Talk to AI Like an Engineer

Core concept 01

Treat the model like a brilliant new hire, on day one.

Anthropic's own guidance for getting better output starts with a blunt instruction: tell the model exactly what you want to see. If you want a comprehensive answer, ask for a comprehensive one. Most disappointing replies are not the model being bad at its job, they are a briefing problem. A new hire on day one cannot read your mind either, and nobody expects them to.

The fix is the same one that works on a real new hire: be clear, give context, and show what good looks like. Explaining why something matters, not just what you want, helps the model aim at your actual goal instead of a generic one. Think of it less like typing a search box and more like handing someone at the party a clear job before you walk away to refill your drink.

Source: Anthropic, "Prompt engineering best practices for 2026"

Core concept 02

Every prompt that works has the same five layers.

Strip away the hype and Anthropic and OpenAI describe the same shape from two angles. Anthropic's prompting guide recommends separating a prompt's distinct parts, instructions, context, and examples, so the model never has to guess which is which. OpenAI's developer guide lays out nearly the same build order: identity first, who the model should be, then instructions, then examples, then context. Give a prompt a role, the context it needs, a specific task, and the format you want back, and you have covered what both labs actually ask engineers to do.

Of those five, one habit does more work than the rest: giving examples. Anthropic puts it plainly, examples show rather than tell, and they clarify the subtle stuff that is hard to put into words, like tone or a particular structure. A single well-chosen example of the output you want is often worth three paragraphs of description. That is why the line that does the most work in any prompt is usually some version of "here is one I loved, match this."

Sources: Anthropic, "Prompting best practices" · OpenAI, "Prompt engineering"

The five layers of a prompt that works · role, context, task, format, examples

Core concept 03

In 2025 the labs renamed the skill: context engineering.

Here is the shift that separates someone who read one article in 2024 from someone keeping current. In September 2025 Anthropic published a piece calling context engineering the natural progression of prompt engineering. Prompt engineering is writing the instructions well. Context engineering is managing the whole set of information the model can see at once, the system instructions, the tools, the files, the conversation so far, so that the highest signal is present and the noise is not.

For you that is not a harder job, it is a calmer one. The context window is a fixed budget, and everything competes for it: your standing brief, every document you pasted, the entire back-and-forth, all of it. A long meandering thread does not make the model smarter, it crowds the desk. The move most people never make is the highest-leverage one: start fresh, more often than feels natural, with only the few files that matter. A clean window with the right three attachments beats a forty-message thread almost every time.

Go deeper: why the same prompt gets worse deep in a long chat

As a conversation grows, older turns get summarized or dropped to fit the budget, and the signal you carefully set up at the top gets diluted by everything since. The instruction is technically still "in there," but it is competing with far more text than it was on message one. This is the same mechanism behind Lesson 01's forgetful long chat, seen from the writer's side instead of the reader's.

The operator habit: treat a conversation as disposable scaffolding, not a filing cabinet. When the output starts drifting, do not argue with it across ten more messages. Open a new chat, paste the standing brief and the two or three real files, and start clean.

Source: Anthropic, "Effective context engineering for AI agents"

Scroll the diagram sideways →

The context budget · same window, two ways to spend it

Core concept 04

Iteration is the skill, not a sign you got it wrong.

The best prompting pattern either lab describes is not a clever one-shot, it is a loop: draft, review against your own criteria, refine. Anthropic calls this self-correction and names it the most common prompt-chaining pattern, one step generates a draft, the next reviews it, the next fixes it. You can run that same loop out loud in a normal chat: ask the model to critique its own draft against what you actually needed, then revise from the critique.

OpenAI's guidance points the same direction from the production side: pin what is working, test changes, and iterate before you trust a prompt with real work. Correcting a draft out loud, in plain words, teaches the model your standard faster than rewriting the whole prompt from scratch. Once a prompt earns its keep two or three times in a row, it stops being a one-off message and becomes a template worth saving, which is exactly what the "make it yours" build below produces.

Sources: Anthropic, "Prompting best practices" · OpenAI, "Prompt engineering"

Core concept 05

Read the bad answer. It tells you which layer was thin.

A disappointing reply is not a dead end, it is a readout. Each way an answer misses points back to one of the five layers, and the fix is a one-line addition, not a rewrite from scratch. This is the same diagnostic habit as Lesson 01's fault map, aimed at the brief instead of the agent.

