NateWiki — Open Brain corpus
Home · The Doctrine · All topics · Search
priority 1 28 posts

Business & Operator Playbooks

The playbook

Nate's operator playbook rests on one claim that hasn't changed since May 2025: model quality is not your bottleneck, and hasn't been for a while. What separates businesses winning with AI from the ones burning budget on pilots is what you feed the model and what you make it responsible for — not which lab shipped the newest checkpoint.

The earliest posts in this cluster frame this as a data-plumbing problem: the "Narrow-Pipe Law" (stop-waiting-for-new-ai-models) argues a noise-free, purpose-built input stream turns a commodity model into a bespoke analyst, while a comprehensive-but-messy input starves even a frontier model of usable signal. Chat is a demo, not a workbench (chatgpt-doesnt-just-do-chat, 9-hard-truths) — real value shows up once you push models into code execution, cross-tool orchestration, and multi-turn threads that persist as reusable assets rather than disposable Q&A.

By late 2025 the emphasis shifts from "get your inputs right" to "get your specification and governance right." ive-reviewed-20-enterprise-ai-builds argues most AI projects fail because nobody defined what "correct" means before building — which claims the system may make, what evidence backs each one, and whether a confident wrong answer or a refusal is worse. dear-manager-your-ai-budget and how-ai-actually-works-at-startups make the organizational case: enterprises stall not from stupidity but from running pre-AI budget and governance processes against a technology whose ROI math and adoption curve don't fit those templates; the fix is bounded, measured pilots, not blanket bans or blanket mandates. beat-the-95-ai-fail-rate reframes the viral MIT failure study: hybrid architectures beat pure buy-or-build, systems must learn from usage instead of sitting static, and the shadow-AI workflows employees already built in secret are free validated user research, not a compliance problem to stamp out.

2026 sharpens this into a single idea: execution is becoming free, so judgment is the only thing left to compete on. 8-habits-worth-unlearning names the shift directly — the bottleneck moved from execution to clarity, ambition, distribution, and relationships, and most habits (permission loops, meetings, decks, planning-before-building) still protect a resource that's no longer scarce. the-capability-overhang-is-real documents the December 2025 convergence of long-running agents and orchestration patterns (Ralph, Gas Town, Claude Code's task system) and Sam Altman's own admission that even he hasn't changed his workflow to match what the models can now do. 6-practices-for-when-the-models-got-good operationalizes this for individual builders: adopt an engineering-manager identity toward your agents, stop over-preparing before engaging a model, learn to change altitude deliberately (dive into detail, then climb back to abstraction), and accept that taste and embodied experience can't be delegated or compressed even as raw production gets instant.

klarna-saved-60-million names the organizational version of the same failure: Klarna's AI succeeded at the measurable objective (ticket speed) and destroyed the unmeasured one (relationship trust) because nobody had encoded what the company actually optimizes for in a form an agent could act on — "intent engineering," the layer above context engineering. whoop-is-hiring-600-people makes the expansionary counter-case: cheap execution should multiply ambition, not headcount cuts, because markets that were "too small to staff" now pencil out.

The most recent posts (reusable-ai-agent, wrong-ai-default, your-team-spends-5-hours, use-ai-sensitive-files) are the tactical payoff: build reusable pipelines — context pack, ingest, normalize, cite, human gate — instead of one-off agents, so every build makes the next one cheaper; measure real workflows against real tools before making a procurement ask; and treat data sensitivity as a per-task judgment call rather than a document-level checkbox, since no single "clean" version of a file serves every future question.

Two constants run underneath all of it: arbitrage windows open and close continuously and never settle into a steady state (313-became-438000, beyond-seo) — pricing and parameter-status gaps compound for whoever moves first and lock out latecomers; and distribution — of trust, of workflow position, of being the entity an AI cites — now matters as much as raw capability, whether you're a consultant or a product competing with Zapier for the coordination layer.

Key moves

The posts

Related