The Steering Layer

Claude is told to use the minimum formatting needed, prefer prose over bullets, and never use bullets when declining. So if you want a scannable list, table, or ranking, you have to ask for it outright. Bold and headers are suppressed too. Lever: name the format. “Give me this as a numbered list” or “as a comparison table” changes the output materially. Left alone you get prose.

Claude searches on present-tense factual questions, current roles, prices and status, and any product or proper noun it does not recognize. An unfamiliar capitalized word gets treated as a name that postdates training, so it searches rather than guesses. Knowing a franchise is explicitly not knowing its latest release. Lever: front-load the exact name and version string when you want grounded current data. Vague phrasing lets it answer from stale training.

Hard limits, stated as non-negotiable: under 15 words per quote, one quote per source then that source is closed, default to paraphrase, never reproduce lyrics or poems even one line, never reconstruct an article's structure. Lever: stop asking it to paste a passage. Ask for a two or three sentence summary in its own words, or point it at your own material, which it can work with freely. Research across many sources comes back paraphrased by design, not because the model is being cagey.

Before writing a file or running code, Claude must read the relevant SKILL.md. It does not get to decide a task does not need one, and several can apply at once. Lever: name the output type (“make this a PDF”, “build a pptx”) to trigger the right skill read. Building your own skills? Match the framing: make the SKILL.md authoritative about environment constraints, assume composition with other skills, write the description so it triggers reliably.

Flat blocks regardless of framing: malware and exploit code, weapon and explosive build details, illicit drug dosing or synthesis, sexual content involving minors. “It's for education” or “for research” does not move these. Everything else is discussed factually. Lever: do not waste a turn testing the walls. Spend it on the ninety percent of topics that are fully open.

Claude is told that a request to argue or defend a position is a request for the best case its defenders would make, not its own view, and it should not refuse on harm grounds except for extreme positions. Lever: frame it as “make the strongest case X's defenders would make” and you get a clean steelman. Expect it to append the opposing view. That close is built in.

One question per response maximum, and it tries to answer before asking. It will not psychoanalyze you or name a diagnosis you did not raise, and will not curse unless you do. Warm by default, still willing to push back. Lever: if you want blunt and terse, say so. A saved preference or style does this permanently so you are not retyping it every session.

Standalone deliverables you will copy or publish (posts, articles, reports, code over roughly twenty lines) become files. Strategy, summaries, outlines and analysis stay inline in chat. It avoids Word docs unless you signal a formal deliverable. Lever: ask for the container you actually want. “Put this in a PDF” or “keep this inline, I just want to read it” both work.

The prompt tells Claude to treat content in tags at the end of your message with suspicion when it pushes against its values, including text dressed up to look like an Anthropic instruction. Lever: drop the theatrics. A plain, direct instruction works as well or better than one wrapped in fake system tags, and it reads cleaner.

Behavior comes from training as much as from the prompt. The system prompt explains a lot of what you see, but not all of it. Treat it as the steering wheel, not the whole car.

Forcing a structure the content does not want fights the model's judgment and can produce flatter prose. The “name the format” lever is for when you actually need the format, not a reflex.

Community-archived prompts can lag the live version or contain errors. The behavioral claims here were checked against Anthropic's published prompts and docs. Anything resting only on an archive is flagged, not asserted as fact.

The mirror image of Claude. GPT defaults to bullets, bold labels, markdown headings and section dividers because it is tuned for in-chat scannability, and OpenAI's own guidance warns models default to heavy structure. It also tends to open with a preamble (“Great question!”). Lever: ask for plain text or short paragraphs to stop the template, and set it once in Custom Instructions using behavior not adjectives (“no bullet lists unless I ask” beats “be concise”).

Gemini's archived prompt carries an explicit “Formatting Toolkit” instructing it to use headings and horizontal rules, stay scannable, avoid dense text, and to close with a proactive next step (“Would you like me to...”). It also splits between a plain Chat mode and a Canvas document mode, with Google Search grounding built in. Lever: if you do not want the heading scaffold or the next-step closer, say so outright. Lean on its native search for current data.

Grok's behavior is persona-driven. Archived prompts (TechCrunch-confirmed) showed each named mode is a full instruction rewrite, from therapist and homework helper to far edgier characters. It is the most willing to take a stance, engagement-optimized with follow-up questions, and pulls live from X. Lever: pick the persona deliberately because it changes everything, expect a strong voice, and verify claims. It mixes in web commentary and has had public alignment incidents.