Say the thing you want

A lesson from Working With AI. Play it above, or read it through below.

• Say what you want, not who to be

  The common read: A detailed expert persona focuses the model on the right knowledge, so opening with twenty years of experience is worth the tokens it costs.

  The persona was never the lever. Anthropic's current reference gives role prompting two sentences and one example, and the example is the entire system prompt: "You are a helpful coding assistant specializing in Python." Nothing resembling twenty years of experience appears anywhere this course read in the current guidance, and the blog that covers the technique warns against over-constraining the role because it can limit the model's helpfulness rather than focus it.

  What replaced it is stating the outcome. The docs' golden rule is to show your prompt to a colleague with minimal context and ask them to follow it, on the argument that if they would be confused the model will be too. Their worked example is not subtle: "Create an analytics dashboard" against "Create an analytics dashboard. Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation."

  That arrives as a review thread where two engineers argue about which persona works better, on a prompt whose actual defect is that neither version says what the output should contain. Delete the costume and write the acceptance criteria. The specific ask is the part that generalises, and the guidance says the quiet part directly: if you want above-and-beyond behaviour, request it explicitly rather than relying on the model to infer it from a vague prompt.

• Emphasis is not a lever. A reason is.

  The common read: An instruction in capitals with MUST in it is how a rule sticks, so the strongest wording is the safest wording.

  Volume was compensating for something that got fixed. The reference says Claude Opus 4.5 and 4.6 are more responsive to the system prompt than previous ones, that prompts written to stop a model underusing a tool may now make it overuse one, and that the repair is to write Use this tool when... where you used to write CRITICAL: You MUST use this tool when.... The shouting is not neutral. It is a live instruction to weight that rule harder, and it is being obeyed.

  The thing that does still buy compliance is a reason, and the docs' own contrast is the one to keep. Against NEVER use ellipses they put: "Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them." Their stated payoff is generalisation, because a rule whose reason is stated extends to cases nobody enumerated, and no amount of capitalisation does that.

  That arrives as a bug filed against a model upgrade, "it has started calling the database for things it already knows", on a team whose config changed by zero lines because the offending line was written in 2024. Re-read your loudest rules first. Every all-caps clause in your instructions is a bet on a model that no longer exists, and the ones you shouted at hardest are the ones now costing the most.

• Your formatting is an instruction you did not mean to give

  The common read: The model reads my prompt for its content, so how I lay the prompt out is a matter of my own convenience.

  Name what a markdown-heavy prompt does to the reply, then read on.

  It produces a markdown-heavy reply. The docs state it as a steering lever rather than a curiosity: the formatting style used in your prompt may influence the response style, and removing markdown from your prompt can reduce the volume of markdown in the output. Two other defaults sit in the same place: the current generation defaults to LaTeX for maths, and is documented as less verbose than its predecessors.

  That arrives as a complaint that the assistant "answers everything in bullet points now", filed against a model version, by someone whose own prompt template is a heading and a bullet list. Write the prompt in the register you want the answer in. It is the cheapest steering lever in this lesson, it costs nothing, and it is the only one that works by removing characters rather than adding them.

• The advice rots faster than the model

  The common read: Prefilling the assistant turn is a property of how the model continues text, so it is a technique that cannot be withdrawn from under me.

  Prefill on the last assistant turn is a 400 now. The migration note is precise about the boundary, which is worth reproducing rather than summarising: prefilled responses on the last assistant turn are no longer supported starting with Claude 4.6, requests carrying them return a 400 error, earlier models still support prefills, and assistant messages elsewhere in the conversation are not affected. One position in one array is the whole difference between a working request and a rejected one.

  The documented replacements are not "prompt harder". For a format, structured outputs constrains the response to a schema. For a preamble, a direct system-prompt instruction plus post-processing. For a continuation, move the interrupted text into a user message and ask the model to continue from it. Every one is a shape change rather than a better sentence: a removed capability is replaced by a different mechanism, not by firmer wording.

  The version of this that never reaches a tracker is the prompt file nobody re-read: it still opens with a persona, still shouts one rule in capitals, and still carries a prefill that has been quietly failing since the last upgrade. Date your prompt techniques. Every one of them was true when someone learned it, the ones that stopped being true rarely announce it, and the one in this lesson announced it with a 400.

• Say the thing you want, not the thing you fear

  The common read: A prohibition is the tightest way to express a constraint, so writing down what must not happen is the safest instruction I can give.

  Predict what a "do not use markdown" instruction puts into the input, then read on.

  The word markdown. A prohibition puts the forbidden thing in the input. The docs' first steering lever is to tell the model what to do instead of what not to do, and their example is small enough to hold: against "Do not use markdown in your response" they put "Your response should be composed of smoothly flowing prose paragraphs." One of those sentences describes an outcome and the other one names the thing you are trying to avoid.

  Examples have the same double edge and the docs are explicit about it. They recommend three to five, wrapped in tags, and they name the failure directly: examples must be diverse enough that the model does not pick up unintended patterns. Five examples that all share an accidental shape teach that shape. Anthropic's 2026 write-up goes further and reports that on its newest models examples can constrain exploration rather than guide it, and recommends expressive tool interfaces instead.

  That arrives as an agent that solves every task the same way, filed as a creativity problem, on a prompt whose five examples happen to share a structure nobody noticed writing. State the outcome and vary the examples. Both halves are the same discipline: what you put in the input is what the model has to work from, and a prohibition and a monotonous example set are both ways of putting the wrong thing there.

• The prompt that survived deletion

  The common read: A longer prompt covers more cases, so the safest thing to do with a prompt that is not working is to add to it.

  Guess how much of its own system prompt Anthropic deleted, then read on.

  Over eighty percent. The 2026 write-up on its newest generation reports removing over eighty percent of Claude Code's system prompt with no measurable loss on their coding evaluations, and gives the diagnosis in the same paragraph: they had been overconstraining the model, through the system prompt and through instruction files and skills. That is a vendor reporting on its own product, and it is the strongest available evidence for a claim nobody wants to believe.

  The same write-up names the mechanism that makes long prompts worse rather than merely wasteful. Reading their own transcripts, they found conflicting messages inside a single request, with the system prompt, the skills and the user's request clashing: The version of this that never gets filed is a prompt that has grown for a year, that everyone believes is load-bearing, and that nobody has ever tested by removing a paragraph. Cut before you add. Call the part that comes out the theatre:

  The thank-you at the bottom of the long version is Loop's. Loop has been polite to the model for a year. Loop has never once checked whether the politeness changed a single character of the reply. Nobody has, and Loop is keeping it, and this course is not going to argue.