🧠 7 Prompting Strategies from the Anthropic Prompt Engineers

Kia ora, Namaskaram 🙏🏾

Does your prompting still feel hit-or-miss?

When you're working with ChatGPT, Claude, or any advanced language model, your results completely depends on how well you prompt.

Prompting isn’t guesswork — it’s a skill you can sharpen.

These 7 rules from the Anthropic team’s podcast will help you craft better, more trustworthy prompts.

1. Write your shitty first prompt (and then iterate)

Prompting is an experimental craft. Get something down, then test and improve it.

🧠 Insight from the Anthropic team:
Zack Witten (Prompt Engineer, Anthropic) and Amanda Askell (Head of Fine-tuning Team, Anthropic) both rely on iteration — write hundreds of prompts to discover what works for you.

💻 Example prompt:

Summarise this research paper on electric vehicle adoption for a general audience [insert paper].

Prepare a first draft — we’ll refine the summary together.

2. Make your goals explicit

If your goal isn’t clear to the model, your output won’t be either.

🧠 Insight from the Anthropic team:
Zack Witten (Prompt Engineer) and Amanda Askell (Head of Fine-tuning Team) emphasise that prompting is a form of clear communication. State your task, constraints, and expected output format.

💻 Example prompt:

You are a behavioural scientist. Suggest low-cost interventions to increase electric vehicle adoption in New Zealand. Focus only on behavioural levers, not technology or infrastructure changes. 

Provide 3–5 ideas with a brief rationale using behavioural science principles.

3. Context — don’t assume the model knows your world

The model doesn’t share your knowledge, team norms, or audience.

🧠 Insight from the Anthropic team:
Alex (Head of Anthropic Developer Relations) compares prompting to briefing a capable but uninformed temp: you must supply the context. Amanda Askell (Head of Finetuning Team) says removing assumptions is one of the hardest — and most important — parts of prompting.

💻 Example prompt:

In New Zealand, electric vehicle uptake is uneven—especially outside major cities. Key barriers include perceived high costs, range anxiety, limited charging access in rural areas, and a strong preference for large vehicles like utes, particularly in farming and trades. 

Outline the most relevant behavioural barriers and drivers in this context.

4. Ask the model to think step-by-step

Reasoning improves quality, accuracy, and transparency.

🧠 Insight from the Anthropic team:
Zack Witten (Prompt Engineer) recommends this as a go-to instruction for complex tasks. Amanda Askell (Head of Fine-tuning Team) finds that structuring the model’s reasoning often improves performance and clarity.

💻 Example prompt:

Identify the three best low-cost behavioural interventions to increase EV adoption in New Zealand.

Think step-by-step:
1. First, outline key behavioural barriers and drivers. 

2. Then, use a simple framework (e.g. impact × scale) to prioritise and select the top three interventions. 

3. Briefly explain your reasoning.

5. Evaluate — don’t just assume it got it right

Read carefully. Check edge cases. Be a critical reviewer.

🧠 Insight from the Anthropic team:
Amanda Askell (Head of Fine-tuning Team) encourages stress-testing prompts. Zack Witten (Prompt Engineer) notes that good-sounding answers can hide subtle but serious errors.

💻 Example prompt:

Critically evaluate each proposed intervention to increase EV adoption in New Zealand. Score for impact, feasibility, and context fit (1–5). Identify what could go wrong, who it might miss, and suggest improvements.

6. The mythical better prompt — don’t chase perfection

There’s no single “perfect” prompt. Stop over-polishing.

🧠 Insight from the Anthropic team:
David Hershey (Technical Lead, Customer Work at Anthropic) sees many users stuck in search of the “mythical better prompt.”

💻 Example prompt:

Based on the behavioural goal of increasing electric vehicle adoption in New Zealand, suggest three ways to test behavioural interventions. Focus on rapid, low-cost experiments that can be run in two weeks and generate useful learnings.

7. When in doubt, ask the model to help you prompt

The model can help you write better prompts — just ask.

🧠 Insight from the Anthropic team:
Amanda Askell (Head of Fine-tuning Team) often asks Claude to revise unclear prompts. Alex (Head of Developer Relations) treats Claude as a “prompting assistant” — great at diagnosing ambiguity and offering better phrasing.

💻 Example prompt:

Help me write a clear prompt that gets the model to think like a PhD-level behavioural scientist. The goal is to generate low-cost, high-impact interventions to increase EV adoption in New Zealand, using behavioural models and frameworks, and grounding ideas in the local context.

⚡ Last spot left: Join the AI Playbook for Behaviour Change

✅ Turn AI into your personal PhD-level Behavioural Designer
✅ Get the ChatGPT Behaviour Change Playbook
✅ Run rapid behaviour change experiments with AI

Final call to join my upcoming 2-week course.

P.S. Only 1 of 20 places left — get in quick!