# SIG-P4B 2026-09-21 — Session Notes

Participants:
  - rafa (UTC+1)
  - Rhea Myers | UTC-8
  - Sachin
  - PAtwater | UTC-8 | 31P

---

## 📖 The Reading
Not clearly identified in the discussion. This transcript is from a breakout session focused on a hands-on project idea (building an AI "factory" for job searching) rather than a discussion of a specific shared text. There are passing references to a main "Kafka" room, but no reading is named.

## 🧭 Overview
This breakout conversation was a working brainstorm about designing an AI-powered pipeline ("factory") to assist with job searching. Participants shared their personal job-hunting strategies and mapped them onto a possible automated workflow: gathering context (CV, LinkedIn, contacts), searching/reducing the field of jobs, tailoring CVs and outreach, and identifying the highest-leverage next actions. The tone was practical and speculative, ending with the group agreeing this could be a good project to "one-shot" before returning to the main session.

## 💡 Key Points & Themes

- **rafa — closed-loop job search design:** Inputs include your CV, target job type, and your network (exporting LinkedIn/email contacts to see where people work, industries, and roles). Emphasized finding "weak network ties." It's a closed loop: once you target open jobs/companies in your network, you go back and update your CV — and possibly even do a project over the next month to strengthen your portfolio for a target role.
- **rafa — avoid over-optimizing / define the MVP:** A common pitfall in AI projects is over-engineering. The MVP is essentially what "Launch series" already does (auto-tailoring a CV); improvements (finding better-suited companies, checking what past coworkers did) come as later iterations.
- **rafa — archetype-based approach:** Rather than going job-by-job, an LLM could surface ~4 job "archetypes" most related to your network and CV, pull benchmarks (example pay, example companies), and let the human choose a direction. Then it could recommend actions per archetype: auto-apply, reach out to specific people, or flag where you're under-qualified.
- **rafa — outreach tactic (from an Instagram video he made):** Search jobs, have the AI find the hiring manager, read their last three posts, and draft a personalized message based on those posts — a way to get outside the job portal.
- **Rhea — differentiation strategy:** Starts by "panicking," then calms down and looks for a "statistically improbable phrase" in her resume — unusual skills rather than common titles (PHP developer, web designer). She projects where she differs from the field to find jobs she's uniquely qualified for, then rewrites her CV to lead with what each specific job wants.
- **Rhea — a three-stage mental model:** (1) research/mapping stage (gathering details), (2) reduce stage (narrowing to the jobs you actually want — she mentioned ~200), (3) reformat CV. She framed this as expressing a personal brand / unique selling point to stand out — both to a human you hope to meet and to the LLM that screens applications first.
- **Rhea — anecdote on framing:** Years ago she had a job porting games to the Macintosh; after losing it, she applied for games programming jobs and got nothing — but once she reframed her CV around "Macintosh porting," she got many offers. Same experience, different emphasis matched to active areas of the market.
- **Sachin — using Claude as a coach:** Described using Claude to identify her profile and possible job titles when she didn't know what to apply for, feeding in links to platforms (Indeed, PolyJobs, others) and getting a list of possible roles based on skills/CV. Asked whether applications could also be automated. Noted she'd write three different CVs (web developer, architect, designer) for the same person, highlighting different qualities in each.
- **rafa — keep the story simple:** Beyond filtering CV content, the factory should help articulate the simplest version of your personal narrative. He gave his own example: moving from HR → strategic consulting → product advisory (semi-technical) → progressively more technical crypto work, with the CV telling that coherent story.
- **rafa — his ideal one-shot design:** (a) gather all context with minimal effort (log into LinkedIn, do the searching), and (b) tell him the single highest-leverage activity to do right now to increase his odds. Example: if interested in robotics product management and you've never built a robot, the factory should recommend building one and writing a speculative product design to add to your CV.

## 🔀 Questions & Disagreements

- **Not a sharp disagreement, but a contrast in approach:** rafa's design leans toward mass-applying (at least initially) and automation, whereas Rhea explicitly does *not* mass-apply — she's careful and selective. rafa acknowledged this difference and suggested a "human check" within the protocol to reconcile the two.
- **Open technical question (Sachin):** Can applications be submitted automatically with Claude, not just the search/matching? Left partly open; rafa cautioned he "wouldn't apply for the same job" the same way and steered toward tailored CVs.
- **Data access question (rafa):** How much data can actually be pulled from LinkedIn? Noted LinkedIn may not have an API, but browser-control tools could work. Playwright was raised for scraping, with uncertainty ("don't know if it will work well for this").
- **AI-detection risk (rafa):** A concern that an AI-generated CV/message could be filtered out by AI-detection tools (mentioned Pangram) — you need to be careful the "factory" doesn't produce output that gets you screened out. Related worry that the LLM reading your application first will throw out "boring" responses.

## 🔗 References Mentioned

- **Claude** — used as a coaching/profile-identification tool (Sachin).
- **"Launch series" / Claude with browser control (cloud.co / "cloudcloud.co.list")** — mentioned as auto-tailoring CVs and controlling Chrome for clicking through/scraping LinkedIn. (Exact product name unclear in the transcript.)
- **Playwright** — plugin/tool for scraping data (rafa).
- **Pangram** — an AI-detection tool that could flag AI-generated applications.
- **Job platforms:** LinkedIn, Indeed, "PolyJobs" (spelling uncertain).
- **rafa's own Instagram video** on the hiring-manager outreach tactic.
- **"Kafka" room** — the main session room the group returned to (likely the session name/topic).

## ✅ Action Items & Next Time

- The group agreed the job-search "factory" is a good candidate to **"one-shot"** (build as a quick prototype), but no owner or deadline was assigned.
- Participants returned to the **main ("Kafka") room** at the end of the breakout (per Sachin and Rhea).
- No specific next reading was named in this segment.

## ⭐ Memorable Quotes

- **Rhea Myers:** "When I'm searching for a job, I first panic... I then calmed down. I always look for it's almost a statistically improbable phrase that's in my resume."
- **Rhea Myers:** "As soon as I changed my CV to Macintosh porting, I got lots of job offers."
- **rafa:** "You should go build a robot and write a... speculative product design for it and add that to your CV."

*Note: This was an auto-generated, overlapping transcript, so some attributions (especially in the tightly interleaved rafa/Rhea exchange) and product names are approximate.*