NateWiki — Open Brain corpus
Home · The Doctrine · All topics · Search
priority 1 32 posts

Prompting Technique Guides

The playbook

This is Nate's core discipline, and reading all 32 posts in order shows a real evolution rather than a static bag of tricks — the same way a craftsman's tool list changes as the material changes. Late 2024 through early 2025 is fundamentals: stop treating ChatGPT like a search box, and use a six-part skeleton (role, task, context, format, examples, constraints) that shows up, relabeled, in nearly every later post. February 2025's "Spellbook" organizes technique into four pillars — self-review/iteration, multi-stage/multi-model chaining, logic/precision, perspective/style — and by March he's formalizing multi-model orchestration explicitly: use GPT-4o to brainstorm broadly, a structure-oriented model to outline, a reasoning model to fill in logical depth, and a research-grade model to synthesize and cite, switching deliberately whenever a model shows logical errors, quality drift, or a knowledge wall rather than sticking with one model out of habit.

Spring 2025 is where the craft gets genuinely systematic. He builds a "four-block" blueprint (Purpose, Instructions, Reference, Output) plus two meta-switches — MODE (reflection/action/agentic) and EFFORT (how hard to think) — specifically because reasoning models like o3 default to bias-for-action and anchor hard on your first framing; you have to explicitly declare intent before the model will explore instead of execute. He also derives, from actual transformer mechanics, why prompt order matters as much as content: early tokens steer the whole probability space, output-format specs act as soft stop-tokens and belong last, and long chats need guardrails re-pinned every dozen or so messages because attention decays toward recency.

By June 2025 the single biggest turn happens: prompting itself gets reframed as one layer inside a bigger discipline he calls context engineering — the deterministic context you write versus the probabilistic context an agent autonomously discovers via search and tools, where your literal prompt can be a fraction of a percent of what the model actually processes. That reframing recurs and sharpens for the rest of the run. August's GPT-5 system-prompt teardown reveals a deliberate "bias to ship" (one clarifying question, then execute, no permission-seeking) that rewards front-loaded assumptions and tool-policy declarations over conversational back-and-forth. September's meta-prompt framework (role, objective, methodology, output contract, constraints, uncertainty protocol, self-check) and the "prompt doctor" post's six diagnosable failure modes (underspecification, regeneration-scope confusion, reasoning collapse, hallucination, consistency drift, context overload) both push toward treating prompts as enforceable specs, with structured-output schemas as the escalation path when prose isn't reliable enough.

Late 2025 keeps decomposing prompting into more specialized moves: Claude Skills package reusable methodology as loadable files rather than re-typed instructions; "Goldilocks prompting" argues most tasks want a ~500-token middle ground of principles and heuristics, not procedures, reserving fully deterministic prompts for machine-to-machine handoffs; the Gemini-3-vs-GPT-5.1 piece frames model choice as an entropy tradeoff (feed messy multimodal chaos to the context-entropy specialist, feed clean inputs and a hard task to the task-entropy specialist); and the six-stage prompt-lifecycle piece names the field's actual gap as intent formation, not wordsmithing — nearly all tooling serves drafting-through-deployment, almost none helps you figure out what you want before you start typing.

The most recent posts (Feb–Mar 2026) are the payoff of that whole arc. "Why your AI output feels generic" names the mechanism directly — RLHF optimizes for a median rater, so default output is calibrated to nobody — and pushes past prompting entirely into memory, persistent instructions, tool/app connections, and style profiles as compounding steering levers. "Prompting split into 4 skills" then reframes the entire cluster's history as a stack: Prompt Craft (the 2024-era conversational skill) is now table stakes; Context Engineering curates the environment; Intent Engineering encodes organizational values so agents don't optimize the wrong thing (the Klarna speed-over-relationship failure is the cautionary case); and Specification Engineering — writing complete, self-contained, verifiable specs that survive multi-hour unsupervised agent runs — is the frontier skill, because synchronous chat-style correction stops working once agents run for hours without you watching. The latest post pushes one layer further still: your rejections of AI output, not your prompts, are the compounding asset, because the constraint you articulate when you say "no, this is wrong because X" is the only part of the interaction that doesn't evaporate.

For an operator running a fleet of agents, the through-line is: prompting-as-clever-wording was always a bridge skill. What survives every model generation is structure, explicit intent, enforceable specs, and a captured record of what you rejected and why.

Key moves

The posts

Related