Files
WebinarNotes/raw/sources/Yulia interview.md

56 lines
6.1 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
## Executive summary
An HR/recruiting lead (SPEAKER_00) brings a senior programmer/AI consultant (SPEAKER_01) three team pain points, and he demonstrates that all of them are solvable today with Claude plus reusable "skills." The firm conclusion: their two recruiting problems collapse into one — build a searchable candidate knowledge base (record interviews → auto-transcribe → auto-generate profiles) — and the next step is a webinar he'll plan over the coming weekend, with a possible paid build of an HR tool to follow.
## Who's who
- **SPEAKER_00** → HR/recruiting lead (female), collecting her team's AI pain points and organizing the webinar. Self-described beginner; refers to Nina, Sergiy and "Andryuk" as the people who approve budget, and to "Jean-Luc/Zhenlyuk" who posts about AI.
- **SPEAKER_01** → Senior programmer (~20 years), works on "Inspectron" and its clients. Heavy Claude user on the $200 max plan; proposes the webinar and offers to build the HR system for money.
## Key themes
**AI usage has distinct levels, and most people are stuck at level one.** He maps a ladder: web chatbot → built-in memory → Claude Code/Cowork with local file access → CLAUDE.md → skills → Obsidian knowledge base → RAG. The practical ceiling for non-programmers is CLAUDE.md plus skills.
> "Usually you need CLAUDE.md and skills. That's your maximum."
**Skills are reusable process memory — solve first, skill-ify after.** The recurring mistake is building the skill up front. Instead, solve a task with Claude once, then ask it to package that into a skill it can rerun on command.
> "The mistake people always make is they create a skill first and then try to shove it somewhere… You first solve a task with Claude; the moment you reach the final solution, you say — now create a skill from this."
**Both recruiting problems are the same problem: a candidate base plus search over it.** Recording interviews, auto-transcribing (with speaker diarization), and generating standardized profiles feeds a knowledge base you can later query by skill — including finding people you already spoke to who match a niche stack.
> "You need a candidate database and you need search over that candidate database."
**AI automates all information work; the human keeps the human parts.** His central thesis for the webinar — the residual human role is connection and communication, not information processing.
> "The human's role in this process is just to be human — introductions, communication… Because our reality is all information work. And AI can replace all of it."
**AI prices will rise, not crash — so invest in the skill now.** He argues the "bubble" pops upward: capable AI gets more expensive, and the leverage shifts to knowing which models to use.
> "What I now buy for 200 will cost about 1,000… Expensive AI will become even more expensive."
**Economics favor renting GPU over owning.** Live demos (RunPod for transcription, Claude driving a browser to pull LinkedIn HR contacts into a Markdown list) frame the cost: ~$0.10 per candidate to rent, versus a $1,5002,000 machine — though a home GPU can transcribe overnight for free at roughly 2× real-time on CPU.
> "It's not per hour of video. The server costs about a dollar an hour, but you only need it for ~6 minutes — so 10 cents."
## Conclusions
- The next webinar will be an overview of the levels of AI usage and how to use them; SPEAKER_01 will draft a plan SaturdaySunday, present it on a call, and there is **no deadline** — SPEAKER_00 was explicit about zero time pressure.
- The team's two recruiting problems are "banal" and easily solvable: a candidate knowledge base plus semantic search, fed by recorded-interview transcripts and standardized profiles.
- Recommended stack: Claude (Claude Cowork for non-programmers), with CLAUDE.md + skills as the realistic ceiling; RAG only matters at corporate scale. He prefers Claude over GPT/Google for professional work; cheaper models (e.g. Qwen) do the same tasks "just worse."
- The immediate good-practice change regardless of tooling: **start recording candidate interviews now**, one file per candidate, named by candidate.
- Renting GPU (~$0.10/candidate) is cheaper than buying hardware, but a home GPU works overnight for free if you already have one.
- A paid build of the HR system is on the table — priced per project, roughly "$5 to $100," with more scoping needed; possibly delivered as a shared monthly service on his subscription.
## Takeaways
- Solve a task with Claude first; only then have it create a skill from the finished workflow. Don't build skills speculatively.
- Turn any correction loop longer than ~3 messages into a skill.
- Use CLAUDE.md for persistent per-project instructions, and add "always keep CLAUDE.md up to date" to make it self-maintaining memory.
- Constrain prompts — your job as a user is to "narrow the variability of interpretation." Vague asks ("make me a dashboard") invite unwanted "fantasy."
- Learn from docs.claude.com; there's no single good course yet. SPEAKER_01 will cover this in the webinar.
- Memorable framing to reuse: "Split yourself into two people — one slightly smarter than you, your manager; one slightly dumber, your subordinate. You become the smarter one; the AI is the dumber one."
- Company-scale idea worth flagging to leadership: a lightweight AI "assistant" that chats with employees and forms an information core — automating most of the PM/status-reporting role.
## Open questions
- Exact scope, deliverables, and price of the HR system — SPEAKER_01 needs more task definition before quoting.
- Webinar date and title — SPEAKER_01 to propose; possibly with pre-recorded segments so nobody waits on live runs.
- Rent GPU vs. buy a machine with a GPU — left as "something to think about," possibly folded into SPEAKER_00's next hardware purchase.
- Larisa's Claude memory complaint — fixable (built-in memory, "remember this," CLAUDE.md), but her specific BA/PM task was never detailed.
- Can it pull a recruiter's full LinkedIn history at scale? No — LinkedIn bans bots; only a shallow pass (~20 contacts at a time) is safe.