Designing prompts that save you hours every week with Dropbox's Morgan Brown

Executive overview

Knowledge workers waste hours on repetitive tasks they barely notice doing. Morgan Brown, VP of Product for AI at Dropbox, built a system of AI workflows — prompt libraries, daily briefings, email agents — to reclaim that time.

The core lever is context: give an LLM your operating principles, your role, and your constraints before asking anything, and it becomes a genuine thought partner rather than a search engine.

The real leverage of AI isn't better answers — it's asking better questions.

Eliminating grunt work with AI

• Meeting transcripts auto-converted to weekly summaries: commitments made, commitments owed, key decisions
• A 12-question document review prompt pre-reads strategy docs and flags gaps before they reach Morgan
• Shared the prompt with his team — they now self-review before sending work up
• A daily 5am agent scans X, Substack, arXiv, YouTube, and podcasts; categorises AI news as world-changing / interesting / incremental / noise; surfaces the top signals in context of his Dash role
• A Claude Code agent reads his personal Gmail daily: categorises, summarises newsletters, flags priority emails, drafts responses

How to write prompts that actually work

• Structure every prompt: context → task → output format → detailed instructions
• Front-load the most relevant context — LLMs read everything in light of what came first
• Assign a domain expert persona to anchor the response (e.g. "You are April Dunford, help me position this feature")
• Specify the output style explicitly: "no fluff, no hyperbole, tell me if the idea is bad"
• Use meta-prompting: talk through your need in ChatGPT first, then ask it to structure a formal prompt; refine iteratively like editing copy

Using AI as a thought partner

• Give the LLM your operating manual first: your priorities, communication style, what adds friction
• Build domain context docs for recurring projects (e.g. CarbScan priorities: accuracy, no false confidence, speed)
• Use divergent thinking prompts first — ask it to lead you through questions, not give answers
• Switch models deliberately: run a session in ChatGPT, then feed the output to Claude and ask what it thinks
• Ask it to role-play stakeholders: how would your CTO, legal team, or go-to-market team react to this proposal?

AI-accelerated experimentation

• Pre-test ideas inside an LLM before involving humans: generate 100 subject line variants, filter to 5 strong ones in 30 minutes
• Simulate target personas with detailed demographic, psychographic, and behavioural context before running copy against them
• Always ask the LLM to cite its reasoning — forces it to distinguish grounded claims from hallucination
• Use LLM simulation as pre-work to make limited human-testing time higher-signal
• Bring real humans in as early as possible: team, customer advisory boards, design partners — LLMs sharpen the question, humans validate the answer
• The weakest link is hypothesis quality: better framing upfront, not more output volume

One piece of homework

• Take 30 minutes and identify one thing you do constantly, almost without thinking
• Break it into steps and find the one slice AI can take off your plate
• Use that small win as the starting point to automate the rest