What Anthropic Engineers Build With Claude Tells You Exactly How to Brief Your Own AI
By Warren Schuitema, Founder | Matchless Marketing | The AI Dad
The people who built Claude are using it to generate HTML instead of Markdown. That one detail from a recent Lenny's Newsletter interview with Anthropic engineer Thariq Shihipar tells you more about how to use AI tools effectively than most tutorials will.
Not because HTML is the point. It isn't. The point is why they made that call, and what it reveals about how the best AI practitioners think about output formats, tool constraints, and briefing their AI to do exactly what they need.
Here's what this means for your business.
The Format You Request Is a Strategic Decision, Not a Default
Most people accept whatever format their AI tool spits out. They ask for a content plan, they get bullet points. They ask for a proposal, they get paragraphs. They treat the output format as something that just happens.
Anthropic's engineers don't think that way. When Shihipar explained why his team uses HTML output instead of Markdown in certain Claude Code workflows, the reasoning was functional: HTML renders predictably in the environments where the output actually gets used. Markdown can behave inconsistently depending on the parser. So they specify HTML to get reliable, controlled results.
The lesson for you: the format you request from your AI tool is a design decision. If you're using ChatGPT or Claude to draft client reports and you're getting output that needs heavy reformatting before it goes anywhere useful, you haven't given enough thought to the output spec. You can ask for output in a specific format, with specific structure, designed to slot directly into your workflow with minimal cleanup.
Next time you write a prompt, add a line at the bottom: "Format the output as [X], structured for [specific use case]." You'll cut your editing time significantly.
Expert Practitioners Use Constraints Intentionally
Here's what stood out to me in how Shihipar talks about his workflow. He doesn't just use Claude. He constrains it. The HTML decision is one example of a broader pattern: expert AI users impose structure so the tool's output fits their system, rather than adapting their system to whatever the tool produces by default.
This is the gap between people who find AI useful and people who find it transformative for their operations.
If you're a coach running a small practice, here's a concrete version of that same principle. Instead of asking Claude to "write me a follow-up email for a client after a session," you give it a constrained template: "Write a follow-up email using this structure: one sentence acknowledging the session theme, two action items in bullet format, one closing line that references their stated goal for next month. Keep it under 150 words. Match this tone: [paste a previous email you wrote that you liked]."
That's the same move Anthropic's engineers are making. You're not leaving the output format to chance. You're specifying it because your workflow depends on consistency.
Claude Code Isn't Just for Developers — It's the Preview of What's Coming
Claude Code is Anthropic's agentic coding tool. When Shihipar talks about how his team uses it, he's describing workflows where Claude is taking multi-step actions, making decisions inside a pipeline, and producing outputs that feed directly into other processes.
Most small business owners aren't there yet, and that's fine. But it's worth paying attention, because Claude Code represents the leading edge of how AI tools are evolving. The pattern — AI as an active participant in a workflow rather than a one-shot answer machine — is already available to you in lighter forms.
If you're using n8n or Make to automate parts of your business, you're already working in this paradigm. You're building sequences where one output becomes the next input. Claude Code is just a more powerful, more agentic version of that same logic applied to complex tasks.
The practical move right now: start treating your AI tool like a step in a workflow, not a search engine you occasionally consult. What does it need as input? What exact output do you need it to produce? What happens to that output after it leaves the AI? Answering those three questions will change how you prompt.
What the People Building These Tools Are Actually Optimizing For
Shihipar's interview is worth reading in full if you want to understand the current state of AI-assisted development. But the meta-lesson, the one that applies whether you're writing code or writing client proposals, is this: the engineers closest to these tools are optimizing for reliability, not capability.
They're not trying to use Claude for the most impressive possible task. They're engineering their workflows so that Claude produces consistent, predictable, usable outputs at scale. HTML over Markdown isn't a technical flex. It's a boring, practical choice made by people who care about whether the thing actually works in production.
That's the orientation worth borrowing. Less "what's the coolest thing AI can do" and more "how do I set this up so it works the same way every time."
One Thing You Can Do Today
Pick one prompt you use regularly — a client email, a content draft, a summary of your notes — and add an explicit output spec to it. Define the format, the structure, the length, and the use case. Run it three times and see whether the output is consistent enough to use without editing.
If it's not, tighten the spec. That iteration process is exactly what Anthropic's engineers are doing at a more complex level. You're doing the same work at a scale that fits your business.
The people who built these tools aren't using them casually. Neither should you.