## Executive summary
Eugene is shaping an internal webinar around one thesis: non-programmers can and should build their own small AI tools, because the old belief that "software is slow and expensive" is now false. A one-button local tool he built — record → diarize → transcribe → structured summary — already replaces hours of manual write-up, and Nina (an HR recruiter) confirms the demand with concrete pain points: job descriptions, interview transcripts, and candidate summaries. Both agree the real blocker to adoption is friction, not willingness — colleagues would use these tools "if it were simple."

## Who's who
- **SPEAKER_00 → Eugene ("Женя")** — senior software engineer (20+ years, ~20 commercial), works at Inspectron on Edge Compute / IoT; author and presenter of the AI webinar; heavy Claude user; evangelist for personal AI tooling ("harnesses") and the idea that "every person becomes a business."
- **SPEAKER_01 → Nina** — HR/recruiter at Virtido; ultra-trail runner (50+ km); among the most AI-forward recruiters on her team; uses paid ChatGPT, LinkedIn Recruiter, and Manatal (ATS). Enthusiastic practical adopter and supportive foil who supplies the real-world use cases.

## Key themes

**AI is a multiplier, and "it makes mistakes" is a weak objection.** Both frame AI as amplifying human work rather than replacing judgment, and both bat away the reliability complaint by comparing it to human error rates.
> "AI is a multiplier — it doesn't cancel the brain, it just helps the brain do the work it already has faster and better."

**For recruiters, the transcript matters more than the AI summary.** Nina's biggest pain is interview write-ups; note-taking during a call wrecks the conversation and memory fails — especially on cross-timezone calls at 5 a.m. Automatic transcription is the part that actually changes the work.
> "This transcript is honestly the most important thing. Because when you take notes yourself while talking to a person, it's useless — I don't concentrate on the conversation, I'm afraid I didn't finish writing down his previous sentence."

**Build your own tools — the "software is slow and expensive" belief is dead.** The webinar's central pitch: most routine-task tools you can now create yourself, and running locally on a subscription removes cost, licensing, and security concerns.
> "Most of these tools you can build yourself… humanity still carries this outdated belief that software is slow, that software is expensive. No."

**Agents plus a real browser beat paid sourcing tools.** Eugene demos Claude Code driving a live browser to source candidates/companies in the background — a practical, cheaper end-run around LinkedIn Sales Navigator, with anti-bot risk at high volume.
> "Find me companies doing Computer Vision, find the most senior reachable contact, save the link to their profile — but only companies under 200 people. And it works in the background."

**Skills turn personal expertise into a transferable, packageable asset.** Do the task through AI, then freeze it into a "skill" — plain text zipped into a folder — that a brand-new hire can run to produce comparable output, cutting onboarding and spreading responsibility.
> "'Create a skill for this.' After that you have a folder you can zip up and hand over… and a person with not even a third of your HR experience can deliver a decent result."

**The payoff is lower cognitive load, not just more output.** Nina could go from 5 to 25 candidates a day, but the deeper win both land on is freeing mental capacity for higher-value work.
> "Even if your output doesn't change, your cognitive load changes — your life overall gets better." / "More creative brain power."

## Conclusions
- The webinar thesis holds up: ordinary users have real, repetitive problems (JDs, interview write-ups, LinkedIn outreach) that small AI tools solve **today**, not eventually.
- Adoption is blocked by friction, not resistance. Nina's colleagues aren't opposed — they lack a simple, all-in-one path (and not all have paid GPT). "They'd use it if it just transcribed everything for them."
- Local + subscription beats API/SaaS for this class of tool: no per-call cost, no license violation, no login/security surface to worry about.
- Human review stays essential, but the bar shifts — from "redo the AI's work" to "did I feed it the right inputs" (e.g., don't leak a client name into a prompt).
- Skills are the real leverage and a de-risking tool: package once, hand off, and take a vacation without being on-call — though Eugene and Sebastian disagree on whether skills are personal IP or employer work product.

## Takeaways
- **The tool:** one red button → record → diarize + transcribe + analyze → structured report (who's who, key themes, conclusions, takeaways, open questions). ~10 minutes to process 1 hour of audio, locally on a GPU, on a Claude subscription.
- **Don't "teach the AI" abstractly** — do your real task through it, watch the result, correct it, then freeze that into a reusable skill.
- **Organize by project/agent**, not one catch-all chat — one responsible agent per project; Eugene ran 7 in parallel without losing track.
- **Knowledge base (Karpathy-style, via Obsidian):** ingest sources, query, and let synthesized answers auto-feed back into the base. Eugene plans to share this file as a post-webinar "gift" attendees can try the same day.
- **Recruiting sourcing** via Claude Code on a live browser (e.g., Computer-Vision companies <200 people, Dutch region, top contact + link) — but watch anti-bot patterns at team volume.
- **Multilingual output** (English/German/Ukrainian) is a genuine unlock for Nina.
- **Consent first:** tell candidates you're recording/transcribing before the call.
- **Webinar:** first one in English, titled *"From a chatbox to your own operating system."* Keep it aimed at non-programmers, not another dev talk.

## Open questions
- How to bridge Eugene's programmer-grade demos to a non-programmer's actual workflow — he admits the simple examples "hang in a vacuum" with no clear on-ramp.
- Can the transcribe→summarize tool plug into **Manatal** (and sync across recruiters, e.g., via git) rather than only LinkedIn?
- Will LinkedIn flag automated browsing at HR-team volume? Eugene hasn't hit limits but doesn't know the threshold.
- Are skills "intellectual property" you own, or work product owned by the employer — unresolved with Sebastian.
- Larysa (now in project management) hasn't been interviewed yet — her PM use cases remain to be gathered.