Jess Martin
banner
jessmart.in
Jess Martin
@jessmart.in
1.7K followers 350 following 640 posts
🛠 Tools for a better tomorrow 🧠 Architect-philosopher 💼 Building https://worksquared.ai 🏠 https://jessmart.in 📍Fuquay-Varina, NC
Posts Media Videos Starter Packs
90% of the time when an LLM fails to do a thing for you, it's a lack of proper context, not model capability.
Reposted by Jess Martin
rooms are where life happens. we create, workshop, collab, party, veg out, bring up our kids... in rooms

and we do it together

why would I want a computer strapped to my face? instead: a computer we can walk into... and walk out of

I am all-in on augmented environments.
if you do this, please publish it somewhere! would v much appreciate it.
one of the few: gpt-5-pro and GLM 4.5 have 'em too
first pelican with speed lines?
- daily / weekly / monthly retrospectives - I can just talk and it turns it into a structured format
- decision log - extract from my notes any decision I've made recently, add to a decision log file, set a date to check in
- also give it access to my github via `gh` so it creates issues for me
After talking to Claude for 20 minutes planning out work for today:

> Priority order? This is a lot for one day. If you had to rank the top 3 must-dos, what would they be?

🥺 Yep, you're so right Claude.
Why does Claude Code operating directly on files in my Obsidian vault feel so much better / different than Claude Desktop + an Obsidian MCP server?
For sure! Reach out!
Yes! I already implemented something I've wanted for *years*: a decision log with a "review date" so that I can check in regularly.

And it was ~10 minutes.

I'm really excited to see where this all leads.
I finally did the thing: pointed Claude Code at my Obsidian Vault's markdown files.

Oh. My. Goodness. This is great!
I think this succinctly explains the rapid adoption of LLMs in a variety of domains.

There's something tipping point for accuracy where the generality wins become more important than 100% deterministic predictability.
My hunch is that humans will be able to accept a surprising amount of nondeterminism once error rates reach acceptable levels. We accept a lot of nondeterminism from other humans, and understand that 100% reliability comes at the cost of flexibility.
I'm finding it's more valuable lately to spend my time on "build the software factory" than to actually build the software directly.

Improving the quality of work and parallelization across my AI agents is a huge force multiplier.
...starting in 20 minutes!
worktrees are great, but I got tired of manually setting them up.

conductor.build gives me a GUI for managing 3-4 concurrent worktrees, while still having full file-system + IDE access.
Conductor
Run a bunch of Claude Codes in Parallel. By Melty Labs.
conductor.build
Good times! Cool to see Claude Code one-shot a @livestore.dev Kanban app.

I have a pretty good idea of the overall architecture of Claude Agents SDK, and I'll be able to get into the meat of it during the next livestream.
🎥 Starting now!

I'm livestreaming building a LiveStore app that uses the Claude Agents SDK as the "agentic loop" and interacting with the LiveStore app through custom tools.

Should be fun! Come hang out.

www.youtube.com/live/5-1TF51...
Coding With Jess: LiveStore and the Claude Agents SDK
YouTube video by Coding with Jess
www.youtube.com
Probably gonna livestream building a LiveStore app with the Claude Agents SDK.
"active" in the same sense Times Square is active