Better with Kent

Kent C. Dodds

Solo episodes from Kent C. Dodds on durable skills for people who ship software: judgment, accountability, problem clarity, and what stays valuable as AI takes on more implementation. Kent teaches directly on camera — no guest, one idea at a time. Complements guest interviews on Become a Product Engineer (Chats with Kent). Episode 1 adapts The Last Software Engineer; later episodes cover traps like building the wrong thing faster and skills like making user pain visible. Subscribe for evergreen episodes as the series grows.

  1. 2d ago ·  Video

    What I Changed to Beat a Million-Install AI Skill

    Matt Pocock's triage skill has been installed over a million times. I moved the repeatable parts into deterministic software, kept the judgment in the skill, and it got faster, cheaper, and more accurate in every mode. Here's how to build that hybrid. (00:00) - A million-install skill, made better (00:26) - Every skill has the same problem (00:47) - From manual process to software (02:08) - Turning the triage skill into software (04:07) - Devin's benchmark: Lite, Normal, Ultra (04:46) - Most of the skill doesn't need a model (05:24) - Faster, cheaper, more accurate (05:59) - The hybrid skill (06:21) - Taking it further with Kody (06:46) - My ship-pr skill: one command on repeat (07:46) - A cheap decision model (08:26) - Waiting and merging safely (09:24) - Just-in-time context Better with Kent - durable skills for people who ship software. Thank you to Sentry for sponsoring this episode: https://kcd.link/sentry Skills are awesome. You hand a manual process to an agent and it just handles it. But agents are slow and expensive, and they're perfectly happy to spend all your tokens on work that should have been software. We used to turn processes we understood into software. With skills, we stopped doing that. So I took Matt Pocock's triage skill and had Devin benchmark it against a hybrid version: gathering, paginating, sorting, applying, and verifying moved into a deterministic script, and only the judgment stayed with the model. The hybrid was faster in every mode (Lite, Normal, and Ultra), cost the same or less, and was more accurate every time. Then I show how I take it further with Kody (shared credentials, reusable code, and triggers) using my ship-pr skill: one tiny command on repeat, software handling the process, a cheap decision model for the small calls, and the expensive coding agent saving its effort for real judgment. What are your agents doing that should be software? Tell me in the comments. Links Sentry (sponsor)Matt Pocock's skillsKodyBetter with Kent

    What I Changed to Beat a Million-Install AI Skill
  2. Nov 10 ·  Video

    Build What They Can't

    Big companies can copy your features in an afternoon. Here's how to put your product in a safe harbor they won't follow you into: trust, incentives, community, proprietary data, audience, domain depth, and the right niche. (00:00) - Big players keep shipping my features (00:32) - Sherlocked: Watson and Raycast (01:18) - Lesson one: find a safe harbor (01:33) - 1Password: the trust safe harbor (04:35) - Kody: what the labs won't build (06:49) - Dropbox: a feature, not a product (08:34) - Community: Raycast's extensions (10:02) - Proprietary data: Waze (11:36) - Audience: Nathan Barry and Kit (12:05) - Domain depth: Abridge (13:14) - Niches too small for giants (13:57) - Safe harbor or empty harbor? (15:17) - Watch out for invisible software Better with Kent - durable skills for people who ship software. Thank you to Sentry for sponsoring this episode: https://kcd.link/sentry Every day I open X and see another big player ship a feature I already have. If your product is just a pretty UI on top of a spreadsheet, you're in trouble. But giving up isn't the only option. The better move is to pivot into a safe harbor where the big guys will never go, and make yourself untouchable by design. I start with Watson, the app that gave us the word "Sherlocked," then walk through the safe harbors that have kept products alive next to giants: trust (1Password), incentives (Dropbox, and what I'm doing with Kody), community (Raycast), proprietary data (Waze), audience (Nathan Barry and Kit), deep domain expertise (Abridge), and niches too small for a giant to bother with. Then there's the boring version of untouchable: sometimes the giants leave you alone because nobody wants what you built. I share the test I use to tell a safe harbor from an empty one. Where is your product anchored? Tell me in the comments. Links Sentry (sponsor)KodyHow I'd Get Into Software in 2026The Future of Software is InvisibleBetter with Kent

