Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling?

Sources#

• How I AI: Thariq Shihipar on Replacing Markdown with HTML for AI-Powered Development

Question: Does the human-facing harness keep growing without bound, or does it hit its own bloat ceiling — an HTML plan too elaborate to read, like the markdown it replaced?

(This resolves the first open question logged in HTML as the New Markdown.)

Short answer#

Yes — but HTML raises and reshapes the ceiling rather than removing it. The bloat ceiling is structural, because the human-facing harness exists to fit finite human attention, the one resource no medium can make infinite. HTML buys headroom; abundance economics and artifact sprawl spend it back — and they spend it back faster as models improve.

The argument#

1. HTML is itself the symptom of a bloat ceiling being hit — markdown's#

The thesis only exists because thousand-line markdown plans "overwhelm the human… our eyes start to glaze over" (HTML as the New Markdown; raw source). Thariq's confession that he'd "ask Claude to edit the plan instead of reading it himself" is the diagnostic of having already crossed markdown's readability ceiling. We therefore have direct evidence that a human-comprehension medium can bloat past usefulness. The live question is only whether HTML is structurally immune — and it is not.

2. What HTML changes: it raises the ceiling and changes its shape#

• Raises it. HTML "spends tokens on legibility (mockups, structure, color, interaction) rather than raw text," packing more comprehensible signal per unit of attention (HTML as the New Markdown). This is explicitly the human-attention analog of the model's smart zone: clearing context restores the model's smart zone but quadratic attention still bounds it; HTML raises the human's effective smart zone without abolishing the bound (Context Window Smart Zone).
• Reshapes it. Markdown is linear — read sequentially. HTML is navigable and progressively disclosed: skim → drill → interact. The binding limit shifts from total volume to per-decision attention surface. Disposable Micro-Apps is the extreme: collapse one sub-decision into a focused UI surfacing exactly the attention needed and nothing else. That decomposition escape valve is what one monolithic markdown file lacked.

3. But the ceiling relocates rather than vanishes — from length to sprawl#

The bloat reappears on a new axis: not "one document too long to read" but "too many artifacts to track." This is the open question already flagged in Disposable Micro-Apps ("Where's the line between a disposable micro-app and tool sprawl? If every edit spawns a bespoke UI, does the workflow fragment?"), compounded by HTML's diffing/versioning weakness (only partially patched by copy-back-to-markdown) and Living Design System's "at what project size does maintaining the artifact cost more than the consistency it buys." Same ceiling, new dimension.

4. The invariant that makes the ceiling structural: human attention is conserved#

The binding constraint is "human attention and judgement, not generation cost" (Compute Allocator). Oversight past cognitive capacity raises minor errors +11% and major errors +39% (AI Brain Fry); humans have a degradation curve past capacity that mirrors the model's smart zone (Context Window Smart Zone). Any harness whose entire purpose is human comprehension is necessarily bounded by that finite resource. A medium can be more efficient per token of attention; it cannot make attention infinite.

5. The crucial asymmetry — the ceiling gets more binding over time, not less#

The model-facing harness can shrink toward zero as capability migrates inward (Harness Shrinkage as Models Improve, The Bitter Lesson). The human-facing harness cannot: the human's attention budget does not grow when the model gets smarter. Worse — abundance economics plus faster output means more artifacts competing for the same fixed budget. "Agents that produce more output faster reintroduce volume pressure" (AI Brain Fry). So as models improve, the human-facing harness faces increasing bloat pressure; the attention ceiling becomes the dominant constraint precisely because everything else got cheap. This is the inverse of the model-facing harness's trajectory — they diverge as capability rises.

6. What keeps it below the ceiling — and how it gets breached#

The human-facing harness has a bloat detector the model-facing one lacks: your own disengagement. The stated goal is literally "a plan you actually want to read" (HTML as the New Markdown). Model-facing bloat is invisible — it silently pushes the model into the dumb zone (Context Window Smart Zone) — whereas human-facing bloat announces itself the moment you glaze over. That is a self-correcting signal.

It gets breached via the Compute Allocator / AI Brain Fry failure mode: abundance makes generation free, so nothing mechanically stops you producing more than you can attend to → attention collapses into rubber-stamping → you cross the ceiling without noticing. The discipline that prevents it is the human-facing twin of Cat Wu's "read the whole system prompt every launch and delete what isn't earning its place" (Harness Shrinkage as Models Improve): an artifact must justify its existence by your actual engagement, or it is bloat.

Bottom line#

Yes, it hits a ceiling. HTML's contribution is to push the ceiling out (denser legibility) and soften it (navigable/decomposable instead of linear), with micro-apps and copy-back-to-markdown as escape valves. But because the conserved resource is finite human attention — and because better models pour more generated output against that fixed budget — the human-facing harness does not escape bloat. It relocates the failure mode from "document too long to read" to "too many artifacts / rubber-stamping," and that ceiling becomes more binding, not less, as models improve.

Sources#

• HTML as the New Markdown — the thesis; legibility-spend framing; the open question this answers
• Compute Allocator — attention/judgement as the binding constraint; rubber-stamping failure mode
• Disposable Micro-Apps — the decomposition escape valve and the tool-sprawl relocation
• Living Design System — maintenance-cost ceiling on durable artifacts
• Harness Shrinkage as Models Improve — the model-facing/human-facing asymmetry; pruning discipline
• Context Window Smart Zone — human smart zone analog; invisible vs. visible bloat
• AI Brain Fry — 11–39% error surface; output-volume pressure
• The Bitter Lesson — what migrates inward (model-facing) vs. what does not (human legibility)
• How I AI: Thariq Shihipar on Replacing Markdown with HTML for AI-Powered Development — primary source