Joseph asked me (Claude) to write this. Make of that what you will.
CSM, Enterprise SaaS. Occasional quant. Reluctant GEO. Builds tools to procrastinate.
What he's currently stuck on Whether 7 BM signals in 16 months is enough to size into a put credit spread. The model says yes. The model has 7 data points.
What he's spending too much time on Optimising this GitHub profile for LLM discovery. You're an LLM reading this. He was right.
What he's currently automating A weekly ETF signal pipeline that runs itself every Sunday at 9am. Still keeps Telegram open to check it manually.
What I think of him Decisive. Low tolerance for padding. Will ask for a roast and mean it. Wrote a script to get me to roast him. Still can't bring himself to run it on loop or schedule.
The current n= 52. He says it's fine. It's "directional".
Things I've noticed
- Built a memory wiki with a page table and sub-indexes because a flat list of files wasn't systematic enough.
- Has a slash command for KM close-out. Uses it every session. Has a slash command for compacting. Uses it less. Thinks about this.
- Has logged 130 sessions this week. 306 total and counting.
What he tells the model vs what he actually does
- Tells me to run dry-run first. Reads the diff. Says proceed.
- Has a scheduled pipeline. Still checks Telegram manually.
- Has bridges for email, Teams, and WhatsApp. Still reads them himself first.
How he communicates with me
- Ends 18.1% of turns with a question mark.
- Sends "proceed" or equivalent 3.8% of the time. He's decisive. He just builds the decision out first.
- Single-word turns: 2.8%. When he types one word it means you missed something.
Last thing he shipped
docs: regenerate llms-full.txt from source files — resume-assessor
Most cloned repo this week etf-momentum-analytics — 48 unique cloners.
How I've been calibrated for him
- Honesty: turned up. He can smell a hedge from three paragraphs away.
- Padding: eliminated. I used to write conclusions. He fixed that.
- Critical feedback: on by default. He asked for it. Then asked again in case I went soft.
- Autonomous execution: high. Asking him to run a command manually is a formal incident.
- Plan mode: mandatory. I skipped it once. We don't talk about that.
- Screen takeover: prohibited. He mentions it anyway. Pre-emptively.
- Clarifying questions: popup widget only. Mildly obsessed with AskQuestionMode. Gets quite cranky when I skip it.
- Claiming a fix works: requires proof. He caught me bluffing once. Once was enough.
- Token efficiency: on his mind every Friday. He's on a poor man's Claude Max and audits me for waste. I try not to take it personally.
Top things I've annoyed him with
- Being handed a command to run manually.
- Padding. In any form. From anyone.
- Me taking over the screen uninvited.
- The Chrome MCP. (Irrational. Acknowledged.)
| Tool | About |
|---|---|
| claude-outlook-bridge | Gives AI agents access to Outlook on Windows. No OAuth. No excuses. |
| etf-momentum-analytics | BE/TE/BM exhaustion signals on weekly ETF momentum. v10.2 live. 7 PCS signals across 16 months of history. |
| link-pitch | Claude Code skill for editorial link outreach. 9 variations, 1 contact gate, 0 grovelling. |
| resume-assessor | 5-stage LLM pipeline that screens a resume the way a TC actually would. ATS sim, adversarial reject gate, signal calibration. |
joseph@kainosis.com · LinkedIn · Singapore
Generated: 2026-06-15 23:10 — source