The prompt triage · symptom, layer, one-line fix
The answer was	The thin layer	The line to add
Generic. True of anyone in your field, specific to no one.	02 · Context	Paste a real example of your own: your numbers, your last good one, your actual words.
The right topic, but not the thing you needed made.	03 · Task	Name the deliverable exactly: "the gallery delivery email," not "help with emails."
Right content, wrong length, tone, or shape.	04 · Format	Add the constraints: "120 words, warm, one link, no exclamation points."
Close on the words, off on the voice. Not how you sound.	05 · Examples	Show one you loved and say "match this." Voice is caught from samples, not described.
It keeps forgetting a rule you set early in a long chat.	01 · Role, and the window	Move the rule into the standing instructions, and start a fresh, shorter conversation.

Notice the pattern: four of the five fixes are things you add, and none of them is "try harder." Better output is a briefing skill, and it is almost entirely additive, which is why it is learnable in an afternoon.

The brief you leave with

The five-layer brief, as a fill-in-the-blank

Write this once for the task that eats the most of your week and you never start from a blank box again. Paste it as your standing instructions, then just add today's context on top.

01Rolewho the model is: "You are a [your role]'s [assistant / studio manager / editor]."
02Contextwhat it needs to know about your world, and the 2–3 real files you will attach
03Taskthe exact deliverable, named: "draft the [thing], ready to [send / post / use]"
04Formatlength, tone, and hard rules: "[N] words, [tone], [one link / no emoji / your constraint]"
05Exampleone past piece you were proud of, pasted in full, with "match this" — the single highest-leverage line in the brief

Fresh from the lab

What changed this month

Anthropic just showed its own team deleting prompts instead of piling more on. On July 24, the Claude Code team published "The new rules of context engineering for Claude 5 generation models," saying it cut more than 80% of Claude Code's system prompt when it moved to Claude Opus 5 and Claude Fable 5, with no drop in coding benchmark scores. The extra rules had been written to guardrail an older, less careful model, and the newer one simply didn't need them anymore. Anthropic also shipped a new command, claude doctor, that scans your own system prompt, skills, and CLAUDE.md files and flags what's safe to cut.

OpenAI's new prompting guide has one instruction: stop explaining how, just say what. The official guidance for the GPT-5.6 family, published July 9, calls it outcome-first prompting: state the goal, the constraints, and when to stop, then let the model choose its own path instead of walking it through each step. In OpenAI's own coding-agent tests, trimming prompts down to that shape improved eval scores by 10 to 15%, cut token use by 41 to 66%, and cut cost by 33 to 67%. "Think step by step," the line that used to genuinely help, is now something GPT-5.6 already does on its own, so leaving it in just spends tokens on an instruction the model no longer needs.

Sources: Anthropic, "The new rules of context engineering for Claude 5 generation models", July 24, 2026 · TechTimes, "GPT-5.6 Prompting Guide: Lean System Prompts Now Outperform Elaborate Scaffolding", July 15, 2026

Vocabulary

Six words that make prompting click

These build directly on lesson one's ten. Learn these and you can read any AI settings menu, from Claude to ChatGPT, without translating.

Context engineering

Managing everything the model can see at once, not just the sentence you type. Anthropic named it the natural progression of prompt engineering in 2025. In practice it is one habit: keep the window full of signal and start fresh before it fills with noise.

System prompt

The standing instructions a model reads before every message in a conversation, the role and rules layer. In Claude Projects and custom GPTs, this is the "instructions" box you fill in once.

Few-shot examples

One or more samples of the exact output you want, pasted right into the prompt. The model matches the pattern instead of guessing at your taste.

Custom instructions

ChatGPT's version of a standing brief, background about you and your preferences applied to every new chat, up to 5,000 characters on paid plans as of July 2026.

Persona

The role you assign the model: "you are a wedding photographer's studio manager." A clear persona narrows the model's choices before it writes a single word.

Iteration

Correcting a draft out loud and running it again, the fastest way to teach a model your standard. Every good template started as an iterated draft.

Deeper definitions, straight from the builders: Anthropic prompt engineering docs · OpenAI prompt engineering guide

Make it yours

Build your prompt kit

Same five layers, your exact pain. Pick your lane and build three reusable templates aimed at the writing or thinking task eating the most of your week. Budget about an hour, once. A free Claude or ChatGPT account is enough to start.

For female founders

I'm the CEO, the marketer, the bookkeeper, and the intern, and I write every proposal, post, and follow-up from a blank page because I've never saved a version of my own voice.