Other

/audio-dashboard

Build audio narration dashboards. Triggers: STT-after-TTS exact word-timing, real word-click-seek read-along dashboard, AfterCode workflow, Lakebed delivery.

$ golems-cli skills install audio-dashboard
18 evals
fixtures
1 workflow

Updated 4 days ago

Building a dashboard?

Run /audio-dashboard creation-steps: read and follow workflows/creation-steps.md, the ordered procedure from intake to delivery. This is a workflow argument to the skill, not a separate catalog command. This skill owns all build instructions; orc only relays the filled intake block and Etan's extra instructions.

Ownership and delivery

skill-creator owns this skill, templates and BUILD gates. NarrationLayer owns the Lakebed app, relay/media server, account answer sync and UploadThing, and is the ONLY publisher. Builders never deploy a dashboard shell themselves. Audio may use CDN/relay or tailnet media fallback (8447); tailnet may host redirects, not the shell. Audio delivery follows the html-dashboard Lakebed/allowedEmails standard through NarrationLayer; see publish contract v2.

Proof images in this repo are mock only. golems is public. Every proof screenshot, diff image, GIF or video committed under evals/proof/ (or anywhere else in golems as proof) is rendered from a synthetic fixture such as evals/fixtures/card-visuals.html. Anything rendered from one of Etan's real pages (his words, his data, a real briefing) stays in the private skill-creator repo or in docs.local, never in a golems commit; a PR links to it there, it does not attach it. (Etan, 2026-11-13: "skillcreator is where we upload my own things modified since its private, golems is public and should have mock/censored versions".)

Purpose

This is the one canonical golem-powers skill for audio narration dashboards, and it is self-contained and transferable: the whole engine is vendored under vendor/ and driven by skill-local scripts under scripts/. Runtime/build paths do NOT clone or reach into $NARRATIONLAYER_ROOT, $AGENT_HTML_ROOT, or $SKILL_CREATOR_ROOT; the opt-in maintenance lint reads only the checkout named explicitly by NARRATIONLAYER_UPSTREAM. The only external requirements are machine-local binaries/daemon (verified by scripts/bootstrap.mjs) and the NarrationLayer Lakebed publish handoff.

The invariant: never ship read-along timing from WPM math, line duration, fake words, placeholder recap text, or old GolemPlaylist generators. A correct dashboard keeps immutable raw Whisper words.raw.json for gates, separate script-first repaired words.json for display, real transcript text, and word-click seek to the clicked timestamp.

Workflows

/audio-dashboard:creation-steps