NateWiki — Open Brain corpus
Home · The Doctrine · All topics · Search
priority 1 24 posts

Task-Specific Prompt Packs

The playbook

This cluster is Nate's grab-and-go layer: instead of teaching prompting as a skill in the abstract, he ships a finished, copy-paste artifact for one concrete job — Excel, PowerPoint, negotiation rehearsal, image generation, accessibility audits, financial decisions — and uses the writeup to explain why the prompt is built the way it is. Read across all 24 posts, a single design philosophy holds the whole cluster together, and it sharpens steadily from April to December 2025.

The foundation, laid down in the very first post (the 16-prompt work stack, April), is an explicit prompt architecture: context first (who/what/why), then defined output format, then an instruction to be interrogative (ask clarifying questions instead of guessing), then constraints, reference examples, iterative refinement, an assigned role, edge-case stress-testing, tone/depth, and a final evaluate-and-refine loop. Almost every later pack in this cluster is a specialization of that same skeleton. The negotiation-rehearsal prompt (July) turns it into an explicit state machine — a builder phase that asks nine questions one at a time with confirmation echoes, then a hard handoff into a host phase that runs the simulation — because letting one prompt try to gather info and roleplay five characters at once collapses into generic mush by round three. The AI-tutor post (July) runs the same state-machine idea through two deliberately different builds (hard mode: batch all questions upfront for power users; easy mode: single-question, 250-word micro-lessons for beginners) to show that how you sequence disclosure is a bigger lever than anything else in the prompt.

A second throughline is constraints-over-instructions: the business-writing pack (October) argues good prompts should spend far more words on what to exclude than on what to include, because models default to comprehensive and vague, and only hard limits (word counts, banned sections, required self-checks) force real prioritization. The anti-slop pack (October) and the deliberate-practice pack (December) both push this further into self-scoring — bake a concrete rubric into the prompt so the model grades its own draft against named failure modes before a human ever sees it, because human review cannot scale to 10-100x AI content volume.

A third throughline, which hardens over the year, is matching the tool to the job rather than treating AI as one monolithic thing. Custom Instructions (May) get an explicit decision grid — use them only when a task is both high-frequency and low-volatility, otherwise you get tone bleed into unrelated chats. The Excel/PowerPoint packs (September–October) show Claude and Copilot have different strengths (Claude for complex multi-tab logic and prompt adherence; Copilot for live point-edits) and that prompts must be written around the finished deliverable (what board-ready artifact, for whom) rather than the generic task (basic formula help). By December, this culminates in an explicit framework: decompose any workflow into five cognitive-task types — extraction, synthesis, analysis, generation, validation — and route each to whichever model tests best for that type, since asking which single model to use for an entire workflow is the wrong question; multi-model routing costs more per month but eliminates the cleanup hours that make single-model output worthless.

A fourth throughline is fighting the model's trained agreeableness. The too-helpful-to-think post (June) names sycophancy directly — RLHF teaches models that having conviction is misaligned — and supplies eight adversarial personas (Harsh Academic Reviewer, Devil's Advocate, Pre-Mortem Analyst, Red Team Commander) to force real pushback. The financial-decision pack (July) and the negotiation pack both operationalize the same insight: use AI as an analyzer that maps trade-offs, never as a decider that hands you an answer, because trusting a model's first note of agreement on a decision that matters is worse than getting no answer at all.

By late 2025 the packs get more ambitious in scope but not different in kind: the Nano Banana Pro posts (November-December) extend the same structure-beats-vibes logic to images — a plain-English-to-JSON translator prompt so non-coders get reproducible, scoped-editable visual output — while the shopping and focus-management posts (November-December) show the same one-request-hides-several-jobs decomposition principle applied to everyday life, not just enterprise workflows. The consistent lesson Nate keeps re-teaching: specificity, phased sequencing, and self-validation aren't optional polish — they're the entire difference between a prompt that works once and a system you can run for months.

Key moves

The posts

Related