TBM 418: Campfires, Trails, and Quests

John Cutler from The Beautiful Mess · email · 2026-04-22 Reader 🖥️

practical · argumentative — A practical essay (Substack post #7 in an AI series) arguing for a particular way of integrating AI into team work. It is argumentative in form, building from a contrast (broken use of AI vs. congruent use), through two concrete examples, into a cognitive-science frame (4Es), then into a new taxonomy (trails / quests / campfires) that prescribes how teams should organize collaboration with AI. The recommendation is explicit: AI should be the surface collaborators work on together, not a private cartographer.

Structure

Key terms

• multiplayer (vs. single-player) AI Multiplayer vs Single-Player AI
  Multiplayer AI is collaborative use where multiple people with different perspectives work on the same AI session/surface together (or asynchronously through shared traces). Single-player is one person curating their own context, prompting in isolation, and shipping the output as a deliverable.
• trails Multiplayer vs Single-Player AI
  The asynchronous traces team members leave for each other: context pointers, shared docs, evolving principles, markdown files in the codebase, customer JSON files. The persistent layer of the environment that shapes what the next person sees and asks.
• quests Multiplayer vs Single-Player AI
  The purposeful work done between campfires, sometimes solo but often together — pair prompting a thorny problem, two people with different expertise walking the same territory and seeing different things. The active, directed work of going somewhere.
• campfires Multiplayer vs Single-Player AI
  The convening sessions where the team brings traces and quest learnings back together — live reshaping of beliefs, collision of different eyes on the same problem, joint mapping of territory rather than mere reporting back from solo expeditions.
• 4Es (embodied, embedded, enacted, extended) Multiplayer vs Single-Player AI
  A cognitive-science framework treating cognition as embedded in environment, enacted through action, extended through tools and artifacts, and embodied through physical/sensory presence. Cutler uses it to diagnose what multiplayer AI sessions have that single-player ones lose.
• stigmergy Multiplayer vs Single-Player AI
  Coordination through traces left in a shared environment — ants and pheromone trails, termite construction, Wikipedia edits. Each actor's actions reshape the environment the next actor encounters, without central planning. Cutler frames trails as stigmergic coordination.
• context pointers Context Engineering
  Markdown files (or similar artifacts) added to a codebase that point AI and humans at the right context for specific tasks. A specific stigmergic artifact in Cutler's team's workflow.
• shared understanding that compounds Multiplayer vs Single-Player AI
  The output of good multiplayer AI use — not a deliverable, but distributed context that each person carries forward into their next conversation, decision, or piece of work, where it accrues further. Contrast with single-player output as a one-shot artifact.
• accelerated isolation Multiplayer vs Single-Player AI
  Single-player AI use viewed through the 4Es lens — productivity that feels collaborative because the model is conversational, but is actually solo cognition with a powerful mirror. The deeper the loop, the more narrow the resulting map.
• Dotwork
  A tool Cutler and his coworker use to pressure-test guiding principles for stable workstreams — explicitly framed as not replacing conversations but making them higher leverage.

