Live presentation → article → channel pipeline

A pipeline that turns a single live talk into three artifacts with one human-review step: (1) the slide deck itself is generated from the speaker's working docs as context; (2) during the talk, a live transcript captures everything said into a context-loop file; (3) immediately after the talk, a content skill writes a publishable article from the talk + slides + transcript, attaches a matching cover via the banner skill, and posts to the founder's channel after one human approval. End-to-end: a 5-minute review window between "talk ended" and "post published".

A founder who gives a one-hour talk normally spends 4–8 hours converting it into a publishable article — and most never do. The compounding cost is enormous: every talk that doesn't become an article is wasted distribution. This pipeline reduces the article step to ~5 minutes of human time and reliably produces an article with the speaker's verbatim voice (because the transcript is the source, not a paraphrase). Over a year of monthly talks, that is ~50 hours of founder time and 10+ extra published posts.

1. Pre-talk: deck from context. The speaker drops research notes, working docs, and prior posts on the topic into a folder (talks/{date}-{topic}/context/). Run a skill: "Generate a 12-slide HTML/web presentation on {topic} using only facts in context/. Mark anything not in context as [PLACEHOLDER]. Style per presentation-style.md." Output: an HTML deck the founder can open in a browser and edit.
2. During talk: live transcript loop. Use a live transcription tool (Krisp, Whisper-based local transcribe, or any browser-based live captioner) configured to write a rolling transcript to talks/{date}-{topic}/transcript.md once per minute. This file becomes the single source-of-truth for what was actually said — not what the slides claimed would be said.
3. Post-talk: article skill. Invoke aim-content (or equivalent) with the talk folder. The skill reads transcript.md, the slide deck, and context/, and produces articles/{date}-{topic}.md shaped as a long-form post: headline, hook, structured sections matching the talk's actual arc, embedded pull-quotes from the transcript, calls-to-action.
4. Voice rule. The article skill reads content-voice-{channel}.md (case #27) so the published voice matches the channel, not the talk's spoken cadence.
5. Cover banner. The same skill triggers the banner skill (case #26) with the article title + topic, producing a cover that matches the channel's visual brand.
6. Single human review. Founder opens articles/{date}-{topic}.md, the cover, and the preview. One pass to catch hallucinations and tighten the lead. Edits go straight into the file.
7. Publish. Skill posts to the channel (Telegram MCP, Substack, LinkedIn, or plain HTML on a static site) with the cover, the article, and an OG-preview verified before posting.
8. Archive loop. The talk folder (context/, transcript.md, articles/{date}.md, deck HTML, cover PNG) stays in the workspace as future context for the next talk on a related topic — this compounds the speaker's content graph.

• 3-hour presentation prep workflow. When the article exists already, presentation prep collapses to: (a) talk file → outline (15 min); (b) outline → slide structure (30 min, Claude); (c) Reveal.js render (10 min); (d) operator-edit pass on visuals (1 hour); (e) speaker-notes rehearsal (1 hour). The discipline is to NOT iterate the article during prep — freeze the article first, then prep.

Deck generation prompt:

Generate a 12-slide HTML presentation on {topic}.
Use only facts from the files under talks/{date}-{topic}/context/.
For any claim you cannot ground in those files, write [PLACEHOLDER: <claim>].
Style: read talks/presentation-style.md.
Output: talks/{date}-{topic}/deck.html
Each slide is a section with id="slide-N"; speaker notes go in <aside class="notes">.

Article skill instruction:

Read:
 - talks/{date}-{topic}/transcript.md (verbatim)
 - talks/{date}-{topic}/deck.html (structure)
 - talks/{date}-{topic}/context/ (factual grounding)
 - content-voice-{channel}.md (voice rule)

Produce articles/{date}-{topic}.md as a long-form post:
 - Headline (matches the talk's strongest claim).
 - Hook (200 words, drawn from the transcript's actual opening).
 - 4-6 sections matching the talk's actual arc, not the slide titles.
 - At least 3 verbatim pull-quotes from the transcript.
 - Closer that previews the next talk in this series.

After writing, run fact-check loop per content-voice protocol.
Then trigger banner-skill with title + 3-word topic, save cover to articles/covers/.

• The transcript is the source of truth for the article, not the slides. If the speaker improvised, the article should reflect that — do not have the skill "smooth over" improvised sections back into the deck's outline.
• Don't auto-publish. One human review pass is non-negotiable — talks routinely contain off-the-cuff claims the founder would not put in writing.
• Don't generate the article from the deck alone (no transcript). The output is generic and loses the voice. The transcript file is the differentiating asset.
• Anonymize any client/customer references in the transcript before the article step if the talk was internal-flavoured.

• Lighter: drop the deck generation step; only run the post-talk article pipeline on transcripts from existing talks. Wins ~80% of the value.
• Heavier: for a podcast series, chain three talks → three articles → a synthesis post that cross-links them, generated weekly.
• Internal-only: instead of publishing externally, save the article to the company wiki / internal knowledge folder. Every team meeting becomes a searchable artifact.
• Reveal.js in-code variant. Instead of (or in addition to) PPT export, generate Reveal.js HTML directly from the structured talk file. The skill takes the same input and outputs a slides.html runnable in any browser, with a --style parameter (e.g. editorial, keynote, cli-deck). Replaces a separate slide-design step for technically-comfortable presenters and integrates cleanly into the static-host flow (Netlify / Vercel). A franchise founder used this to ship 3 presentations in one week.