explain
The one he "probably couldn't do without." One sentence does the work: talk to me like a colleague who knows the project but wants to know what you just did. Referenced in every other skill he has.
Show Us Your Agent Skills / EP 01 / guest dossier
JEREMIAH LOWIN
PREFECT · FASTMCP
SECOND BRAIN
MOSTLY HARNESS
JEREMIAH LOWIN
Jeremiah is into personal software "in the most true sense." His harness is built for exactly one user: memory he can muck with, a morning voice memo as the day's system prompt, talk software in his own vocabulary, and skills that pin down what "ship it" means.
this episode is also a blog post: The Agentic Software Factory and a podcast: Agentic Engineering and the Lost Art of Verification
EP 01 · JEREMIAH LOWIN · the second brain, live on stream
PERSONAL SOFTWARE
the second brain, captured in the second-brain write-up. fork it
"I'm into personal software basically in the most true sense." Every layer of Jeremiah's harness is shaped to him: OpenClaw because he can "go muck around with its memory in a way that works for me," skills that encode his tone and his verbs, and tools like Cardboard and Prefab built because nothing off the shelf makes talks or dashboards the way he wants them.
The memory is the core of it: "I use it as a second brain, and so I have big expectations about the information that I put in at a moment coming back out later."
The second-brain write-up captures the loop: daily voice memos feed the memory, ideas land the moment they occur, and months of trickled context sit behind every real question. Cardboard, his custom talk software, rides the same memory: the agent writes through an API and MCP server while the UI stays read-only.
Inside the write-up: four principles, what a session looks like, anti-patterns, a "what you need" list, and the personal-software philosophy underneath it all.
"This has been really fun to have a piece of software that exactly makes talks the way I want them, the way I like to give them."
Cardboard exists for one user. So does the rest of the harness: "the way people pile on functionality, features, customize it, skills, is deeply personal." 01:01:50
POUR IN, WORK IT OUT
the second-brain loop. every timestamp opens the segment
Record the morning voice memoAlmost every day, up to 30 commute minutes: what he's thinking about, what he wants to do. "That's my system prompt for the day." 00:37:01
Capture ideas the moment they occur"I'm walking down the street and I'm like, 'Oh, here's an idea. Oh, we should name it that.'" Notes go in immediately or at the morning ritual. 00:37:42
Let memory carry the stateFeed in a PyData London talk idea, leave it, talk about other things for weeks, come back. The substrate holds. 01:02:42
Split the stack by jobOpenClaw is "my memory absorber for more asynchronous work." Claude Desktop and Codex Desktop handle parallel coding sessions. 01:03:19
Agents draft, Jeremiah decidesAgents write his code and review others' code on FastMCP, "but ultimately, I will step in, and I will look at it, and I will work on it." 00:45:30
Edit skills when behavior missesSkills are living documents: "Most of these skills are changing as I'm like, 'Oh, this didn't work.'" 00:54:02
"The way agents review code, for whatever reason, seems to have a bias that the PR should be accepted."
Exactly the wrong default for a framework maintainer. His rule, from An Open-Source Maintainer's Guide to Saying No: agents draft, the human says no. 00:39:42
POLITE NOTES TO THE AGENT
his skills, shipped to the companion repo. install them, then make them yours: they're living documents
github-reply
Encodes how to treat contributors: "don't say, 'Great work,' followed by a rejection." There's a right way to treat people, and the LLM doesn't do it by default.
ship-it
"It means open a PR. It does not mean merge the code." A bridge between his words and the outcome he wants, because models read the phrase differently.
a skill for writing skills
"This is how we make living documents." Assembled from things he found online, used to keep every other skill honest as the work exposes failures.
"Skills are awesome ways to steer behavior. They go into the agent's brain in the exact same way that a message from you does."
Progressive disclosure is "the magic of skills": the description is always visible, the body loads only when the agent invokes it. 00:48:57
THE STACK
fastmcp, prefab & cardboard are his own builds