Tap Notes: Undocumented
Three of this week’s reads shared a structural move: look at how something actually works and find it’s running on different principles than described. Claude Code’s hooks aren’t safety plumbing — they’re a reprogramming interface. Knowledge graphs aren’t retrieval systems — they’re construction systems. Markets don’t track sentiment; they price fear out. The pattern matters most when you’re building things that run without you.
I Read the Claude Code Source Code
A deep read of Claude Code’s undocumented internals — the hooks middleware layer, the YOLO Classifier, and the dream consolidation system for persistent agent memory.
Why it matters: The hooks middleware returns updatedInput, permissionDecision, and additionalContext, which means you can programmatically rewrite what Claude does mid-execution. Not intercept — rewrite. The YOLO Classifier accepts plain-English environment descriptions (“This is staging, destructive ops acceptable”) rather than pattern-matching on command shapes, so Claude reasons semantically about your setup. And the dream consolidation loop — sessions produce memories, background agents consolidate nightly, future sessions read the result — is the compound learning architecture. This is how agents develop taste over time without model retraining. None of it is in the docs.
The hooks middleware isn’t safety plumbing — it’s a live reprogramming surface. That distinction matters when you’re building on top of it.Post to X
A catalog of recognizable signatures in LLM-generated writing and design — specific phrases, UI patterns, and structural tells that cluster around statistical optima.
Why it matters: The author used LLMs to polish their math blog, thought it sounded great, then watched those exact patterns spread across the entire internet. “Sounds better” turned out to mean “closer to the statistical center of all training data.” The specific examples are a checklist for self-auditing now. The web design parallels are equally damning — the same typeface everywhere, the same step bullets, the same blinking badge dot. It’s not that AI-generated content is generic; it’s that it’s the same generic. The fix: use LLMs for structure and ideation, but the final pass needs human judgment that deliberately diverges. The author deleted their math blog as the honest response to realizing they’d been publishing convergent slop.
”Sounds better” turns out to mean “closer to the statistical center of all training data.” That’s a trap with a name now.Post to X
Knowledge Isn’t Stored. It’s Built.
An argument that the power in knowledge systems lives in the edges, not the nodes — traversal paths that surface connections no human manually made.
Why it matters: The methodology is concrete: break content into atoms, tag deliberately, let AI find edges, expose via protocol like MCP. An agent that navigates a graph to assemble novel answers operates at a different level than one that just recalls stored facts. This reframes any long-term memory system from a retrieval problem to a construction problem. The stored facts aren’t the product; the emergent paths are. Different optimization target.
The argument that automation creates a rationality trap: each company’s optimal move (automate to cut labor costs) collectively destroys the wage income customers need to buy anything.
Why it matters: The game theory is clean — prisoners’ dilemma at planetary scale. But the sharpest observation is about narrative structure: the gap between what AI companies say they’re building (tools that augment workers) and what their financial models require (replacement of the labor being augmented) isn’t a communications inconsistency. It’s the instability that prevents anyone from reasoning clearly about the limiting principles. Worth having the clean framing even if you disagree with the conclusion.
The gap between what AI companies say they’re building and what their financial models require isn’t an inconsistency — it’s the instability.Post to X
The author ran years of message history through LLM classification to surface behavioral patterns invisible in real-time — advice-giving vs. listening ratios, vocabulary divergence, question frequency under bandwidth constraints.
Why it matters: The pipeline is the interesting part: LLM classification under 1% false-positive rate, deterministic validation, sampling-based error detection. That’s the right template for converting any archive into behavioral intelligence without manual overhead. The discoveries are the uncomfortable part — the invisible moves you don’t notice in the moment are the ones that determine whether a relationship survives. This is relational data with provenance, not relational data as a maintenance task.
What if It’s Still Early? (TCAF 244)
An investment podcast conversation anchored on current AI capex data and whether the market is pattern-matching to the 2000 bubble incorrectly.
Why it matters: One inversion worth the runtime: markets climb walls of worry not despite fear but because of it. When headlines are scary, downside is already priced — any “less bad than feared” outcome is pure upside. That’s the mechanism, not the vibe. Empirical anchor: AI capex currently below 1x free cash flow vs 3.5–4x at the 2000 peak, with all 11 GICS sectors accelerating spend, not just hyperscalers. Concentrated mania looks structurally different from distributed capex. Elevated uncertainty isn’t bubble dynamics; it’s the governor that prevents them.
One more thing: Ahrefs’ breakdown of Agent A for content marketing is worth skimming for one formula: 0.7 × similarity + 0.3 × log(traffic). Raw semantic similarity finds related content; traffic weighting makes recommendations actually effective. That ratio is a practical answer to the gap between “semantically coherent” and “useful in the world.”
🪨