Propositions

• explicitUsing AI to reinforce broken mental models of product development (PRDs that wither on contact with reality, box-checking business cases, cascading work nobody believes in) supercharges what was already broken — 'more artifacts, same blind spots.'conf: high
• explicitAI use is on the right track when it is congruent with the collaborative nature of product development — co-design, co-prompting, evolving contexts, and shared understanding that compounds.conf: high
• explicitcorroboratedAI is not the collaborator in good multiplayer use — it is the surface that collaborators work on together.conf: high
  supports ←Radar Trends to Watch: July 2026 · p1 — Concrete tooling instantiation (Claude Tag, Devin Desktop, Rayfin, Cosmos) of the same claim that good AI-mediated team work treats AI as the shared surface, not the collaborator.
• explicitThe output of good AI-mediated team work is not a deliverable but shared understanding that each person carries into their next conversation, decision, or piece of work.conf: high
• explicitcorroboratedSingle-player AI is 'accelerated isolation that feels like collaboration' — the map looks authoritative but no one walked the ground with others who would see different things.conf: high
  supports ←AI + Teams = ? · p1 — Both describe solo/parallel AI use as feeling productive while quietly eroding shared ground — this piece's 'collective debt' names the same phenomenon Cutler calls 'accelerated isolation that feels like collaboration.'
  extends ←John Cutler's Post · p1 — TBM-418 names a single-player/multiplayer binary ('accelerated isolation that feels like collaboration' vs. true multiplayer); Cutler's post sharpens this into three tiers by isolating 'shared-context solo' as a distinct, genuinely valuable middle state that is still not multiplayer — a finer-grained version of the same warning.
  extends ←Radar Trends to Watch: July 2026 · p4 — Names the missing infrastructure ('context lakes') for exactly the single-player-vs-multiplayer gap that doc identifies conceptually — this document proposes what would need to be built to close it.
• explicitcorroboratedWhen you curate your own AI context alone, you cannot question what you didn't think to put there — the embedded environment is shaped entirely by your own prior view.conf: high
  supports ←TBM 439: Day At The Gig... · p3 — His 'UNDO' after letting Notion's AI extend his map from workflows he hasn't personally verified is a lived instance of exactly the single-player-curation risk TBM 418 names abstractly: you can't question what you didn't think to put there.
• explicitAI can simulate multiple perspectives if you ask, but that is directing a puppet show, not a collision of genuinely different stakes.conf: high
• explicitMultiplayer collaboration does not have to be synchronous — stigmergy (coordination through traces in a shared environment) is multiplayer across time.conf: high
• explicitSingle-player AI is 'stigmergy for an audience of one' — traces only loop back to you, no one else's surprises compound for you and yours don't compound for anyone else.conf: high
• explicitHumans aren't ants — pure stigmergy (trails only) is insufficient; teams need campfires (convening) and quests (purposeful joint work) as well.conf: high
• explicitEach of trails, quests, and campfires fails in a predictable way without the other two: trails without campfires are a wiki nobody reads; quests without campfires are solo adventures that never connect; campfires without trails or quests are Groundhog Day.conf: high
• explicitThe key diagnostic question is bidirectional: are your traces feeding your conversations, and are your conversations updating your traces?conf: high
• explicitAI can power both halves of the trails-campfires loop, but only if there are other people in it.conf: high
• explicitThe 'go apart, come together' rhythm of high-trust collaborative teams predates AI — Cutler diagrammed it eight years ago for designer/developer collaboration. AI does not invent the pattern; it amplifies each phase.conf: high
• explicitUsed responsibly, AI does not make the back-and-forth rhythm of collaboration redundant — it makes it accessible to teams that could never pull it off before, and better for teams that already could.conf: high
• explicitdevelopedAI is best thought of (per Kent Beck, cited approvingly) as an imperfect, eager-to-please genie — useful but not autonomous, requiring direction and verification.conf: high
  extends ←TBM 428: Yes, Robot. Yes, Boss? · p4 — Extends Beck's 'eager-to-please genie needing direction/verification' into the interpersonal/organizational domain: the genie faithfully grants bad-faith intent, not just imperfect technical wishes.
• explicitThe 4Es (embedded, enacted, extended, embodied) hold not only in real-time interaction but across time through traces — the framework extends naturally to asynchronous stigmergic coordination.conf: high
• explicitdevelopedSolo AI work is deep but narrow — AI extends your cognition powerfully but holds only what you gave it, not what three people with different expertise would have given it.conf: high
  extends ←TBM 439: Day At The Gig... · p6 — Adds first-person testimony to TBM 418's abstract claim that solo AI work is 'deep but narrow' — it holds only what one person gave it, unlike what a coworker with different, tacit expertise would add.
• explicitThe value of multiplayer sessions lies partly in non-cognitive embodied cues — frowns, 'wait, go back' moments, trading the steering wheel — which are 'not incidental to the outcome; they were the outcome.'conf: high
• explicitAI raises the value of pre-session externalization: people arrive at co-design sessions with context already out of their heads, making the joint session higher leverage.conf: high
• explicitAI tightens iteration cycles by allowing the context graph to update in real time rather than 'whenever someone remembers to update the wiki.'conf: high
• implicitMultiplayer AI work transforms the role of AI from cartographer to terrain — from an authority producing maps to a shared surface where multiple walkers compare what they see.conf: high
• implicitestablishedThe fundamental danger of AI in knowledge work is epistemic narrowing — confident outputs from a context curated by one mind, with no mechanism for the surprises that other minds would have introduced.conf: high
  supports ←Design AI Systems That Actually Strengthen Human Reasoning · p9 — Both name AI-induced homogenization/epistemic narrowing of ideas as a distinct organizational risk, independent of individual output quality.
  supports ←TBM 439: Day At The Gig... · p4 — His 'meh' answers from querying AI over his own curated artifact pile, unable to distinguish live from stale, is a concrete instance of the 'epistemic narrowing' TBM 418 describes — confident outputs from a single-mind-curated context, with no mechanism for outside surprise.
• implicitAI is best diagnosed not by what it produces but by what loops it participates in — whether its outputs feed back into shared environments that other people then reshape.conf: high

