Augment Speakers and Panelists

Status: Draft (2026-10-26). Written for event-hog during the 7x24 Exchange Fall Conference; not yet run end to end against a real speaker.

Event apps give you a name, a title, a company, and two lines of bio, usually as a screenshot. Listening and taking notes at the same time is hard, so the team needs to know who someone is before or during the session, fast. This skill turns that thin bio into a short profile, in a fixed order:

1. Social accounts: where to follow up or reach them.
2. Features: coverage about them, or by them.
3. Mentions: they're quoted or named in a broader story.
4. Video: talks and interviews, usually YouTube.

0. Resolve the configuration first

The method here is the same for everyone. Which research tools to use is not. One person has Firecrawl and Tavily, a team has Origami, another has only the web search built into their chat app.

• Connected to event-hog: call get_config with skill: "augment-speakers-panelists". It returns the organization's defaults merged with this person's overrides: research_tools (in order of preference), limits, and where results are saved.
• Otherwise: read config/private.json next to this file (shape in config/private.example.json).
• If neither exists, ask once which research tools the person has. If event-hog is connected, save the answer with set_my_research_tools so they're never asked again.
• Use the tools in the order listed. Fall back to the next one when a tool is missing, errors, or returns nothing useful. The client's built-in web search is always the last resort. Never refuse to start because a preferred tool isn't there.

Defaults when the configuration doesn't say: 10 features and mentions combined, 5 videos, look back 3 years.

1. Read the input and anchor the identity

From the screenshot or text, extract exactly what's shown: name, title, organization, the session or panel title, its time and room, and any bio lines. Don't fill gaps from memory.

Confirm identity before researching. Common names collide constantly ("Ryan Bernstein" is several people). The anchor is name + organization, plus title or location when shown. Every result must match the anchor:

• High: the account or article names the same organization or role.
• Medium: same name and industry, organization not stated. Keep it, marked (probable match).
• Low: same name only. Leave it out.

If the very first searches can't find any high-confidence match, say so and ask before continuing. A confident profile of the wrong person is worse than none.

2. Social accounts (first, always)

Look for, in this order: LinkedIn, X/Twitter, Bluesky, a personal site or newsletter (Substack, Medium, blog), GitHub (for technical speakers), a YouTube channel, Mastodon or Threads. Include Instagram or TikTok only if they're clearly professional.

• Report only accounts you actually found and opened, or that a reliable page links to (their company bio, the event's speaker page). Never construct a handle or URL from the name.
• Confirm each one against the anchor: bio mentions the organization or role, or the company site links to it.
• No personal or family accounts, home addresses, or personal phone numbers, even if found. This is professional information only.

Then the organization's accounts: its LinkedIn company page and its X account (plus YouTube if it posts talks). Posts about the session tag both the speaker and the organization, so both need to be ready. Take them from the organization's own site (footer links) or a confirmed search result, never from a guess at the name.

3. Features, then 4. mentions

• Feature: the piece is about them or by them: an interview, profile, podcast episode with them as a guest, an op-ed or byline they wrote, a talk write-up centered on them, a company announcement that's mainly about them.
• Mention: they're quoted or named in a story about something else.

Search mainstream and trade press (for data centers and energy, for example: Data Center Dynamics, Data Center Frontier, Utility Dive, trade association sites), plus podcasts.

Selection, when there are many:

1. Relevance to their session or panel topic first. If they're speaking on AI-era power density, a 2025 interview on cooling beats a 2026 item about a promotion.
2. Then recency: newer first.
3. Cap features and mentions combined at the configured limit (default 10). Features usually earn more of the slots, but a highly relevant recent mention can outrank an old feature.

For each item: date, outlet, title, URL, and one line on why it matters for this session.

5. Video

Talks, panels, interviews, webinars: usually YouTube, sometimes Vimeo or the event's own site. Same rules: relevance to the session, then recency, capped (default 5). Include date, title, channel, URL, and length if shown.

6. Output, then save

Show a card readable on a phone, in this order:

**<Name>** · <Title>, <Organization>
Session: <title> (<day, time>)
Why they matter here: <one line>

Social: LinkedIn · X · …              (only what was found and confirmed)
Organization: LinkedIn page · X · …
Features (n): date · outlet · title — why it matters
Mentions (n): …
Video (n): …
Not found / not confirmed: <what you looked for and didn't find>

Then save it:

• Connected to event-hog: call save_speaker_profile. It writes speakers/<name>.md in the event workspace and records each link as an observation (has_social, featured_in, mentioned_in, has_video) with its URL as the source, tied to this event. Pass the organization's accounts as organization_socials; they're recorded on the organization. Saving the same person again adds to the profile instead of duplicating it.
• Otherwise: write the card to a markdown file where the user keeps event notes, with every URL kept.

Discipline

• Only report links you actually retrieved. If a tool returns a snippet but the page wouldn't load, mark it (not opened).
• Say what you didn't find. "No X account found" is information; silence is ambiguous.
• Dates matter. An undated item goes after dated ones and is marked (undated).
• Speed over completeness at the event. If time is short, do social accounts and the 3 most relevant features first, save, and offer to finish later.
• Speakers aren't contacts. A profile doesn't mean the team has met them. Use log_contact only after a real conversation.

See also

• event-prep: the prep map this enriches; speakers ranked there are the first to augment
• crawl-fetch-ingest: the deeper ingest workflow for firm and portfolio pages
• lossless-crm-interface-guidelines: provenance habits (fetch first, offer model knowledge rather than smuggling it in)