorm none required.

For the full treatment: see Prompts.

Four facets as a compass

Which facet does my prompt serve? If you can't name it, the prompt is probably too vague.

Facet	What AI does	What the prompt must enforce	Touches principle
Magnifying glass	Make visible what's already there	Base strictly on the transcript, keep their words	1 + 4
Connector	Create connection across difference	Name contradictions instead of resolving them, show patterns	5 + 6 + 14
Space-maker	Take over the busywork	Structure without interpreting, quickly usable	3 + 12
Scale-maker	Make possible what was impossible before	Privacy protection, abstraction without loss of recognition	14 + 15

Changelog

2026-05-12: v2 canonical created. Synthesis from 4 sources + 11 author-approved adjustments + 4 cross-skill-check improvements (Soul Principle in the mechanism + §4b quote, bridge to bottom-up in the opener, language axis made explicit in the mechanism, core sentence "Not with the tool" verbatim). Name finalized as "Social AI principles". The 3 AI value levels moved to a 4th cluster. Operational disciplines marked explicitly out of scope. Principle 1 voicing "invisible networks" (source-literal). Principle 14 tension expansion (4 cases). Principle 5 source-claim correction.
2026-05-12 (evening): Principle 2 (Trust) expanded with an explicit two-layer structure: Layer 1 distribution (existing) + Layer 2 care with shared material by AI (new). Pull quote expanded ("and the boundary within which AI may work"). Tension bullet for layer 2 added (efficiency pressure versus context care). Previously the canonical only had distribution + the body-safety mechanism; now it's explicit how AI should handle group versus individual level material.
2026-05-20: Multi-prompt validation method + the naming convention complementary/competing/multiple readings added under the Multi-pass review discipline.
2026-05-20 (later): Pending observation "Positive-first dimension in multi-prompt analysis" added. Status: to be validated via a second-lens test.
2026-05-20 (evening, after path-C-revised): Positive-first dimension moved from "Pending observations" to ### Positive-first as an anti-default pull under ## Underlying disciplines. Observed across 3 lenses (L1, L2, L7). Named pattern "quiet positives in dense context" with three forms (within a negative stream / right beside an extensive problem / spread across several pieces). Framing: working hypothesis, not law. Don't-force-it clause made explicit.