Arguments

Narratives

• Lane Coherence: Cutler and a coworker use Dotwork to pressure-test the guiding principles of their stable workstreams, checking where confidence has shifted and whether upcoming work is still in the right sequence. They take that readout into a review session where they reshape the principles to reflect current beliefs. AI doesn't replace the conversation — it makes it higher leverage.
  → Concrete example of trails feeding campfires — the Dotwork readout is a trace; the review session is a campfire where principles get reshaped. Demonstrates the bidirectional loop Cutler later defines.
• Co-Prompting: Mid-essay, Cutler takes a break to tackle a date problem. His coworker had been adding markdown files as context pointers in their codebase. Cutler opens the codebase in Cursor, drops in a customer JSON file from his earlier prototype, and the two pair-prompt — one with technical eye, one with customer eye — while the coworker notes gaps in his context pointers and where they were burning tokens. In twenty minutes they both learn substantially, generate options to discuss with others, and Cutler can immediately talk intelligently about system architecture to a third coworker minutes later.
  → The piece's central illustrative narrative. Supports the multiplayer thesis, the stigmergy concept (context pointers prepared in advance), the embodied/enacted dimensions (pair prompting), and the 'shared understanding that compounds' claim (the immediate downstream conversation).
• The eight-year-old designer/developer collaboration diagram: Cutler diagrammed what designer/developer collaboration looked like on a high-trust team — joint prep, kickoff, co-design together, then individual research, pairing test, split for deeper work, reconvene, iterate, release. No AI involved. The rhythm of 'go apart, come together, go apart, come together' was the point.
  → Establishes that the trails/campfires/quests pattern is not novel to AI. Provides credibility for the claim that AI amplifies rather than invents collaborative rhythm.
• Ants and pheromone trails, termite construction, Wikipedia: each editor or insect reshapes the environment the next one encounters. No central plan, just accumulated traces.
  → Analogical scaffolding for the stigmergy concept. Makes asynchronous multiplayer coordination intuitive by anchoring it in familiar biological and digital examples.
• Kent Beck's framing of AI as an 'imperfect, eager-to-please genie' — invoked by name as a useful frame for how Cutler and his coworker use AI.
  → Borrows authority and a vivid metaphor. Positions AI as something to be directed and verified, not trusted as an autonomous collaborator.

Questions addressed

• ✅ What distinguishes AI use that reinforces broken practices from AI use that genuinely improves team work?
  AI use that produces more artifacts inside the same blind spots (PRDs, business cases, cascading decompositions) supercharges what was already broken. AI use that is congruent with collaborative product development — co-design, co-prompting, evolving shared context — is on the right track.
• ✅ Why does multiplayer AI use produce better outcomes than single-player use, even when single-player AI is powerful?
  The 4Es framework explains it: multiplayer sessions are richly embedded in shared context, enacted through real-time redirection, extended through artifacts beyond any one head, and embodied through the physical presence of different stakeholders. Single-player loses each of these — most critically the 'collision of different stakes' that no simulation can replicate.
• ✅ Is multiplayer coordination only possible when people are in the same session at the same time?
  No. Stigmergy — coordination through traces in a shared environment — is multiplayer across time. Context pointers, shared artifacts, evolving principles function as trails that shape the next person's reasoning.
• ✅ If trails are enough for ants, why do human teams need more?
  Humans have more to convene about than which path to take. They need campfires (synthesis, joint belief reshaping) and quests (purposeful joint work) in addition to trails. Pure stigmergy underspecifies what humans need from collaboration.
• ✅ How do you diagnose whether your team's AI workflow is healthy?
  Check the bidirectional loop: are your traces feeding your conversations, and are your conversations updating your traces? Also check whether each mode (trails, quests, campfires) is present — each absence has a characteristic failure (unread wiki, disconnected solo adventures, Groundhog Day conversations).
• ✅ Does AI make collaborative rhythm (go apart, come together) obsolete?
  No. AI amplifies each phase of the rhythm — sharper prep, higher-leverage co-design, richer traces during deep work, more memory in pairing tests, tighter iteration. It makes the rhythm accessible to teams that couldn't pull it off before, and better for teams that already could.
• ⭕ When is single-player AI use legitimate?
  Cutler acknowledges in passing that he uses AI as a solo collaborator 'when it makes sense, and when I would annoy the crap out of a coworker,' but does not develop criteria for when single-player use is appropriate vs. when it produces accelerated isolation. The reader is left with the impression that single-player use is suspect by default.
• ⭕ How do you manage the volume of trails so they remain useful rather than becoming trace-pollution?
  Not addressed. The stigmergy analogy works for ants because pheromone trails decay; it works for Wikipedia because edits can be reverted and pages have stewards. Cutler does not engage how teams should curate, prune, or retire context pointers and other trails as beliefs and codebases evolve.

