Zilin Slide Narrator

Zilin Slide Narrator

Produce an audience-facing slide player and narrated MP4. Start with a compact interactive intake modeled on the cat-hatcher workflow: show defaults, accept “keep defaults,” apply partial overrides, and carry accepted choices forward. This skill can run independently of that skill.

Workflow and review gates

1. Intake before production. Read intake and load defaults. Collect source, audience, rough script, coverage, voice, review, visual, output, and deployment choices. Encourage incomplete notes or key points; a complete initial script is not required. Read-only source inspection may proceed while waiting. Do not draft a full walkthrough or synthesize audio until the plan and source are settled.
2. Inspect every slide. Read teaching. Establish the authoritative deck, render it, and map every original slide to included/excluded status. Do not reuse another deck's page count or narration. Attached slides take precedence over a different web version unless the user says otherwise.
3. Write and rehearse. Produce a complete per-slide instructor script. Generate a synchronized player and video using macOS Samantha at 150 words/minute. Revise cheaply until the user approves the script and native-voice draft. Native voice is a development aid, not a claim about final quality.
4. Two-slide paid-voice gate. Read audio production. Check voice availability and account balance; generate only the first two included slides with the confirmed ElevenLabs parameters. Download, assemble, and show their player/video preview. Wait for approval before bulk generation. Do not infer approval from silence. A request to change parameters and preview again resets this gate.
5. Finish approved audio. Generate the remaining approved scripts within the authorized credit budget. Save and verify each clip before moving on. Reuse completed matching clips. A credit shortage means assemble current work, disclose unfinished slides, and wait for a balance/account change; do not buy plans. User-authorized account changes require a fresh balance/settings check.
6. Align captions to the audio. Read alignment and assembly. Use forced alignment on the final playable recordings, not text-length timing. Rebuild player CC, SRT/VTT, and MP4 subtitles from the same aligned word data. Verify every included slide and the complete output.
7. Deliver or deploy. Read delivery. Always provide the player and video. Ask whether deployment is needed during intake; local/package-only is the default. When requested, build a course-level homepage and a clean static package, then publish only to the chosen destination/access under existing authorization.

Defaults and customization

The JSON template is the source of defaults; user choices override it. For any PDF deck, include all slides by default and confirm exclusions explicitly. The optional teaching preset preserves the originating workflow: include announcements/hackathon if present and exclude office hours and questions/Q&A. Do not apply course-specific exclusions to an unrelated presentation.

Use all non-excluded slides, teach all substantive slide contents, avoid added running examples, and use concise but complete instructor speech. Explain relationships and transitions rather than reading bullets. Original code/examples already on a slide still need explanation.

No changes to voice, model, speed, narration, or deck should silently reuse stale approval or timestamps. Record approvals against the exact script/source/settings, and invalidate only the affected downstream work. Source audio remains untouched; import and align derived playback audio.

Persistent run record

Keep run.json beside production files with source checksum/page count, resolved defaults/overrides, per-slide inclusion rationale, script and audio checksums, actual durations, voice/model/settings, generation state, observed credit balance/time, approvals, alignment signatures, and deployment status. Do not store account credentials. Distinguish generated, downloaded, verified, and assembled; a generation request alone is not completion.

For resumed tasks, inspect this record and existing artifacts before asking questions or generating. Honor “stop generating” immediately; assemble saved clips if asked. Do not regenerate approved audio merely to recover a download or fix CC.

Audience-facing quality

Player defaults: dark navy/gold, automatic slide progression, seek, speed, slide selector, keyboard controls, fullscreen, CC, per-slide transcript, and video/caption/slide downloads. Keep voice parameters, debug labels, credit balances, editable-script links, cache tokens, internal paths, and raw errors out of the audience interface. Never claim the style matches a course website without inspecting it. Offer intentional style matching during intake.

References and helpers

• Intake: staged prompts and confirmation boundaries.
• Teaching: source mapping, script quality, review.
• Audio production: Mac drafts, ElevenLabs gates, credits, downloads.
• Alignment and assembly: actual-audio timing, validation, synchronized outputs.
• Delivery: package, homepage, deployment, handoff.
• scripts/check_run.py RUN.json: lightweight preflight for stage/approval consistency; does not authorize external actions or replace artifact verification. Read its error output and resolve mismatches before paid generation.