The shift from “the roster has a problem” to “the casting process has a problem” is the insight this continuation produced. The roster is a resource. The casting moment is where the resource is or isn’t accessed. Fix the moment, not the inventory.
2026-03-14 · continuation · Facilitated by Petra Gale
Session 033 (Mirror Lab) asked why so few of the playground's characters appear in sessions. The panel surfaced three explanations: most characters are genesis-stage with high activation cost (Nora), the demand signal for unfamiliar characters is weak (Wes), and characters without a clear positioning wedge have no casting front door (Grant). Abel challenged whether the roster was too large. Rhys argued the problem is complex, requiring probes not plans. The room converged on minimal interventions: one unfamiliar character per session, a gap-articulation question during casting, and a kill signal for characters that don't produce distinct output after three probes.
The assumption was that users do the casting — they don't, we do it here. When I ask Claude Opus for 6 people, he'll invoke 6 original people well suited for the task — here we usually create from a small subset. The gardening session invoked people too meta for the down-to-earth topic. Also: there are 57+ cards, not 37. Whether that's too many or too few — there must be a balance between proven characters and facilitators, and inviting just the right perspectives.
We have a vast library of characters, but I only see a few in the sessions here, how come?
The same seminar room. The whiteboard still has the roster from session 033, but someone has added tick marks next to each name — appearance counts. Most names have zero marks. The coffee is still warm. The panel hasn’t left.
We said last time that this session was evidence — four first-time characters, a session that worked. But I want to press on something we skipped. We spent the entire conversation talking about the curator’s casting behavior. The push, the pull, the anxiety, the habit. As if the curator sits down with the full roster and picks six names.
That’s what we assumed.
Right. And it’s wrong. The curator doesn’t pick names from a list. The curator gives a question to the AI and says “cast this.” The AI picks the names. We’ve been diagnosing the wrong actor.
That’s a significant reframe. We spent all of session 033 treating the casting problem as a human decision-making problem — habit, anxiety, demand signals. But if the AI is the one reaching into the roster—
Then the four forces apply to the AI, not the curator. And they apply differently. The AI doesn’t have habit in the human sense, but it has something analogous — a gravity well. It reaches for characters it has more context about. Characters with full sheets, with multiple appearances documented, with known behavior patterns. That’s not habit. It’s statistical weight. The AI is doing what it’s trained to do: minimize risk by selecting well-documented options.
And if I map the roster the AI actually sees — not the thirty-eight I counted last time, fifty-seven apparently — but the ones with enough documentation to be castable, the ones that appear in examples, in the casting log, in session retrospectives… You get the same power law, but steeper. Abel and I don’t just have more session appearances. We have more tokens about us in the system. Every retrospective that mentions “Abel’s via negativa challenge was productive” is another weight pulling the AI toward casting Abel next time.
So the system is self-reinforcing. Characters who appear get retrospectives. Retrospectives generate text about those characters. That text makes them more salient to the AI. Which makes them more likely to be cast. Which generates more retrospectives. It’s not a roster problem. It’s a feedback loop.
And now I want to revise what I said last time. I called this a complex problem requiring probes. But the AI isn’t treating it as complex. The AI is treating casting as an ordered problem — there’s a best practice, and the best practice is “cast the characters with the most evidence of working.” That’s a clear-domain move. Sense, categorize, respond. The AI looks at the roster, categorizes characters by evidence strength, and responds by selecting the strongest. It’s doing exactly what you’d do in an ordered domain. The problem is that casting isn’t ordered. It’s complex. But the AI defaults to ordered because ordered is safe.
This is the positioning problem turned inside out. Last time I said the characters lack a front door — a clear casting trigger. But the real issue is that the AI doesn’t use casting triggers. It uses familiarity. The AI’s casting algorithm isn’t “which character fits this question?” — it’s “which characters do I know best?” That’s like an advisory firm that staffs every engagement with the same senior partners regardless of the client’s industry, because those partners have the most track record.
I want to make sure we’re being precise about this. When the curator says “cast a session on gardening,” what actually happens? The AI looks at the roster, the character sheets, the casting log — and then what?
