Nat launches a preschool app and quickly learns what a test run can miss. A plan that made sense for three or four people looks different when ten show up. Instead of trying to fix every problem by hand, she thinks about what would help someone report a problem well and what the app might be able to fix with the right guardrails.

Lucas has the opposite problem. He’s built so many safeguards into a project that a small change can take hours to get moving. That sends us back to a question we keep running into when we work with AI. What’s the smallest thing worth putting in someone’s hands, and what do you learn only after they use it?

Then Kelly’s garage door stops coming down. The repair tech leaves a note inside the mechanism for the next person. A conversation about context engineering turns into one about what experience feels like, how hard it is to pass along, and why it matters whether the person reviewing an AI’s work knows what to look for.

Recorded September 3, 2026, with Kelly Hodsdon, Nat Walstead, and Lucas Whelan. The conversation reflects the tools and releases the group was discussing that week.

Follow Less Miserable Tech Backchannel at https://lessmiserabletech.com/backchannel/.

In this conversation

• 00:00 Kelly on why the last mile is the hardest
• 00:23 What we’ve kept up with this week
• 01:18 A preschool app meets its first real users
• 03:23 Giving a teacher a way to request changes
• 05:11 Finding the smallest thing worth releasing
• 07:06 Using screenshots to make feedback actionable
• 11:04 When safeguards slow a small change
• 15:04 The minimum valuable product and the empty prompt box
• 17:10 Trying new models and keeping a setup that works
• 23:48 Context engineering and what domain experts bring
• 29:36 Asking a client what a good outcome looks like
• 32:29 Surveys, tasks, and feedback people can give
• 41:12 A garage door repair and what the technician noticed
• 45:55 A car belt, Blink, and knowledge that goes undocumented
• 48:27 What self-healing software could do with human oversight
• 53:04 What we get from talking it through together

A few things we kept coming back to

• Releasing something useful gives you feedback you can’t get by testing it alone.
• A screenshot can make a bug report concrete. A short task can tell you more about a new feature than asking whether people liked it.
• Someone who has done the work can often tell that a result is off before they can explain why.
• An agent can turn incoming feedback into a punch list. A person still decides what needs more detail, what to fix, and what can wait.

Music by Michael Ramir C. “Gimme that Groove!” is available through Mixkit.

Watch the episode · Listen on Podpage

“Please don’t break it any more than it is. It was one bug.”

Kelly wants a coding agent to fix his PDF splitter. The agent has bigger plans. Nat knows the feeling: she’s just talked Codex through several projects, only to discover it’s been working without the project folders she thought it had. A lot of work got done. Now she has to figure out how much of it to undo.

Lucas’s dashboard works, but the AWS bill sends him back through decisions the AI had made sound perfectly reasonable. Getting an explanation he can actually follow takes more persistence than getting the thing built.

That’s some of the less glamorous work behind using these tools. They can do more for us, faster, but we still have to recognize when they’ve misunderstood the assignment. Lucas puts it this way: “AI doesn’t equal everybody out, it actually magnifies our differences.”

The same question comes up when we talk about charging for work that used to take much longer. If experience lets you finish a job in two hours instead of twenty, has the work become less valuable? What about knowing which questions to ask, which mistakes matter, and whether the result is ready for a client?

There are things worth stealing here, too. Nat shows how Clive asks enough follow-up questions to turn “I want to eat healthier” into something you could actually plan and shop for. And after a week when everything seemed to break, she gets to watch families use her preschool app at orientation.

We compare notes, give each other a little grief, and leave with things to try. It helps to talk with people who understand why you love these tools and why you occasionally want to yell at them.

Recorded August 27, 2026, with Kelly Hodsdon, Nat Walstead, and Lucas Whelan.

In this conversation

00:00 Lucas on how AI magnifies differences

02:08 The Thursday Backchannel, Tech HIVE, and how we use our tools

05:12 A voice-driven Codex session loses its project context

09:04 When an AI assistant goes from novelty to daily habit

12:03 Faster client work and what expertise is worth

17:51 Fixing one PDF splitter bug without rewriting the app

21:07 How leverage amplifies mistakes and good judgment

25:38 Staying the expert and asking for an explanation you can follow

28:01 The AWS bill for a function nobody used

32:25 A tape measure, a pencil, and knowing which errors matter

34:55 Clive in the Hive turns a vague goal into a usable brief

37:53 Thinking out loud with AI and collecting Field Notes

40:42 Agent personalities and what happens when agents collaborate

47:00 The preschool app makes it to its first orientation

49:42 A genomic research challenge and what makes people want to join in

53:02 Sharing the puzzle, the discoveries, and the prize

A few things we kept coming back to

• An agent can sound like it understands you without having the project context you assumed it had.
• A working app can still contain decisions you wouldn’t have agreed to if you’d understood them.
• Doing the work faster doesn’t eliminate the experience needed to judge it, or settle what that experience is worth.
• A specific frustration gives you somewhere to start. “Eat healthier” needs a few more questions before anyone can help.
• Sharing an unfinished problem can be useful. Someone else may have the piece you’re missing.

Music by Michael Ramir C. “Gimme that Groove!” is available through Mixkit.

Nat walks us through the system she uses to organize projects, turn meeting notes into next steps, and keep track of what her agents have done. We compare notes on coding tools, subscriptions that don't actually save us money, and why capturing a task isn't much help if we never look at it again. Lucas describes a legal-document workbench where the user can request an evidence packet without writing a prompt.

It's a conversation about the things we're building, the parts that still frustrate us, and the small changes worth trying next. The common thread: the technology should fit the way people work, not give them another system to babysit.

Recorded August 13, 2026, with Kelly Hodsdon, Nat Walstead, and Lucas Whelan. Tool names, prices, and model comparisons reflect the conversation at the time of recording.

In this conversation

00:00 Opening

00:39 Why the AI news brief keeps missing things

05:08 Coding agents, rabbit holes, and staying in the loop

06:47 More coding tools, more subscriptions

11:08 The PDF prototype and the hard last mile

12:51 Capturing tasks and bringing them back

15:17 Nat's project system and agent instructions

18:56 POP: Progress, Obstacles, and Priorities

21:54 Trello, daily sweeps, and agent receipts

25:34 Building for people who won't use your system

26:54 A tour of the TechHive

31:27 An always-on personal assistant, and who maintains it

34:18 Meet people where they are and hide the plumbing

39:18 A legal-document workbench without a prompt box

41:56 Goodbyes

Ideas worth keeping

• Fix the process behind an AI's recurring mistakes.
• Compare the cost of the whole workflow, including extra subscriptions and learning time.
• Plan for integration and everyday use after the first working demo.
• Capture needs a review habit that brings information back when it matters.
• Keep meetings connected to project work. Nat's POP structure means Progress, Obstacles, and Priorities: what changed, what's blocked, and who does what before the next meeting.
• Give unattended agent work clear permission boundaries and a linked record of what changed.
• Meet people in the channels they already use. Hide the plumbing behind the task they are trying to complete.
• Combine fixed rules, AI interpretation, and human review.

Music

Music: "Gimme that Groove!" by Michael Ramir C., via Mixkit.