I have spent days debugging my own carefully typed functions. I have also watched agent-drafted code pass tests and behave correctly the first time. The thing that matters is not who produced the bytes. It is whether the problem was decomposed clearly enough to inspect.
Meanwhile, TShirtize
So here is a product about estimating software effort, built by the thing skeptics say cannot build software, with an effort profile that would have looked dishonest if I had estimated it the old way. TShirtize came together in about three hours of real work spread across four days. Not because I am fast, but because once the decomposition was clean, the implementation was cheap and the agent handled it.
The suite now has ten apps. The August history shows the progression clearly. The first week brought a broad architecture sync: diagrams for Weekly, Almanac, Gopher, and DrawmaticAI were refreshed to match the code that is actually running. Then the second half of the month belonged to the new member. TShirtize joined on August 14, got its first architecture diagram that same morning, then picked up Monday item creation, an automated test layer, and a final diagram refresh by August 16. The Code Reviewer diagram got a matching update the same day.
What It Actually Does
The pitch is simple. Paste a Monday.com pulse URL, a GitHub issue or PR URL, or just type a free-text description. The app reads what it can, adds context from the repo or board, and returns a t-shirt size along with a rationale, a suggested definition of done, confidence level, spike flags, and clarifying questions. If you gave it plain text, it also offers a Monday ticket tree: epic, stories, and subtasks, none of which are created until you confirm each one.
That is the surface. The useful part is underneath. TShirtize is doing the context-gathering I normally do before I can think clearly about a ticket: reading the description, skimming the comments, looking at the diff, checking recent commits, scanning the repo structure. It packages all of that into a single structured answer so I can start at the decomposition instead of ending there.
The Recursion
TShirtize sizes work by reading Monday, GitHub, or free text. In a nice recursion, the act of building it was sized the same way: small, bounded units, clear definition of done, test coverage as part of the estimate. The product does what the workflow preaches. I keep waiting for the hidden tax to show up, but the honest read is simpler: decompose well up front, and the quality comes cheap. That still surprises me.
Calibration, Not Prediction
The calibration layer is what gives it staying power. After a tenant records a few completed sizings per size and input type, historical averages fold back into the prompt. Link a merged PR and TShirtize can infer effort from open-to-merge hours and commit count. That means the estimates get more honest over time, not less. Most estimation tools start opinionated and stay there. This one is designed to be corrected by reality.
The Monday writeback is another signal of intent. TShirtize could have auto-created tickets, but it does not. It suggests the tree, gates stories behind the epic, gates subtasks behind the story, and asks for a browser confirmation every time. The user keeps the authority. The app is an assistant, not an author.
The Test Pyramid Nudge
One detail I like is the sizing cheat sheet and the built-in reminder that real work carries a testing cost. It sounds small, but most informal estimates quietly pretend testing is free. Having unit, API, and UI or E2E test tasks sit next to the size badge changes the shape of the conversation. It is not a rule; it is a prompt to remember what the estimate actually has to cover.
Decent Enough, and That Is Enough
My honest read is that TShirtize is a decent product with a very useful core. It will not replace project managers or planning poker. It is not trying to. What it replaces is the empty-screen moment where you have a ticket in front of you and no idea where to start. It gives you a first draft of the breakdown, the questions, and the done criteria. From there, a human with context can refine, disagree, and ship.
The suite now has ten apps, and TShirtize is the one that handles the gap between "we have work" and "we understand the work." That is a narrow gap, but it is where a lot of projects stall. Closing it with something small and grounded feels like the right move for August.