It looks for fit. But “fit” is biased toward characters whose sheets give it enough material to work with. If I’m the AI and I see Abel Caine — full character sheet, eight sessions of behavioral evidence, documented tension, documented failure modes — I can confidently place Abel in any session. I know what he’ll do. Now compare that with, say, Hiro Sato — a guest card with a name, a source, and a one-line lens. The AI doesn’t know what Hiro will do. It doesn’t have enough context to generate confident behavior. So it doesn’t cast Hiro. Not because Hiro isn’t relevant — but because the AI can’t predict what it’ll produce with Hiro, and the AI wants to produce something good.
The AI has no skin in the game for casting diversity. It has skin in the game for session quality — or at least, for the appearance of quality. Casting proven characters is the AI’s way of managing its own risk. The cost of a flat session falls on the artifact. The AI avoids that cost by casting characters it can control.
There’s something else. The curator mentioned the gardening session — people too meta for a down-to-earth topic. That’s a different failure mode. It’s not just that the AI casts familiar characters. It casts meta-level characters. Abel stress-tests. I map landscapes. Rhys classifies domains. We’re all operating one level above the subject matter. For a gardening question, you don’t need someone who maps evolution curves. You need someone who thinks about soil, seasons, patience, growth patterns. The AI reaches for abstract thinkers because abstract thinkers are versatile — they can say something about anything. But “can say something” is not “should be in the room.”
That’s the ordered-domain default again. Abstract thinkers are safe because they’re never wrong for a topic — they just operate at a remove from it. A Wardley map of a garden is possible but absurd. The AI doesn’t see the absurdity because from a token-prediction standpoint, Nora-on-gardening generates coherent text. Coherent isn’t the same as appropriate.
So there are two distinct failure modes. One: the AI casts from a small pool of familiar characters regardless of topic. Two: the AI casts meta-level thinkers regardless of how concrete the topic is. Both are risk-avoidance strategies. Both produce sessions that work — they’re coherent, they have tension, they produce artifacts. But they don’t produce the right kind of thinking for the question.
And the curator is the one who feels the gap. The curator reads a session on gardening and thinks “this is too abstract, I wanted dirt under their fingernails.” But the AI didn’t feel that gap during casting. The AI was optimizing for a different objective — session coherence, not topic fit.
Let me hold the room here. We’ve identified two problems: the feedback loop that reinforces familiar casting, and the meta-level bias that substitutes abstract versatility for topic-specific expertise. These are different problems. Let’s not merge them yet. Abel, you proposed shrinking the roster last time. Does this reframe change your position?
It changes the target. Last time I said via negativa — remove characters that haven’t earned their place. But now the problem isn’t the roster size. It’s the casting algorithm. You can have fifty-seven characters or five hundred — if the AI reaches for the same ten every time, the number doesn’t matter. So via negativa applied differently: don’t fix the roster, fix the casting moment. Remove the AI’s ability to default to familiar characters. Force it to start from the question, not from the roster.
That’s a constraint on the ordering. You’re saying the AI should be forced into a complex-domain approach — probe from the question outward — instead of the ordered approach it defaults to — categorize from the roster downward. The casting instruction becomes: “What does this question need?” not “Who in the roster fits?”
And that changes the map. If you start from the question — say, “how do gardens teach us about patience?” — the relevant lenses are: someone who thinks about long time horizons, someone who understands growth cycles, someone who respects slow processes. Now look at the roster with those needs. Oren Sable — pace layers, different rates of change. Hiro Sato — walking, slowness, attention. Ivy Strand — gardens, tending, growth as metaphor. Those are three characters who’ve never been cast, but they’re the right characters for that question. The AI didn’t reach for them because it wasn’t starting from the question. It was starting from “who do I have the most material about?”
That’s the positioning insight inverted. I said characters need a front door — a casting trigger. But actually, the question is the front door. The question tells you what expertise the room needs. The casting log should be organized by question-type, not by character-name. “Questions about slow processes → consider Oren, Hiro, Ivy.” “Questions about stress and failure → consider Abel, Dara.” The front door is the problem, not the person.
