A short note on why the format you send a thing in matters more than the thing itself, and a workflow shift that's about to take work off your desk.
Vault notes, agent files, memory logs — they all stay markdown. Forever. But the moment a human needs to actually read, decide, and respond? That's an HTML moment.
Grep-able. Diff-able. Loadable into any AI session. Survives the apocalypse and the next Obsidian update. This is the bedrock — and it never moves.
Visual hierarchy. Mockups. Diagrams. Editable interfaces. Sections you actually scroll. The medium humans engage with when the stakes are decision, not storage.
A 1,000-line markdown plan gets skimmed or skipped. An HTML plan with sections, cards, mockups, and visual hierarchy gets read. If you stop reading the plan, you stop steering the agent.
The Loop ClosesAn 8-hour agent run = $500 of compute. You're now an allocator, not a typer. The plan is where allocation happens. Unreadable plan = blind spend.
$$$ Per HourASCII mockups in markdown are pitiful. HTML mockups, charts, diagrams — same effort for the model, dramatically more communicable for the human reading it.
Free UpgradeDon't like a section? Don't re-prompt in text. Ask Claude to build a custom editing interface for that one piece. Throwaway by design. Micro-software on top of micro-software.
See Pattern 02 ↓Same plan, two formats. Which one would you actually finish reading?
Multifamily-only → multi-audience via URL params. Luxury, rooftop, live builds.
/projects · ?type=luxury-home · ?type=rooftop · ?status=live
Old routes → new param-driven equivalents. Configured in astro.config.
// The prompt that got you here
"Create an HTML plan file with mockups, code excerpts, and whatever else gives me max context. I trust your judgment on structure."
A section of the plan is wrong. Don't re-prompt in text. Ask Claude to build the ideal interface for editing it. Try it below — click options, copy the output.
selected: Option 02 - Blueprint ($29,000) scope: SD + DD position: recommended status: confirmed
// The prompt that built this
"Create an editable HTML artifact for [this section]. Make a custom UI that helps with structure but gives flexibility. Design the ideal interface for this problem."
Drop pins. Add notes. Copy the comment thread back as your revision prompt. Hover the pins below.
You're building a hillside legacy home, not just another lot. The two-rooftop terrace system isn't a perk, it's the spatial argument the architecture is making.
Most LAs would lay out beds and call it a plan. We see your 12,000 SF as a single cinematic gesture: arrival, ascent, vista, retreat.
This proposal lays out three ways we could partner — each builds on the last, none of them dilute the vision.
Build the review interface you actually want — not the one the tool gives you.
The vault is the substrate. The substrate doesn't get an upgrade. If you convert these to HTML, you break grep, diff, and AI loading. Don't.
One test. Compare review time. Port if faster. Ignore if not. Markdown plans still work for small-scope specs — this isn't a religion.
Use: "Create an HTML plan file with mockups, code excerpts, and whatever else gives me max context. I trust your judgment on structure."
Did you actually read it? Did revision land in fewer cycles than the last markdown spec?
If faster → port the pattern. No further ceremony. If not → ignore. Done.