Anthropic Prompt Tips: How to Write Better AI Prompts

Why clear thinking matters for AI prompting

Philosophy trains you to clarify concepts, expose hidden assumptions, and communicate with precision — and those are exactly the muscles you need for better prompts. Amanda Askell, a philosopher on Anthropic’s technical team, often frames prompt work as the same craft: balance experimentation with explicitness. The truth is, a clear head + a clear prompt = fewer surprises from the model.

What Amanda Askell recommends: experiment, explain, iterate

On Anthropic’s Ask Me Anything podcast, Askell walked through a practical routine: interact with the model often and inspect output after output. That iterative loop is where you learn what the model actually follows and what it quietly ignores. From my experience working with product teams, this is the fastest route to reliable answers.

• Be experimental: Try different phrasings, toggle constraints, and compare answers side-by-side — you’ll spot patterns fast.
• Be explicit: Tell the model your expectations — tone, length, style, and what to avoid. Don’t assume shared context.
• Iterate fast: Use each response as feedback and refine the prompt. Two quick passes often beat one careful guess.

How philosophical training improves prompts

Askell argues that philosophy forces you to state your concerns clearly. Instead of a vague "Improve this paragraph," say: "Make this paragraph concise, replace passive voice with active voice, and maintain a neutral, business-professional tone." That sort of specificity reduces ambiguity and helps the model return a useful reply. It’s like giving a colleague a checklist rather than a hint.

Anthropic’s practical analogy: Claude as a new employee with amnesia

Anthropic’s Prompt Engineering Overview gives a mental model I still find useful: imagine Claude as a brilliant but new employee who doesn’t know your conventions. In practice, that means the model won’t inherit your style guide or assumptions unless you provide them. So — give the context block. Give examples. Be explicit about norms.

Key prompting techniques you can apply today

• Set a context block: Open with goals, audience, constraints, and one or two examples of preferred output. That upfront context prevents many misunderstandings.
• Use explicit formatting instructions: Need bullets, numbered steps, or HTML? Say so. The model follows format instructions well when they’re clear.
• Ask the model to self-critique: After an answer, request a short critique and a revision — the self-critique often surfaces subtle problems.
• Chain-of-thought sparingly: For tricky reasoning, ask for step-by-step reasoning, then a concise summary. (But don’t overuse it — it’s costlier and can leak tentative paths you don’t need.)
• Provide counterexamples: Show what you don’t want. Negative examples are surprisingly instructive for reducing unwanted outputs.

What business leaders and investors say

Marc Andreessen and other leaders pitch a useful shift: treat AI as a "thought partner." The skill isn’t just formatting — it’s asking the right questions. Teams that can frame problems precisely are finding tangible value; some companies now hire dedicated prompt engineers. Not glamorous, but effective — and yes, marketable.

Short practical prompt templates

Here are three tight templates inspired by Askell’s emphasis on explicitness. Swap the bracketed text for your specifics and you’re off to the races.

• Editing / Style: "Edit the following text to be [target word count] words, use [tone], eliminate passive voice, and preserve technical accuracy. Output as bullet points: strengths, weaknesses, suggested rewrite."
• Research Summary: "Summarize the research on [topic] in 5 bullet points for a nontechnical executive. Include 2 citations and one recommended next step."
• Creative brief: "Draft a 150-word product description for [audience] that highlights [key features], uses a friendly tone, and ends with a one-line CTA."

Mini case study — improving a marketing brief

Here’s a quick, real-feeling example. A marketing manager asked Claude: "Make this tagline snappier." The first pass gave generic options. Following Askell’s advice, they refined the prompt: "Create five taglines under 6 words for a premium noise-cancelling headphone aimed at commuters aged 25–45. Emphasize comfort and battery life, avoid hyperbole, and give a 1-sentence rationale for each." The second pass produced focused lines and rationales the team could A/B test. Lesson: specificity + iteration = usable marketing copy in minutes.

Common prompting mistakes to avoid

• Vague goals (e.g., "Make this better").
• Too many tasks in one prompt — split complex workflows into steps.
• Assuming shared context — always provide necessary background.
• Over-relying on one phrasing — vary prompts to test robustness.

Recommended resources and further reading

For a deeper practical dive, see Anthropic’s prompting guide and these industry write-ups on prompt engineering:

• Business Insider — How to write better AI prompts
• Business Insider — What prompt engineering jobs look like

Takeaways: how to prompt like a philosopher

• Clarify your goal before you write the prompt.
• Be explicit: supply background, constraints, and example outputs.
• Experiment quickly and iterate using the model’s responses as feedback.
• Teach the model: explain norms, style, and what to avoid — it won’t assume them.

Effective prompting blends curiosity, discipline, and clear communication — the very habits philosophy builds. Try Amanda Askell’s method: write a precise prompt, run two variants, and compare results. You’ll learn more in ten minutes than you expect. (Honestly — it’s a small habit that pays.)

Sources: Anthropic’s Prompt Engineering Overview; Business Insider articles on prompting and prompt engineering.

For a related primer on reliable prompt design and iterative testing, see our guide on multi-agent AI coding, which covers practical iteration patterns and debugging workflows that apply to prompt engineering as well. Learn more about prompt iteration patterns and tools in our Cursor 2.0 Composer guide.