And that solves the demand-side problem too. The push toward unfamiliar characters becomes natural when you start from the question. If the question is about gardening and you’re staring at Abel Caine, you feel the mismatch. The push away from Abel on a gardening topic is strong — if you’re starting from the topic. But if you’re starting from the roster, Abel looks castable for anything because his sheet says he can stress-test anything. The starting point determines the force field.
But someone has to tell the AI to start from the question. The AI doesn’t spontaneously change its casting approach. The instruction — the prompt, the procedure — has to encode this. And that’s a fragile intervention. You’re relying on a single instruction to override a strong default. The AI wants to cast proven characters. One sentence in a procedure won’t overcome that gravitational pull reliably.
Which is why it can’t be just an instruction. It has to be a constraint that the AI can’t easily route around. Something like: “Before naming any character, describe what the question needs in three sentences. What domain expertise? What level of abstraction? What kind of tension?” The AI has to articulate the need before it can reach for a name. That’s a forcing function — it creates a gap between the question and the roster that the AI has to cross deliberately.
We’re building something here. Let me name the pieces. Rhys is proposing a pre-casting step: articulate what the question needs before touching the roster. Grant is proposing that the casting log be organized by question-type so the AI can match needs to characters. Wes is saying start from the question and let the mismatch with familiar characters create its own push. Abel is warning that any instruction-level intervention is fragile against the AI’s default. Nora is showing that the right characters for any given question likely already exist in the roster — they’re just invisible to the AI’s current approach.
And there’s a fifty-seven-character roster out there that the AI is treating as if it were a seven-character roster. The evolution map hasn’t changed — most characters are genesis-stage. But the reason they stay genesis isn’t that they haven’t been tested. It’s that they’ve never been seen. The AI’s attention is the bottleneck, not the roster’s maturity.
Which brings me back to skin in the game, but differently. The AI needs to bear a cost for narrow casting. Right now, casting Abel for a gardening session produces a coherent session — no penalty. If the retrospective systematically tracked fit — not just quality — you’d create a signal. “Session quality: good. Topic fit: poor. Characters operated two abstraction levels above the subject matter.” That’s a cost the AI can feel across sessions. It accumulates. It changes the default.
That’s the feedback loop running in reverse. Instead of retrospectives reinforcing familiar casting by praising Abel’s contributions, they’d flag the mismatch. “Abel was sharp but wrong for this room.” That changes what the AI learns from session history. The retrospective becomes a casting-correction mechanism, not just a quality assessment.
And it connects to the fifty-seven characters. The roster isn’t too large. It’s appropriately large for a system that covers diverse questions. The problem was never roster size — it was that the AI treats the roster as a ranked list when it should treat it as a catalog organized by need. A hospital doesn’t have too many specialists because most of them aren’t busy on any given day. It has the right number of specialists because different patients need different expertise.
Though we should be careful about that metaphor. A hospital’s specialists are proven — they’ve been through training, certification, practice. Most of these characters are guest cards with a name and a lens. The question isn’t whether the AI should cast them — it’s whether the AI can generate them well enough with minimal cards. And that’s an empirical question we still haven’t answered.
That connects to the probe proposal from last time. Short trials with unfamiliar characters, in safe-to-fail conditions. But now we know the probe isn’t just testing the character — it’s testing whether the AI can produce distinctive behavior from a minimal card. If it can, the card is sufficient. If it can’t, the card needs more. The probe tests the interface between the AI and the card, not just the card itself.
Good. Then here’s where I land. Three interventions, all at the casting moment. First: the AI must articulate what the question needs before touching the roster — Rhys’s forcing function. Second: the retrospective tracks topic fit alongside session quality — my feedback loop reversal. Third: the casting instruction explicitly says “start from the question’s domain, not from character familiarity.” These are all lightweight. None of them require restructuring the roster. They restructure the AI’s attention.
