TBM 439: Day At The Gig...
John Cutler from The Beautiful Mess · email · 2026-09-11 Reader 🖥️
AI SlopContext EngineeringKnowledge BaseMental Models & UnderstandingMultiplayer vs Single-Player AIOrganizational Legibility
AI can accelerate the mechanical formalization of sensemaking artifacts a person has already understood, but it cannot solve the harder problem underneath — telling which of an organization's proliferating artifacts still reflect current reality — so genuine organizational context remains something you have to actively create through your own exploration and other people's tacit knowledge, not something AI can spoon-feed you.
theoretical · narrative — Cutler explicitly declines to turn the experience into 'a neat argument' and instead narrates one diary-style day. But the narration is in service of understanding, not action-guidance: the day's episodes function as evidence for a set of claims about how context is created, how AI does and doesn't help, and why artifact staleness is hard to detect — made explicit in the closing 'reflections' section. It is theoretical dressed as narrative: the propositions are mostly implicit, carried by the story rather than stated as advice.
Structure
| Part | Role |
|---|---|
| Framing note before the diary (italicized intro) | context |
| Morning: building the 'subway map' by hand, converting it into Notion tables, then resisting letting Notion's AI extend it directly ('UNDO. UNDO. UNDO.') | narrative evidence |
| Midday: querying AI across the pile of artifacts (docs, transcripts, Miro boards, videos) and getting 'meh' answers; the explicit puzzle of figuring out what is 'still valid' | evidence + direct statement of the central problem |
| Context-sharing sessions with veteran colleagues, described as 'extremely helpful,' contrasted with AI's answers | counter-evidence / contrast case |
| Afternoon: PowerPoint friction, Linear tickets found insufficient for a newcomer, code treated as the one definitive artifact, AMP tool aside, Miro-acquisition aside | further narrative texture / minor evidence |
| Closing 'Some reflections' section: on context creation vs. spoon-feeding, the 'context battleground' marketing framing, erosion, and AI Slop Cascade | conclusion — the only place the day's implicit propositions are stated directly |
Key terms
- subway map
A personal, hand-built network diagram (stops, lines, transfer points) that Cutler uses to model a product's end-to-end shape as a customer/data/artifact journey — a private sensemaking tool, not a literal transit map. - sensemaking artifacts
The broad, proliferating class of self- or team-generated documents and diagrams (journeys, map types, one-pagers, atomic insights, goals, 'Jobs,' principles) produced in the course of understanding a domain, as distinct from execution artifacts like tickets. - live (artifact)
An artifact that is continuously updated and still reflects current understanding — as opposed to a 'relic' or 'cave drawing,' a frozen artifact whose accuracy has quietly expired without anyone marking it so. - context battleground
The marketing framing (which Cutler attributes to hearing others use it, e.g. around Notion) that treats owning/storing a user's or org's 'context' as the competitive prize — which he distinguishes from, and finds less useful than, the harder, unmarketed problem of context creation and preservation. - AI Slop Cascade
Cutler's coinage ('a three-times-removed AI Slop Cascade') for context or knowledge that has passed through multiple layers of AI summarization/regeneration, each pass eroding groundedness — contrasted with context built through his own active exploration. - archaeology
The effortful, self-directed work of reconstructing situated, tacit understanding of a product/org from a pile of artifacts plus conversations, when there is no single authoritative source to simply read.
Propositions
- explicitSensemaking artifacts (maps, one-pagers, insights, journeys, etc.) outnumber tickets roughly 8:1 in his experience, and grafting them onto tickets as labels or references always feels disconnected.conf: high
- explicitOne sensemaking artifact is never sufficient — real understanding requires many different, simultaneous artifact types (journeys, maps, docs, atomic insights, metrics, one-pagers, goals, 'Jobs,' principles).conf: high
- implicitHe trusts AI to mechanically formalize a sensemaking artifact he has already personally understood (turning the hand-built subway map into Notion tables), but not to extend that artifact into territory he hasn't personally walked — evidenced by his immediate 'UNDO. UNDO. UNDO.' after asking AI to add lines to the map from its knowledge of other workflows.conf: highsupportsTBM 418: Campfires, Trails, and Quests · p6 — 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.