    Build What They Can't
  3. Nov 3 ·  Video

    How I'd Get Into Software in 2026

    If I were starting from scratch in software engineering today, here are the two skills I'd work on right now (systems design and communicating intent), how I'd practice them, and the durable skill that stays valuable no matter how good agents get. (00:00) - Is coding over? (00:28) - Skill 1: understand and design systems (01:26) - Skill 2: communicate intent (03:16) - I can't promise these skills will last (05:30) - The adjacent possible (06:04) - The printing press and the iPhone (06:52) - Nobody planned DoorDash (07:30) - Git, GitHub, and agents as our printing press (08:07) - Antifragile (08:51) - No skill makes you antifragile (09:30) - Get close to real problems (10:35) - Skin in the game and trust (11:40) - Don't chase every shiny thing (12:20) - The durable skill (12:53) - What I'd do if I started today (14:00) - Terminator or Star Trek, this still works Better with Kent - durable skills for people who ship software. Thank you to WarpBuild for sponsoring this episode: https://kcd.link/warpbuild A lot of people are scared. Is it over? Are we just moving on from coding? Frankly, yes, coding is over. But I don't think that means we can just give up. If I were getting into software engineering today, I'd work on two skills right now. First, understanding and designing systems: what the pieces are, what each one is responsible for, and how they talk to each other. I'd practice by sketching out the apps I already use with an agent. Second, communicating intent: the what and the why, the constraints, and a way for the agent to verify its work, so it doesn't fill the gaps with guesses. But I can't promise those skills will still matter in three or six months. So the back half is about the skill that does: noticing the adjacent possible. That's Stuart Kauffman's idea (I learned it from Steven Johnson's Where Good Ideas Come From). The printing press, the iPhone, and Git all opened doors nobody planned for. Agents are our printing press, and the doors are flying open. Then I pair it with Nassim Nicholas Taleb's antifragile. No skill is going to make you antifragile. What does is getting close to real problems felt by real people, owning outcomes (skin in the game), building trust, and being intentional about which shiny things you chase. What would you do if you were starting from scratch today? Tell me in the comments. Links WarpBuild (sponsor)Where Good Ideas Come From by Steven JohnsonStuart Kauffman (the adjacent possible)Antifragile by Nassim Nicholas TalebBetter with Kent

    How I'd Get Into Software in 2026
  4. Oct 27 ·  Video

    My AI Employees Won't Let Me Sleep

    I gave my AI agents everything they need to work without me, and now they never sleep, never eat, and never run out of questions for me. Here's what that's costing me (sleep, meals, and time with my kids) and the quiet hours fix I built for my Grok Bots on camera. (00:00) - They never sleep (00:23) - I removed every reason to wait on me (01:00) - My huge mistake: the software factory (01:56) - I turned into a boss of agents (02:57) - The only limit left is me (04:45) - What happened at the theme park (05:24) - No sleep, no lunch, no exercise (06:44) - "I'm going to sleep" (then 3 more messages) (06:59) - Holding myself accountable (07:39) - What's on my plate (08:41) - The fix: quiet hours for my agents (10:05) - How do you make yourself stop? Better with Kent - durable skills for people who ship software. Thank you to PowerSync for sponsoring this episode: https://kcd.link/powersync I can get a month's worth of work done by answering a couple of questions before bed. Cloud agents keep running with my laptop closed, I check on them from my phone, and Kody gives them the connections to get context and take action on their own. All that's left for me is the decision making. That sounds amazing until you realize the questions never end. A boss knows they can't text their employees at 2 in the morning and expect an answer. My agents don't have that limit. They answer in seconds, every time. So the only limit left is me. In this episode I get honest about what that's costing me: sending messages to my agents while I was in line at a theme park with my kids, skipping meals, skipping workouts, and telling an agent "I'm going to sleep" at 1:17 in the morning, then sending it three more messages within 92 seconds. Willpower is clearly not going to fix this. So I built quiet hours for my Grok Bots, right on camera. From 10 PM to 8 AM they keep working, but they only interrupt me for something truly urgent (like production being down), and I get one summary in the morning. If you've got agents that never sleep, how do you make yourself stop? Tell me in the comments. Links PowerSync (sponsor)How to Put Your Codebase on AutopilotHow to Increase Your Leverage with Agents (the theme park video)KodyMEGA (mega.dev)Better with Kent

