Robert Blust's profile page and two talks. Self-contained, no build step, no external
assets. Working conventions and the traps that break this site silently are in
CLAUDE.md.
/ profile page — thesis, talks teaser, links
/talks/ talks index — the descriptions live here
/talks/mental-model/ deck: The Mental Model
/talks/essential-complexity/ deck: Essential Complexity
The profile page and the talks index deliberately list the same two talks: the profile
page carries a one-line teaser (title, length, language, link), the talks index carries
the description. Editing a talk means checking both files — see CLAUDE.md.
Each deck is bilingual (German content, English via data-en), self-contained, and
works from file:// as well as a local server.
npm install && npx playwright install chromium
npm run serve # python3 -m http.server 8000
npm run verify # Playwright DOM assertions against all four pages, plus the sitemap
npm run og # regenerate the four 1200×630 og:image share cards
npm run pdf # regenerate both decks' PDF fallbacks
./tts/generate.py --dry-run # narration: what would be billed, and for which slides
./tts/generate.py # narration: generate what changed, both decksRun npm run verify after any change under index.html, talks/, or verify/. Run
npm run og and npm run pdf after a visual change to either deck or to index.html /
talks/index.html — the share cards and PDFs are rendered, committed files, not
generated on demand.
Narration is generated from the speaker notes themselves and cached on a content hash, so
editing one note regenerates one clip. It needs ELEVENLABS_API_KEY, which lives in
~/.zshrc and is therefore invisible to a non-interactive shell — CLAUDE.md has the
one-liner that reaches it, and the rule about never printing it.
The two decks were brought into this repository with their git history intact, as subtree merges. The cut-over runbook's own draft says to confirm this with:
git log --oneline -- talks/mental-model/That command proves nothing.git log with a path filter does not follow history
across a subtree merge commit — it shows only the merge itself, not what was merged in,
regardless of whether the history is actually there. A clean-looking one-line result
from that command is not evidence of anything.
An earlier draft of this runbook recommended git cat-file -e <sha> against this local
repository instead. That is not sufficient, and must not be reinstated.cat-file -e
only proves an object is reachable from local refs on this disk — it says nothing about
whether it survives. git remote -v here is currently empty: nothing has ever been
pushed anywhere. A local object passing cat-file -e is fully consistent with that
history existing on exactly one machine — the one about to have its only other copies,
the two source repositories, deleted. Since that deletion is irreversible, a gate that
can pass while the history has a single point of failure is worse than no gate.
The check that actually protects against the deletion requires two steps, in order:
Push this repository to GitHub first. Until
originexists and this history is on it, there is no second copy to verify against — only local objects that would die with this disk.Verify against the pushed remote, not local
HEADand not local objects:git fetch origin git merge-base --is-ancestor 0b5acb75e48d2a8ad7132fd1b4abc15bf20dc6ea origin/main git merge-base --is-ancestor 77757acea99102d128916e305d9f59e81610d08b origin/main
Both commands must exit 0. --is-ancestor against origin/main confirms each source
repository's original tip commit is not merely present somewhere but actually reachable
from the branch GitHub is serving — which is what has to be true for the history to be
recoverable once the source repositories are gone. Run both, against the remote, only
after pushing, before deleting robertblust/mental-model or
robertblust/essential-complexity on GitHub.
- Commits happen when the user asks; suggest a message, don't auto-commit.
- Never mention closed-source predecessor projects — here, in docs, or in commits.