Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Why this Claude Code engineer uses HTML files as AI specs | Thariq Shihipar (Anthropic)
TL;DR
HTML beats markdown when plans get too long to actually read — Thariq Shihipar says Claude plans can now run to 1,000 lines or cover an hour-long agent run, and HTML makes them more scrollable, visual, and engaging enough that he’ll still stay in the loop instead of blindly trusting the model.
Specs still matter because AI work is really compute allocation — his line is that telling Claude to run for 8 hours is basically authorizing it to spend around $500, so planning, PRDs, and implementation specs remain the place where humans decide what deserves that budget.
The prompt can stay simple if the artifact is right — for brainstorming and planning, he used plain requests like “create an HTML file as a plan” and “include excerpts, mockups, code… whatever is needed to give me maximum context,” arguing that overbuilt prompts often overconstrain Claude.
He uses HTML to generate throwaway ‘micro software’ for decision-making — when he didn’t like one rules table in the plan, he had Claude build a custom editable HTML UI just for that sub-problem, then copied the output back into the main spec.
HTML artifacts become shared working surfaces, not just documents — Shihipar says he sends weekly status updates to his manager Kat as HTML, shares plans as hosted links, and keeps living design systems as design.html files that Claude can reference across repos.
Verification is bigger than unit tests now — he and Claire connect types, validation criteria, synthetic test data, rubrics, and even recorded videos as modern ways to verify agent work, especially with outcome-oriented systems like Anthropic’s newly announced Outcomes.
The Breakdown
“HTML is the new markdown” lands hard
At Anthropic’s Code with Claude event, Thariq Shihipar drops the line that stuck with Claire: HTML is the new markdown. His argument is practical, not trendy — markdown worked when agent plans were short enough to read and edit, but now that Claude can run for an hour and generate giant plans, he found himself not reading them anymore, which he calls “honestly a mistake.”
Why specs survive the agent era
Before getting into tooling, they pause on a bigger point: planning still matters. Thariq says if you tell Claude it can run for 8 hours, what you’re really saying is that it can spend roughly $500, so everyone is becoming a “compute allocator” who has to decide what work is worth the budget. Claire riffs on this as the post-product-manager identity: same core judgment, new resource.
Brainstorming in HTML instead of drowning in text
Thariq shows a real workflow in Claude Code: he asked Claude to brainstorm demo ideas for Claire’s podcast in HTML, and it returned eight visual concepts with mockups, rationale, and risks. The point isn’t that HTML is easier for the model — it’s that he’ll actually scroll and read it, whereas he has a rule of thumb that he won’t read text output longer than one Claude Code screen.
From interview to implementation plan
After picking the “CSV to interactive dashboards” idea, he had Claude interview him to surface unknown unknowns, then asked for a plan in HTML with mockups, code excerpts, and whatever else would provide “maximum context.” The result blended product thinking, file structure, helper scripts, UI examples, and even podcast scripting into one artifact Claire says is exactly the kind of thing she would actually read, unlike a mega markdown blob.
The prompting trick: specify intent, leave an out
A nice moment here is how little prompt-engineering theater there is. Thariq says the sweet spot is giving Claude enough direction without overconstraining it with elaborate personas or rigid scaffolding; in his phrasing, he always wants to “give Claude an out.” That’s why he names must-haves like code excerpts, but ends with “whatever is needed to give me maximum context,” signaling trust in the model’s judgment.
Turning one bad table into a custom app
The most memorable demo is when he dislikes a specific rules table inside the plan and, instead of editing raw text, asks Claude to make an editable HTML artifact just for that decision surface. Claire nails the framing: this is “micro software on top of micro software” — a throwaway custom UI for one tiny planning problem, complete with editable fields and copy-back markdown.
HTML as a collaboration layer, not just a solo hack
When Claire asks whether this breaks in teams, Thariq says the opposite: HTML plans are easier to host and share, so coworkers like Kat or Boris are far more likely to read them. He even says he sends weekly HTML status updates to his manager Kat, generated from Slack, because websites are simply more consumable than static docs.
Living design systems, verification, and orbital compute
They end by extending the idea beyond plans: Thariq keeps living design systems as HTML artifacts that can be passed around repos, while Claire describes building component visualization pages marketers can grab transparent PNGs from. They also tee up a future episode on testing, with Thariq’s memorable line that “test verification is not testing,” then close on Anthropic’s SpaceX partnership and the deliciously sci-fi idea of orbital data centers.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