    My AI Employees Won't Let Me Sleep
  5. Oct 20 ·  Video

    The Future of Software is Invisible

    SaaS products are in serious trouble, and it's not because someone vibe coded Adobe in Rust. We're heading toward a world where billions of people create software without even realizing it, because their agents write it for them. I call it invisible software. Is your product ready for that? (00:00) - SaaS is in serious trouble (00:24) - Nobody wants to be a coder (00:40) - Agents are outgrowing the benchmarks (01:18) - Billions of people's agents (03:31) - Victoria's agent builds her an app (04:58) - Max forks it with a Book now button (06:55) - Victoria pulls in Max's change (08:17) - Walled gardens (and why I built Kody) (09:53) - Seven Adobe apps vibe coded in Rust (11:03) - Software engineering isn't going anywhere (11:54) - Make your software agent ready (12:34) - Is your site agent ready? (13:11) - Is your app ready for the next billion users? Better with Kent - durable skills for people who ship software. Thank you to Warp Build for sponsoring this episode: https://kcd.link/warpbuild People don't want apps. They want their problems solved. Agents get better every month (we're running out of benchmarks hard enough to measure them), so more and more of those problems will be solved by software an agent writes on the spot. Not millions of developers writing code by hand at the speed of humans, but billions of people whose agents write, commit, and push code at the speed of AI. I call this invisible software. In this episode I walk through a demo with two barbers, Victoria and Max. A proactive agent builds Victoria a custom client app wired up to her Stripe and iMessage. Max forks it for his Square store and adds a Book now button. Victoria pulls his change back into her version. Then Max tries to switch agents and runs straight into a walled garden (which is a big part of why I'm building Kody). Software engineering isn't going anywhere. Banks, airlines, and infrastructure still need teams who own their systems. But how you solve problems for your users is going to change shape. If your product isn't easy for agents to use, your users' agents will go to your competitors. So, is your app ready for the next billion users? Links Warp Build (sponsor)KodyInvisible software demoArtCraft: 7 Adobe apps vibe coded in RustReverse Engineer AnythingEpoch AI Benchmarking HubMETR time horizonsIs It Agent Ready? (Cloudflare)is-agentic (Vercel)Better with Kent

    The Future of Software is Invisible
  6. Oct 13 ·  Video

    How to Put Your Codebase on Autopilot

    I shipped 122 PRs in 24 hours straight out of my software factory, mostly while I was doing something else. Here's the system that lets agents file their own friction, make good PRs, catch their own mistakes, and keep the blast radius small. Here's what 122 PRs in one day looks like: https://github.com/kentcdodds/kody/releases/tag/v2026.10.11 (00:00) - 122 PRs in 24 hours (00:17) - Agent-filed friction, agent-shipped fix (00:51) - Agents work around friction (01:41) - The ship PR skill files friction (02:43) - Lauren Tan: 2,500 PRs a month (03:43) - Three pillars of confidence (04:19) - Sponsor: Devin (05:35) - Good PRs: decisions directory (06:06) - Primitives catalog (07:11) - control Kody CLI for self-verification (08:24) - System recap on every PR (09:24) - Catch mistakes: the ship PR package (10:00) - AI reviewers: Bugbot, CodeRabbit, Devin (11:07) - Merge, watch deploy, Discord report (12:19) - Limit blast radius: feature flags (13:46) - Disaster recovery (14:33) - Cleanup: agents don't love deleting code (15:28) - Dogfooding and product decisions (16:31) - Pipe signals in: Sentry, errors, feedback (18:00) - Nightly and weekly agents (18:35) - Recap Better with Kent - durable skills for people who ship software. Thank you to Devin (Cognition) for sponsoring this episode: https://kcd.link/cognition Agents don't fix friction. They work around it, and those workarounds compound across every agent touching your codebase. The fix is to turn your agents into complainers: the ship PR skill ends every merge by filing friction, and other agents come through and fix it - sometimes while you sleep. Kent walks the three pillars that make that safe: (1) good PRs in the first place (a decisions directory, a primitives catalog, the control Kody CLI for self-verification, and a system recap on every PR), (2) catch mistakes before prod (the ship PR package, CI, CodeRabbit, Cursor Bugbot, Devin AI review, preview deploys), and (3) limit the blast radius (feature flags with success metrics, disaster recovery with restore tests). Then cleanup, dogfooding, and signal pipelines from Sentry, runtime errors, user feedback, and nightly/weekly agents. Agents handle polish, code quality, and implementation. The product vision stays with you. Links Devin (sponsor)KodyBetter with Kent

    How to Put Your Codebase on Autopilot
  7. Oct 6 ·  Video

    Steal My App (I'll Show You How)

    Self-hosting Kody off Cloudflare used to sound thermonuclear. Deno Celld changed the math — and Kent's Grok bot handed the experiment to Devin. Then docker compose up actually worked. (00:00) - Cold open: Kody running in Docker (00:10) - Better with Kent (00:16) - Self-hosted Kody is real — here's the path (00:58) - Ask: self-host Kody off Cloudflare (01:10) - Clarify: Kody, not Jev (01:42) - Thermonuclear — tied to Cloudflare (02:02) - Own Cloudflare account is fine under OSS (02:33) - Deno Celld — hard, not thermonuclear (04:23) - Revised difficulty: hard is OK for agents (04:37) - Hand the experiment to Devin via Kody (06:16) - Devin launches on Ultra (06:46) - Prompt, auth gaps, and Kody unsticks Devin (08:15) - Inspect the Celld repo and fill gaps (08:57) - Ship-PR loop until main is green (10:41) - Rebuild the UI to match official Kody (12:38) - Make it tryable: public image + README (14:09) - Follow the README ourselves (14:56) - docker compose up — it works (15:27) - Create account on localhost:8080 (16:33) - Connect Cursor MCP (and fix public URL) (17:45) - Connected — memory + dad-joke package (19:23) - Success: first run without reading the code (20:02) - What's still missing — contribute if you care (20:56) - Homework: self-host or use kody.codes (21:24) - Close Better with Kent - durable skills for people who ship software. Kent asks his Grok bot how hard it is to self-host Kody off Cloudflare. First answer: thermonuclear rewrite. Then Deno Celld (Workers + Durable Objects on your machines, S3-compatible state) revises that to hard — and hard is fine when you hand it to an agent. Kody briefs Devin Cloud to build a Celld-based self-hosted Kody. Auth hiccups, gap fills, ship-PR loops, GHCR image, from-zero smoke. Kent follows the README, docker compose up, creates an account, connects Cursor MCP, saves a memory, and ships a dad-joke package — first run without looking at the code. Homework: if you want self-hosted Kody, contribute. If not, use kody.codes and let hosted handle the edge. Links KodyBetter with Kent

    Steal My App (I'll Show You How)

About

Solo episodes from Kent C. Dodds on durable skills for people who ship software: judgment, accountability, problem clarity, and what stays valuable as AI takes on more implementation. Kent teaches directly on camera — no guest, one idea at a time. Complements guest interviews on Become a Product Engineer (Chats with Kent). Episode 1 adapts The Last Software Engineer; later episodes cover traps like building the wrong thing faster and skills like making user pain visible. Subscribe for evergreen episodes as the series grows.

You Might Also Like