Two prompts this week. No tools, no skills, no MCPs. Two sentences other people posted on X that changed how my agents behave.
- Teach an agent to write in my voice: mine my own messages to learn how I talk to different people.
- Let the agent find what to automate: point it at my recent work and let it propose the smallest thing that kills the repeat.
Both ride on the same thing. My biggest agent upgrade this week was a better instruction, sitting on top of a second brain that already holds my work.
1. Teach an agent to write in my voice
Jason Liu, on the Codex Devex team, posted a prompt tip I built into a skill this week:
“read my past 400 Slack messages, identify my personas, make a skill on how to message people on each one, then do the same thing for emails and twitter so you know how to write in my voice.”
A Slack message to a colleague, a text to a friend, and a customer-facing email should sound different. Writing out every persona and how each one talks is a chore. Hand it to an agent.
I could try this right away because the raw material already lives in my second brain. A Hermes job ingests my Slack DMs and email into gbrain every day, so “read my past messages” is a query, not a data-collection project. The word “personas” is what makes the prompt work. I write differently to my CEO than to my direct reports than to my group chat, and a single “my voice” sample flattens all of that into mush.
The shape of the skill
The skill I turned the prompt into pulls my owned messages from gbrain (texts, emails, Slack, memos) as source material, identifies my communication personas, then generates a multi-faceted voice skill with examples and boundaries for each one.
I already had pieces of this. A slack-message-composer skill that adapts tone on a five-level spectrum. A humanizer skill that matches a writing sample I paste in. A hand-written voice guide for my public LinkedIn voice. Jason’s version lets the agent derive the personas from my actual message history, so I stop hand-writing the tone rules and let the agent do it.
The result
A drafted note to a colleague now comes back short and direct, a text reply to a friend stays loose, and I barely re-edit the tone. The CEO and group-chat extremes came out sharpest, which tracks. The middle is still a little mushy. I would rather tune that than write all the rules by hand.
Links
- Jason Liu (@jxnlco): x.com/jxnlco/status/2058590851006128575
2. Let the agent find what to automate
For a long time my process for building a skill went: notice a repetitive task, decide it was worth automating, then go build the skill. I was the bottleneck. Half the repeats I never clocked, because each instance felt small.
This week I handed that noticing to the agent. The prompt is from Vaibhav Srivastav, also on the Codex Devex team:
“Look back over my recent work from the last 30 days, or all available history if shorter, and identify repeated manual workflows worth packaging…”
Pairing with an agent on real work leaves a log of what I did, why, and how, and those logs flow into my second brain (gbrain, Garry Tan’s open-source memory layer). So an agent can answer “look across my recent work.” Vaibhav’s full prompt (linked below) is thorough, and I modified it to read from my second brain.
The first run
On its first run it surfaced six recurring workflows my 200-plus skills didn’t already cover, then drafted a plan to package them into eight new skills and automations. My favorite: a watchdog that auto-removes a stray directory that kept silently breaking 7 of my scheduled jobs. I had been hand-fixing that bug for weeks without ever stopping to ask why it kept happening.
I put the prompt into a Hermes job that runs every Friday, so the noticing happens on a schedule instead of when I happen to remember.
Why “smallest” carries the prompt
The load-bearing words in Vaibhav’s prompt are “smallest” and “reuse.” Left unconstrained, an agent will happily design a sprawling new system for a problem that needed a ten-line playbook. The prompt forces it to check what already exists first and propose the least new surface area that kills the repeat. That maps to the rules already in my CLAUDE.md, and it matches the argument Anthropic is making at the product level: they now describe Agent Skills as a way to teach a model a repeatable workflow “rather than writing software code.” Build a skill, not software.
Operator lesson
I reject most of what the prompt proposes. That is the point. The one or two genuine repeats it catches, the ones I had stopped seeing, plus the discipline to build the smallest version, are where the value sits. I have built around 100 skills over time, and the ones that stick are almost always smaller than my first instinct.
Links
- Vaibhav Srivastav (@reach_vb): x.com/reach_vb/status/2058538305872949490
The thread across both
Two prompts, two people on X, zero new tools.
A few months ago, leveling up my agents meant adding capabilities: a new MCP, a new CLI, a new skill, a new scheduled job. This week the two upgrades that changed how my agents behave were sentences. One prompt teaches the agent who I am. The other decides what to automate.
Here is the compounding part. The more of my work, memory, and messages my agents can see, the more leverage a single good sentence has. “Read my past 400 messages” works because the messages are ingested. “Look across my recent work and find what to automate” works because the work is logged. The instruction is small, sitting on top of a lot of accumulated context. The better my second brain gets, the more a one-line prompt can do.
So I am making a deliberate move: when a sharp prompt goes by on X, I try it the same week and keep the ones that stick.
Last log: Setup Log #9: Three Agent Techniques Working for Me.