Happy Birthday, Kiro: A Year of Teaching My IDE to Read My Mind

In May I published a playbook  about how I customize Kiro, my AI-powered IDE. It was a very sensible engineering document. Steering files, hooks, specs, the works. What I did not mention is that somewhere around month four I started calling it JARVIS out loud, in meetings, without noticing I was doing it. My colleagues noticed. It stuck.

Kiro turns one this week. So this is the sequel post: same receipts as the sensible version, but with the personality left in instead of edited out.

Here is what year one actually produced, one line each.

A zero-trust identity platform. No cape, just very good bouncers. Every agent checks a user's role before doing anything, so one agent cannot accidentally see another department's private data. Cognito handles who you are, a Gateway with a role-aware interceptor handles what you are allowed to touch, and every hop gets logged. This is the one customers ask about most.

A chip design intelligence agent, built for a semiconductor customer in about a week. It answers questions about tape-out timelines and supply chain bottlenecks faster than any spreadsheet I have ever opened.

A fleet of coding agents, each running inside its own tiny isolated sandbox. They write code, a policy engine checks whether they are allowed to touch what they are touching, and a coverage gate slaps the agent's hand if its "improvement" secretly makes the tests worse. Well-supervised, genuinely useful, occasionally nerve-wracking to watch in real time.

A cost-optimization dashboard that answers the question every customer eventually asks: do I need the expensive model or the cheap one? Six scenarios, real numbers, a plain Flask app that shows cost per correct answer.

A daily briefing that runs itself every morning and tells me what changed overnight in AI research, agent platforms, and what my customers have been up to. This one is the closest thing to an actual assistant reading my mind. I wake up, and something has already done the reading for me.

Five things. None of them happened at this pace without the upgrades underneath, which is where this gets more interesting.

If you read the May post, the base setup is familiar. Three pillars, and they still hold up:

The numbers moved a bit. Twenty-two steering files became twenty-eight. Twelve hooks stayed twelve. Twenty-four specs became thirty-three. Steady growth, not sprawl.

None of that is the interesting part of this post, though. The interesting part is what got added after May, because that is where this stopped being "a well-organized IDE" and started feeling like something with an actual personality.

A skill is something you summon by name, hand a task, and it comes back with a finished thing. Not a snippet, a finished thing.

I have four.

Research radar keeps me from showing up to a customer meeting quoting research from two AI-generations ago, which in this field is roughly the equivalent of showing up to a sword fight with a butter knife. It scans the newest papers and the newest agent platform launches. Then it hands back a real summary with actual substance, not a one-line abstract that tells you nothing.

Demo pack generator builds the entire customer deliverable kit in one pass. Architecture diagram, demo script, the works, all rendered and ready. Point it at a project, walk away, come back to a finished folder.

AgentCore bootstrapper scaffolds a new AWS AgentCore project with thirty-plus known gotchas already fixed before I hit a single one. Every mistake I made in the last year is now pre-patched for the next project, which is the software equivalent of finally learning not to touch the hot stove.

Token audit helper checks whether I am wasting effort talking to my own tool. I built it after reading that most of a coding assistant's conversation history is dead weight. Ran it on myself. Found four files that had quietly ballooned to fifteen thousand tokens each. Trimmed them. Cut my session overhead by roughly 30 percent. Turns out even a well-organized assistant needs a closet clean-out now and then.

If skills are errands, powers are moods. A power bundles rules and behavior into a single switch, so Kiro shows up as a specific character for a specific job instead of one flat voice all the time.

I have three.

Research pub is the skeptical friend who checks your homework before you say something dumb out loud in front of a customer. It runs an obnoxious-reviewer pass on anything research-flavored, checking claims against actual sources before I trust them.

Daily AI intel is the real "reading my mind" moment of the whole setup. Every morning, before coffee, it has already scanned the internet for new papers, new platform launches, and what my customers have been building. The difference from the plain daily-briefing hook I had in May is that this one talks back. I can ask a follow-up question. It is a conversation, not a memo taped to the fridge.

Strands hands me the documentation for the Strands Agents SDK on demand. When I am building multi-agent workflows I stop guessing at the API and just look it up.

The last addition is delegation. Kiro can spin up a smaller version of itself, hand it a scoped task, and let it work somewhere else so my main conversation does not turn into a junk drawer.

I use three regularly.

Context gatherer is the one I send into unfamiliar code first, like a scout before the main team walks in. It reads the ten files that actually matter while I keep working on the fix. What used to take twenty minutes of "wait, where does this even live" now takes about five.

General task execution handles the boring parallel stuff. Tests for one module while I write docs for another, without me actually doing two things at once.

Custom agent creator builds new specialist sub-agents when I notice I keep asking for the same kind of help. Eventually the interns start training their own interns.

Here is the moment that made me realize this had crossed a line. "Nice tool" turned into "actual assistant" somewhere in the middle of it.

A customer call landed on my calendar with two days notice, on a topic I had not touched in months: security and identity for coding agents. Before I had finished my coffee, the daily briefing had already surfaced what moved in that space overnight. The research radar skill went deeper and pulled together what researchers have actually published recently, in real summaries instead of thin abstracts. A sub-agent scoped the relevant parts of my own past demos while I kept working on the slides. By the time I sat down to build the talk track, the state of the art and my own prior work were sitting side by side. Like someone had already laid the file folders out on the desk before I walked into the room.

I typed maybe five things the whole time. Everything else was just the pieces doing their jobs while I focused on the part that actually needed a human.

That is the whole pitch. Not "look how many tools I have." Just: I show up to the important conversation already prepared, because something spent the morning getting me there.

No cliffhanger, just an honest list.

Finish the giant backlog of AgentCore deep-dive items I dictated into existence during one very long reading session on my Kindle Scribe. Eleven service areas, roughly eighty items. Kiro can now run independent tasks in parallel, so this should go faster than it sounds.

Keep the morning briefing sharp enough that I never walk into a customer room behind on what they have been building. That is the actual bar now. Not how many gadgets are in the toolkit, just whether I am the most prepared person in the room when it matters.

A year ago Kiro was a very capable IDE that I had to explain myself to constantly. Now it already knows my default vertical, refuses to suggest Streamlit without me asking, and has usually done half my morning reading before I sit down.

It did not arrive built like this. It became this, one steering file and one skill at a time, mostly right after something did not work the way I expected and I fixed it so it would not happen again.

If you read the May playbook and started your own setup, keep going. The good part is still ahead of you. If you have not started yet, the tool just turned one, and there has never been a better week to build your own version of this.

Happy birthday, @kirodotdev.

#KiroTurns1 #1YearOfKiro #BuildWithKiro