Skip to content

docs: re-cut the demo so a human can actually read it - #59

Merged
Levironexe merged 1 commit into
mainfrom
docs/demo-readable-pacing
Sep 12, 2026
Merged

Levironexe merged 1 commit into
mainfrom
docs/demo-readable-pacing

Conversation

@Levironexe

Copy link
Copy Markdown
Owner

The problem, measured

The old cut ran 15.9s. Measuring the reading window against the content volume:

Beat On screen Characters Chars/sec demanded
chapter 1 terminal 4.5s 322 72
chapter 2 terminal 4.0s 137 35

Comfortable technical reading is 10-15 characters per second. Terminal output is parsed character by character, not skimmed like prose, because file paths and rule ids have no word shape. At 72 the only way to follow the first beat was to pause and step through frames.

What changed

Pacing only. No claim about the tool changed, and the numbers on screen are still the real output of a real run.

Constant Before After
frames between output lines 15 32
hold on the finished screen 60 120
claim card 75 95
Beat Before After
chapter 1 72 chars/sec 37
chapter 2 35 chars/sec 18

Chapter one also spent the viewer's time twice on one idea: direct_db_in_route and direct_db_in_page are different rules that read almost identically on screen ("Database client imported directly in an API route handler" / "... in a page or layout component"). The props builder now skips a rule whose message repeats one already picked, so the sample is one critical and one warning of genuinely different kinds.

Runtime goes 15.9s to 24.1s. Longer, but legible at speed instead of needing a pause.

The poster frame is now the settled check result rather than a black title card, so the video reads as a terminal before it is played.

Verified

node scripts/verify-props.mjs captures/architect passes, so props-dark.json still rebuilds from the raw capture and no number was hand-edited. pnpm run build passes. Frames inspected at the settled point of both terminal beats.

🤖 Generated with Claude Code

https://claude.ai/code/session_015e3bqn8QmG1H2fgHL15RoH

The old cut ran 15.9s and asked the viewer to read 322 characters of
terminal output in a 4.5s window. That is 72 characters per second;
comfortable technical reading is 10-15, because file paths and rule
messages are parsed character by character rather than skimmed. The only
way to follow it was to pause and step through frames.

Output lines now land 32 frames apart instead of 15, the finished screen
holds 120 frames instead of 60, and the claim cards get 95 instead of 75.
Chapter one drops from 72 to 37 characters per second, chapter two from
35 to 18.

Chapter one also showed two violations whose messages read almost
identically ("Database client imported directly in an API route handler"
and "... in a page or layout component"), spending the viewer's reading
time twice for one idea. The props builder now skips a rule whose message
repeats one already picked, so the sample is a critical and a warning of
different kinds. props-dark.json still rebuilds from the raw capture.

Runtime 15.9s -> 24.1s. The poster is now the settled check result rather
than a black card.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015e3bqn8QmG1H2fgHL15RoH
@vercel

vercel Bot commented Sep 12, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
leviron.architect Ready Ready Preview Sep 12, 2026 1:15am UTC

@Levironexe
Levironexe merged commit 948faee into main Sep 12, 2026
3 checks passed
@Levironexe
Levironexe deleted the docs/demo-readable-pacing branch September 12, 2026 01:16
Levironexe added a commit that referenced this pull request Oct 23, 2026
docs: re-cut the demo so a human can actually read it

This branch was successfully deployed

1 active deployment
Preview — 5a088888 Deployed Sep 12, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant