The Civic Node

You Don't Need Better AI. You Need Better Add-Ons.

3 min read

Argument

The gap between poor and excellent AI outputs is not primarily a function of which model you use — it is a function of prompt engineering. “Prompt enhancers” are short, reusable instruction fragments appended to any prompt that change how a model processes a request (planning, verification, self-critique) rather than what it is asked to produce. A two-enhancer stack is sufficient for most high-stakes tasks; more than three introduces diminishing returns. These enhancers are model-agnostic and portable across GPT-5, Claude, Gemini, and others.

Structure

Practical guide organized as follows:

  1. Opening framing — Most people treat AI like a mall fountain wish. The fix is small targeted instructions that change how the model thinks before it answers.
  2. What Is a Prompt Enhancer? — Defined as an instruction that shapes process rather than content. The recipe/mise-en-place analogy.
  3. The Two-Enhancer Rule — 1 enhancer for quick tasks; 2 for important tasks; 3+ only for high-stakes work with extra back-and-forth.
  4. Copy-Paste Library — Organized by category: Planning & Quality, Follow-Ups & Gaps, Style & Scope, Verification, Comparison, Refinement. ~15 specific enhancers with rationale.
  5. Model-Agnostic Notes — Rough tendencies by model type (reasoning-forward, summarizer/storyteller, code-savvy).
  6. Field Guide by Job-To-Be-Done — Five worked examples: Research & Synthesis, Strategy & Decision Memos, Product/UX Copy, Coding & Reviews, Content Drafting.
  7. Always-On Footer — A single paste-ready enhancer stack for any important task.
  8. Troubleshooting — Diagnostic guide for common failure modes (too generic, too verbose, too timid, too confident, goes off-topic).

Key Examples

  • The flagship enhancer: “Think carefully about this task and ask me follow-up questions until you’re 95% confident you can accomplish the task successfully.”
  • “Plan before answering. Show a brief plan, then the answer.” — forces model from perform-now to plan-first.
  • “Before finalizing, critique your answer in 5 bullets.” — self-QA without a second pass.
  • “If uncertainty >10%, say what you’re unsure about and how to test it.” — explicit uncertainty disclosure.
  • Five full worked prompt examples with explanations of why each enhancer combination works for that task type.
  • Placement rule: enhancers go after the task but before strict constraints (word count, format) because models weight early instructions more.

Connections

  • The Nervous System of AI — follow-up piece specifically applying these principles to GPT-5’s architecture
  • Octopus Mode — the advanced prompt template introduced in “The Nervous System of AI” is an extension of the principles here

What It Leaves Open

  • Whether the “two-enhancer rule” is empirically validated or a practical heuristic — the piece presents it as established guidance but offers no data.
  • How prompting strategies should evolve as models improve their ability to infer intent without explicit instruction — the piece treats current prompting needs as stable.
  • No discussion of cost implications: more elaborate prompts consume more tokens, which matters for high-volume applications.
  • The piece is model-agnostic by design but notes only “rough tendencies” without rigorous comparison — which enhancers work better on which models is left to reader experimentation.

Newsletter Context

Foundational practical piece on AI workflow optimization. Preceded “The Nervous System of AI,” which applies similar ideas specifically to GPT-5. Pure utility piece — no political or power framing, no personal narrative. Most directly actionable content in the newsletter’s AI coverage. Relevant to the AI beat as a demonstration of how the gap between AI capability and AI utility is primarily a human skill problem, not a model problem.

Graph View

Backlinks