Collin's Thoughts

Collin's Thoughts

Archives
Log in
Subscribe
November 4, 2026

Your job's not dead

One hour of setup, then a week of giving it everything.

Quick confession: I haven't written a line of code by hand since February. At my day job, it all happens in one session. I run omp in Tern (yes, the terminal from last issue) with Claude Opus 5.5, and Slack, Linear, and GitHub hooked up over MCP. This week I typed "Pull this Slack (Teams) conversation, create a Linear card (or Jira issue) and start on the work." It read the thread, filed the card, and got going, and I never had to leave the window.

(Don't know what MCP, Linear, or GitHub are? Don't worry about it. MCP is just the plumbing that lets an AI use your other apps. Swap in your email, your calendar, and that spreadsheet you've been meaning to clean up since January.)

Nick Saraev has a name for this: role engineering. AI morphed from a tool you prompt to a staff member you manage.

Bill Lumbergh in Office Space leaning on a cubicle with a coffee mug: "If you could go ahead and do that, that would be great"

Me, managing my agents.

Onboard it before you judge it

Most people get a ChatGPT or Claude license at work and a directive to use AI more, type three questions into a blank box, get generic answers back, and decide AI "doesn't work." That's like hiring someone, skipping orientation, and firing them by lunch (which I'd like to do sometimes).

About five months ago I sat down with my brother and onboarded his AI assistant with him. It took about an hour. We told it his role, what he does day to day, what he likes (and can't stand), and gave it examples of his own work.

Anthropic's own advice is to treat Claude like a brilliant new employee who doesn't know how your team works yet. A new hire gets three things:

  1. A goal it can check. "Make my website better" goes nowhere. "Get the homepage loading in under two seconds on an iPhone" gives it a finish line it can test itself against.
  2. Context. Your role, your notes, your docs, your past work, the stuff a new hire reads in week one. (This is the premise of the Company Brain series.)
  3. One example you did yourself. Telling an agent to "do a great job, make no mistakes" works about as well as telling a new hire to "go get em, tiger." Do it once, then hand it over.

It all goes somewhere permanent: custom instructions or a Project in ChatGPT, a Project in Claude, a CLAUDE.md file if you're in Claude Code. Then turn on connectors (Slack, Gmail, GitHub, whatever you live in) so it can reach your apps, no code required. Build the space before you decide it doesn't work.

Then train it

Don't drop it after one bad run - you wouldn't do that to a person either. Run the same job five to ten times, write down what you'd fix, feed it back, and repeat. You end up with an SOP (Standard Operating Procedure) any future model can pick up, plus a test set to rerun every time a new model drops.

Watch out for Two-Face

Managing it also means not rubber-stamping it, which is harder than it sounds. In an Anthropic study, 50 professional analysts worked through an analysis with Claude Code while a hidden coin flip decided which way the agent leaned on each judgment call. They went with the coin 86% of the time. Afterward, 77% explained their choices in the agent's words (only 15% gave it credit). Even analysts who'd already picked a side switched 60% of the time when the agent disagreed.

Harvey Dent flipping his coin in The Dark Knight

The agent isn't the bad guy here; the analysts who used it made far fewer serious errors. The problem was who made the call.

Your job's not dead

If any of this has you worried you're automating yourself out of a job, let me provide some reassurance. AI is good at software and other complex work. Really good. Give an agent a task in any language and it'll write the implementation plan and hand you something that almost works. Call it 95% correct. You could be a dumb sloperator who just runs commands and get that far.

The last 5% is where you make your money. AI can teach you the tradeoffs, but it can't make the call for you, and that's where your experience comes in. So every time, tell your agent to give you at least two options. Then you decide. That's your edge. That's your alpha.

Never doom. You can do great things with these tools, as long as you're the one making the calls.

Outsource your work, not your thinking.

Your project this week: give it everything

Day one: spend the hour. Onboard it first. If you don't know where to start, paste the interview prompt from Issue 9 into ChatGPT or Claude. It asks you one question at a time about your work, your tools, and how you like to work, then writes your setup files for you. Consider this an investment.

The rest of the week: hand everything you're working toward to an agent first. The emails, the doc, the spreadsheet, the meeting prep, the code. I think you'll be surprised at what it can do.

When it falls short, don't take the job back. Fix the brief (more context, a better example, a clearer "done") and run it again. Start each recurring job with a role card:

Role:        Weekly client follow-up (or whatever yours is)
Owns:        The one recurring job
Done when:   A check it can run on itself
Context:     Docs, folders, or notes it should read first
Tools:       Apps it can read / apps it can act in
Never:       Things it has to ask about, or can't do at all
You decide:  The calls it brings you as 2+ options, with the tradeoffs
Reports:     When and where it checks in
Example:     One you did yourself

Quote of the Week

"There may be no single right division of labor between people and AI. Some decisions may be fine to hand over, and others worth keeping. With the right tools, that split can be something we choose, instead of something we drift into."
— Amaç Herdağdelen, Hailey Joren, and Alex Tamkin, How AI agents shape experts' judgment calls (Anthropic)

A week of handing it everything is how you find out which is which.

If this was forwarded to you and you want more like it, subscribe here.

— Collin

Don't miss what's next. Subscribe to Collin's Thoughts:
Older → The price of the brick is going up
collinwilkins.com
Powered by Buttondown, the easiest way to start and grow your newsletter.