- explicitQuerying AI across a heterogeneous pile of artifacts produces mediocre ('meh') answers, largely because it is very hard — for AI and often for him too — to tell which artifacts are 'live' (continuously updated, current) versus relics.conf: highsupportsTBM 418: Campfires, Trails, and Quests · p23 — 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.
- explicitThere is a real risk of building shared understanding on top of already-eroded shared understanding, or worse, on understanding that is glossy and coherent-sounding but ultimately wrong — i.e., fluency is not evidence of validity.conf: highsupportsManagers Are Struggling to Keep Up with the AI Productivity Boom · p8 — His 'glossy, coherent-sounding, but ultimately wrong' worry matches the HBR piece's claim that AI summaries are functionally useless because they flatten signal and slop indistinguishably.
- explicitTen minutes with AI feels expansive in breadth and ability to answer questions, but it does not have the same 'flavor' as ten minutes with a new coworker who holds a lot of implicit context — AI breadth and human tacit knowledge are qualitatively different, not interchangeable, resources.conf: highextendsTBM 418: Campfires, Trails, and Quests · p18 — 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.
- explicitContext erodes continuously and immediately — starting the moment the day ends — so any 'brain'/knowledge repository built from it needs constant replenishment rather than being a one-time construction.conf: highsupportsAn LLM wiki changed how I work · p2 — Both pieces converge on knowledge repositories requiring continuous upkeep rather than one-time construction — Cutler frames it as context eroding daily, Newton frames it as maintenance comparable to a vintage sports car.
- implicitThe current AI/product marketing discourse around 'context' as a battleground repeats the framing of earlier failed knowledge-management efforts (source of truth, knowledge repositories, connected data) by emphasizing context possession/storage over context creation and preservation, and risks failing for similar reasons.conf: medium
- implicitSource code is treated as more definitively trustworthy than most process/sensemaking artifacts, because it is what's actually in production — whereas docs, insights, and journeys can silently go stale without anyone marking them as such.conf: medium
- explicitThe active process of creating context through his own exploration — as opposed to having it delivered by AI — is itself valuable independent of the artifact it produces; he 'feels alive' doing it.conf: high
Arguments
AI, as currently used, cannot solve the central onboarding problem of determining which artifacts in an organization's pile still reflect current reality. strength: medium
Explicit premises:
- Querying AI across his artifact pile yields mediocre ('meh') answers
- It is very hard to tell which artifacts are 'live' versus relics even reading them directly himself
Implicit premises:
- AI has no independent way to verify an artifact's current validity beyond signals already present in the artifact or its metadata
- Confident, fluent AI output can mask staleness rather than reveal it (the 'glossy, coherent-sounding, but ultimately wrong' failure mode)
Inferences:
- A trustworthy AI assistant over organizational knowledge would need explicit provenance/freshness/consensus signals that his current tools don't expose
- The artifact-staleness problem is a documentation-governance problem that AI-generated content amplifies rather than resolves
Real organizational context must be actively created — through exploration, conversation, and 'archaeology' — rather than passively received from AI, and that active creation carries value beyond its output. strength: medium
Explicit premises:
- He 'feels alive' trying to make sense of things
- Ten minutes with a new coworker's tacit context has a different 'flavor' than ten minutes of AI breadth
- Context-sharing sessions with veterans, where he takes furious notes, are described as 'extremely helpful'
Implicit premises:
- Passively consuming AI-synthesized context does not produce the same understanding, ownership, or memorability as self-directed exploration
- Part of what's valuable in a veteran colleague's answer is their implicit, silent judgment about what's still relevant or current — not just the facts they state
Inferences:
- Onboarding approaches that hand new hires an AI-generated 'org brain' instead of time with veterans and self-directed exploration will underperform
- The value of tacit-knowledge transfer scales with how much implicit relevance/currency judgment the veteran silently performs — something hard to elicit from an AI system that wasn't there to see the context change
Narratives
- Across the day, Cutler builds a hand-crafted 'subway map' of the product, converts it to Notion tables with AI's help, then catches himself before letting Notion's AI extend the map itself from its own knowledge of other workflows, hitting 'UNDO. UNDO. UNDO.'
