Prompt engineering best practices are becoming obsolete

Executive overview

Most prompt engineering advice teaches static tactics — roles, steps, chain-of-thought — that are already losing relevance as models evolve. Reasoning models handle chain-of-thought internally; agentic tools generate their own steps. Teaching techniques tied to today's models creates knowledge that expires fast.

Three principles remain useful regardless of model generation: be specific with word choice, provide rich context, and describe what you want rather than prescribing how to achieve it.

The shift from prescriptive to descriptive prompting is the single most durable change to make now.

Viewer Q&A: deep research and PDF filtering

• You cannot set hard parameters (e.g. PDFs only) in deep research tools today.
• You can direct the model in natural language: appending "only use sources ending in .pdf" to a prompt yields ~80–85% PDF sources in practice.
• Tested in Perplexity; the same technique applies to OpenAI and other tools.

Viewer Q&A: connecting AI-generated code chunks

• Chunks are not separate outputs — they come from the same model in a single session.
• The process: Spec (what to build) → Blueprint (how to build it, broken into iterative chunks) → To-do list (implementation order).
• Each chunk adds more detail than the last; the model retains full context across all chunks.
• The to-do list acts as a durable reference the model can return to during builds.

Why current prompting tactics are expiring

• Chain-of-thought prompting was a workaround to make models reason step-by-step. Reasoning models (O1, O3, Claude 3.7, DeepSeek R1) do this natively — the tactic is redundant.
• Inference cost for reasoning models is falling; speed and quality will improve, making reasoning the default mode.
• Explicit steps in prompts are becoming unnecessary with agentic deep research tools — the agent generates, then iteratively expands its own steps based on what it learns mid-task.
• Both OpenAI and Anthropic prompt generators already embed these tactics by default, showing they are institutionalised — and therefore ripe for obsolescence.

Three evergreen prompting principles

1. Be specific with words. Words carry multiple meanings; vague instructions produce vague outputs. "Extremely detailed summary" is self-contradicting; "brief summary" is redundant. Choose words that precisely constrain the model's direction.

2. Provide context. More context yields better results today. Over time, models will pull context from ambient sources (your history, connected tools) rather than requiring you to supply it directly — but for now, more is better.

3. Describe, don't prescribe. Prescriptive prompts tell the model how to do something; descriptive prompts tell it what you want. AI will increasingly know better than you how to execute — your job is to articulate the outcome clearly and let the model determine the method.