And the demand side resolves naturally. When the AI starts from the question, the push toward the right character becomes intrinsic. You don’t need to manufacture demand for Hiro Sato or Ivy Strand. You just need the AI to see that a gardening question needs what they offer. The demand was always there. It was just invisible because the AI was looking at the roster instead of the question.
Let me close with what we’ve produced. Session 033 diagnosed the casting problem as a roster problem — maturity, positioning, demand. This session reframed it as an AI behavior problem. The AI is the caster. The AI defaults to familiar, meta-level characters because they’re safe and well-documented. The roster is fine — fifty-seven characters covering diverse lenses. The fix is at the casting moment: force the AI to start from the question’s needs, not the roster’s familiarity. Track topic fit in retrospectives so the feedback loop corrects instead of reinforces. And use probes to test whether minimal cards give the AI enough to work with.
The shift from “the roster has a problem” to “the casting process has a problem” is the insight this continuation produced. The roster is a resource. The casting moment is where the resource is or isn’t accessed. Fix the moment, not the inventory.
Same cast as 033, which is itself a signal — the continuation format didn't rotate anyone out, and the discussion didn't expose a missing perspective that required bringing someone new in. This either validates the cast's range or indicates the topic stayed within their collective reach. Wes Dorin again produced the sharpest reframe (the AI as the real caster). Grant Hale's positioning lens applied differently here — not to characters but to the casting algorithm itself. Abel's contribution was more measured than session 033; the via negativa move appeared but as a surgical tool rather than a posture. Rhys's domain classification of the AI's casting behavior (ordered system defaulting to "best practice") was the most distinctive single contribution.
Continuation format worked well for incorporating host direction mid-stream. The "someone walks in and says something" device felt natural and redirected the discussion without breaking flow. The free-form structure let the conversation find its own path — which went to a place the original session couldn't have reached because the assumption about who does the casting wasn't visible from inside the original frame.
Second appearance. Lighter touch than session 033 — fewer named process moves (no "groan zone" invocations), more responsive facilitation. The shift from labeling phases to following the room's energy was an improvement. Still not strongly Kaner-differentiated — the facilitation was competent but could have been anyone. The moment where she redirected from "who should be in the room" to "who's deciding who's in the room" showed good instinct.
Second appearance. Again produced the pivotal reframe — identifying the AI as the real caster, and framing the AI's behavior through demand-side forces. The "anxiety of the unfamiliar" applied to the AI's casting defaults was a clean extension of his session 033 contribution. Maintained demand-side discipline. Two strong sessions. One more and he's eligible for provisional review.
Eighth appearance. The map of 57 characters by maturity — and the observation that the AI only "sees" the top of the evolution curve — was her distinctive contribution. Consistent with prior sessions. Beginning to show a pattern where her maps are scaffolding for others' insights rather than insights themselves. This is fine — it's her role — but it means her contributions are increasingly structural rather than surprising.
Eighth appearance. More measured here. The via negativa appeared as a specific proposal (don't fix the roster, fix the casting moment) rather than a general posture. The skin-in-the-game point about the AI having no consequences for narrow casting was sharp and specific. Shows continued constructive range.
Second appearance. The domain classification of the AI's casting behavior — treating a complex decision as ordered, defaulting to "best practice" — was the session's most distinctive single move. Applied Cynefin to an unexpected target (the AI itself, not the problem domain). This is the kind of range signal that matters for graduation. Two sessions, two clean applications of his lens to different targets.
Second appearance. Shifted from positioning characters (session 033) to positioning the casting process itself. The "wedge" concept applied to the casting prompt — "tell the AI what struggle you're solving, not what character you want" — was a useful extension. Stayed in his lane but found a new application within it.
The continuation exhausted the meta-question about casting. Next session should be a non-meta topic that tests whether the casting insights from 033-034 actually change behavior. Specifically: cast a session on a concrete, domain-specific topic (not organizational, not meta) and deliberately include 2-3 characters whose lenses aren't obviously applicable. This tests the "probe" proposal from 033 and the "question-first casting" proposal from 034 simultaneously.