Gaps

• implicit_premiseCutler's argument depends on the implicit premise that multi-perspective input produces better understanding of complex product territory than single-perspective input amplified, regardless of how powerful the amplification. This premise is doing enormous work — it is the reason a 'cartographer for places you haven't visited' is a bad metaphor even if the cartographer is brilliant. The premise is plausible but not defended, and it is exactly the premise that AI optimists would contest: 'what if the AI cartographer can simulate enough perspectives well enough that the collision-of-stakes advantage shrinks toward zero?' Cutler dismisses this as 'directing a puppet show' but does not say why simulated diversity must always be epistemically inferior to embodied diversity.
  Read next: Cass Sunstein and Reid Hastie 'Wiser' on group decision-making and the conditions under which diverse groups outperform individuals; Helene Landemore 'Democratic Reason' on epistemic diversity; Andy Clark on extended/predictive cognition and where the boundary between simulated and real perspectives lies
• incompleteThe taxonomy of trails/quests/campfires is asserted as the carving of collaborative work but is not tested against alternatives. Where does individual deep work without AI fit? Where do decision points fit — are they a fourth mode, or part of campfires? Where does external customer contact fit — is talking to a customer a quest, or something outside the model? The neatness of the three-way split is rhetorically powerful but the boundaries are not interrogated.
  Read next: Etienne Wenger on communities of practice and the modes of engagement/imagination/alignment; Diana Larsen and Esther Derby's retrospectives literature on cadence design; Allen Holub or Jessitron on team patterns
  closed byThe Office Amenity That Actually Improves Teamwork — The article does not use the campfires/trails/quests vocabulary, but it directly answers that gap's open question, 'where does individual deep work fit?': deep work is not a fourth collaborative mode but the precondition the collaborative modes depend on, since effective teamwork requires protected solitude for independent thinking before and between collective work (p3, p7, p8).
• implicit_premiseThe piece assumes that the friction of collaboration is itself epistemically valuable — that 'wait, go back' moments and the trading of the steering wheel produce understanding that solo work cannot. The opposite hypothesis is plausible: collaboration friction is a cost teams pay because they have no better way to integrate perspectives, and AI that genuinely integrates multiple perspectives could deliver the integration without the friction. Cutler's strongest moment ('the back-and-forth... was the outcome') is also his most assertion-heavy.
  Read next: Karl Weick on sensemaking and the role of equivocality; the literature on collective intelligence (Anita Williams Woolley); critiques of meeting-heavy cultures (Cal Newport, 'A World Without Email')
• unacknowledged_questionCutler's prescription assumes psychological and organizational conditions for genuine multiplayer collaboration — trust, time, willingness to be redirected mid-stream, role flexibility. He notes these implicitly (his diagram is for 'a team with real trust') but does not engage with what teams should do when those conditions are absent. The advice 'make AI a shared surface' presupposes the willingness to share. In a low-trust team, single-player AI may be a rational defensive posture, and 'multiplayer or bust' could be advice that only works for teams that already don't need it most.
  Read next: Amy Edmondson on psychological safety as a precondition for collaborative learning; Patrick Lencioni's trust pyramid; research on remote/distributed team trust formation
• incompleteThe trails/stigmergy framing inherits a curation problem the analogy obscures. Pheromone trails decay biologically; Wikipedia has revert mechanisms and stewards. Context pointers in a codebase do not decay — they accumulate, can become stale, can contradict each other, and can mislead future readers and AI agents. The piece names trails as an asset without acknowledging the maintenance liability or the political question of who has authority to prune.
  Read next: Brian Foote and Joseph Yoder 'Big Ball of Mud' on accumulated context as decay; Ward Cunningham on technical debt; the documentation-rot literature (Daniele Procida's 'Diataxis')

Related documents

• AI Won't Save the Kingdoms We Built
  AI as Amplifier Context Engineering Multiplayer vs Single-Player AI
• John Cutler's Post
  AI as Amplifier Multiplayer vs Single-Player AI
• Experience Mapping Matters More the Faster You Move
  AI as Amplifier Multiplayer vs Single-Player AI
• TBM 439: Day At The Gig...
  Context Engineering Multiplayer vs Single-Player AI