→ Shows rather than argues the boundary he implicitly draws between AI as a formalizer of understanding he already has and AI as an originator of understanding he hasn't personally verified — supports proposition 3 and Argument A. - He's flooded with heterogeneous artifacts (meeting transcripts, notes, Miro boards, usability-test videos, 'insert 10 other artifact types') and finds AI queries across them 'meh,' while a couple of live context-sharing sessions with veteran colleagues are 'extremely helpful.'
→ Narratively supports proposition 6 and Argument B — the contrast case that AI breadth and human tacit context are not interchangeable. - Mid-afternoon he hears about the Miro acquisition and notes it's 'one of my favorite ways to work with people' — a brief aside amid AI-tool frustration (PowerPoint's broken AI features, mediocre AI answers over his artifact pile).
→ A small but telling signal of which category of tool he values (collaborative, human-in-the-loop) versus the AI tools frustrating him that day — reinforces proposition 6's human/AI contrast, though thinly. - He has to use PowerPoint for the first time in years, doesn't recognize it, and finds its AI features unhelpful when he just wants to paste without formatting.
→ A minor narrative of AI-feature friction interrupting a simple mechanical task, reinforcing a broader skepticism toward AI feature bloat consistent with his 'rigorous adopter' stance.
Questions addressed
- ⭕ How does someone with zero context make sense of a new product/organization's full end-to-end shape, when tickets, existing docs, and AI aren't built for someone parachuting in?
The closest he gets is a hand-built personal artifact (the subway map) plus live conversations with veterans plus reading the code — not a systematic solution; he explicitly doesn't know if he'll keep the artifact or if anyone will ever look at it. - ⭕ How do you know whether an existing artifact (doc, insight, one-pager) still reflects current reality?
He names the question directly ('One puzzle... is figuring out what is "still valid"') but offers no method beyond talking to people who've been around and doing his own archaeology — no heuristic or signal for staleness is proposed. - ⭕ Will his own subway-map artifact become exactly the kind of stale, un-understandable relic he's worried about inheriting from others once he moves on or forgets its private notation?
Gaps
- incompleteCutler identifies that it's hard to tell which artifacts are 'live' versus 'relics' but offers no concrete heuristic or signal for making that determination — e.g. last-updated metadata, authorship, revision/consensus history, or explicit 'still valid as of' annotations. The piece's central practical problem is named vividly but left without even a candidate method.
Read next: His own TBM 418 'Campfires, Trails, and Quests' — specifically its unclosed gap that trails/context pointers, unlike wikis, have no revert/staleness mechanism — or documentation-rot literature (Diataxis, 'Big Ball of Mud') for maintenance-liability frameworks. - incompleteHe notes that live context-sharing sessions with veteran colleagues are 'extremely helpful' where AI queries over similar underlying material are 'meh,' but never analyzes why — is it the veteran's live relevance-judgment, two-way questioning, social trust, or information that was simply never written down at all? The mechanism behind the contrast is asserted, not examined.
Read next: TBM 418's campfires/trails/quests framework, or Nonaka & Takeuchi's SECI model of tacit-to-explicit knowledge conversion. - incompleteThe piece frames 'context battleground' AI marketing as reminiscent of past failed knowledge-management movements (source-of-truth, connected data) but doesn't engage with why those specific efforts failed — the analogy is asserted rather than argued, leaving open whether AI-era context tools would fail for the same reasons or genuinely different ones.
Read next: Classic knowledge-management-failure retrospectives (e.g. Davenport on why KM initiatives stall) to test whether the parallel holds.
Related documents
- An LLM wiki changed how I work
- Managers Are Struggling to Keep Up with the AI Productivity Boom
- Why AI agents need a Context Lake
- No Vibes Allowed: Solving Hard Problems in Complex Codebases – Dex Horthy, HumanLayer
- Experience Mapping Matters More the Faster You Move
- Presentation: From Retrieval to Reasoning: Building Production-Ready Agentic AI Systems with Knowledge Graphs
- Tony Seale's Post
- What 60 Years of Data Reveals About How Men and Women Experience Leadership
- Adrien Taravant's Post
- AI Won't Save the Kingdoms We Built