The insight

Most prompt tutorials teach tactics. Specific moves for specific problems. You collect them, paste them somewhere, and hope you remember which ones worked.

The Anthropic interactive tutorial teaches mechanics. Not the moves. Why they work. When the output is wrong, mechanics tell you which lever to pull. Tactics just make you start over.

Every prompt is built from five variables:

1. Role
2. Task
3. Format
4. Constraints
5. Examples

The five variables

Role: Who Agent is in this prompt. Not "you are an assistant." Try: "You are a conversion copywriter who specializes in B2B SaaS landing pages." Specificity narrows the output. That's the point.

Task: The single thing you're asking it to do. One prompt, one job. Two jobs means two mediocre outputs.

Format: What the output looks like. Define it explicitly: bullet list, markdown table, JSON, three paragraphs. Skip this and you'll reformat by hand instead of piping it to the next step.

Constraints: What it shouldn't do. Put the guardrails in the prompt, not the edit pass: "No filler," "max 200 words," "only include verifiable sources."

Examples: Show it what good looks like. Most people skip this one. Nothing else moves output quality faster. A concrete example beats a paragraph of description every time. The model pattern-matches to the example before it finishes reading your instructions.

Here's what all five look like applied to one prompt from the toolkit:

Role: You are a customer research specialist who builds buyer personas for B2B SaaS companies.

Task: Build one buyer persona from the market research data provided.

Format: Return a JSON object with these fields: name, job_title, company_size, primary_pain, secondary_pains (array of 3), buying_trigger, objections (array of 2), preferred_channels (array).

Constraints: Base all fields on the provided data only. No invented details. Keep primary_pain under 20 words. Buying trigger must describe a specific event, not a general state.

Examples:

{
  "name": "Operations Olivia",
  "job_title": "Head of Operations",
  "company_size": "50-200 employees",
  "primary_pain": "Manual reporting takes 6 hours every Monday",
  "secondary_pains": ["No single source of truth", "Team uses spreadsheets instead of systems", "Bottlenecks invisible until it's too late"],
  "buying_trigger": "Missed a board deadline because data lived in three places",
  "objections": ["We already have a BI tool", "My team won't adopt another platform"],
  "preferred_channels": ["LinkedIn", "industry newsletters", "peer recommendations"]
}

Notice the Format is JSON. That's not arbitrary — the next prompt in the chain receives this object as its input. That's the whole point.

Why the chain works

Knowing the five variables changes how you chain, not just how you write.

The format constraint on step one is the input shape for step two. If you define the output of your Market Research prompt as a structured JSON object, your Persona prompt can be written to receive that exact shape. No reformatting, no copying and pasting between steps.

The toolkit I built runs six steps: Market Research → Personas → Competitor Gaps → SEO → Content → AI Employees. Each prompt scoped to one output, each format designed to feed the next. I handed Claude the full chain and let it run end to end. It worked — not because any single prompt was clever, but because the handoffs were clean.

The chain is the product.

The part most people miss

Prompt sprawl is real. You end up with a folder of markdown files and no memory of which version of the competitor analysis prompt actually worked, or why you made three variants of the persona prompt.

Treat prompts like code. One job per prompt. Document which feeds which. Version what changes and why.

The AI Prompt Toolkit is 28 prompts built this way: all five variables defined for each one, the chaining workflow mapped out. It's the system, not just the prompts.

Lessons learned

• Generic prompts produce generic output. The prompt is the brief. Specific role, format, and constraints means less editing on the back end.
• Specialize then compound. Get the single version correct before chaining.
• The chain is the product. A single prompt is a move. A system of connected prompts is a workflow. The compound output beats any one.
• Version your prompts. First version is never the best. Track what changed and why. Treat iteration like code review.

Read the rest →

Worth Reading

Prompt Engineering Interactive Tutorial — Anthropic / GitHub The first-principles foundation this issue is built on. Jupyter notebooks you can run locally. Not a listicle.

How to Master Prompt Engineering — exm7777 on X Tactical patterns that pick up where the Anthropic tutorial ends.

Block CEO Jack Dorsey lays off nearly half his staff because of AI — Fortune Block cut 4,000 people — nearly half its workforce — and Dorsey said AI made them unnecessary. His prediction: "Within the next year, I believe the majority of companies will reach the same conclusion." Worth reading regardless of whether you believe the narrative. The loudest CEO yet saying the quiet part out loud.

OpenAI Raised $110B — The Largest Private Round in History — TechCrunch $730B valuation. Amazon put in $50B, Nvidia and SoftBank added $30B each. But there's a clause: $35B of Amazon's commitment only kicks in if OpenAI goes public — or builds AGI. Amazon is literally betting $35 billion on the possibility of artificial general intelligence. Meanwhile, OpenAI signed a Pentagon deal the same day Anthropic got axed — with basically the same safety red lines Anthropic got banned for having. ChatGPT is at 900M weekly active users. 50M paying.

Claude Code Gets Remote Control — Anthropic Docs Start a session in your terminal, pick it up later from your phone or the web app. It keeps working on the original machine. If you're running long tasks, you can finally walk away from your desk without killing the process.

If you're building an AI workflow for your business and the prompt system keeps breaking down before it becomes repeatable, hit reply. That's exactly the kind of problem I consult on.

— Collin