Positioning · 0 to 1 · Claude Code
Winledger
I built a work log for myself in a weekend. A year later I audited it like a product I had to launch, cut my favourite feature, and rebuilt it as Winledger.
2025, rebuilt 2026
Why I Built This
A few weeks into my role as Customer Product Manager at Zuper, a colleague asked me a simple question: “So what have you been working on lately?”
I blanked.
Not because I hadn’t done enough. The opposite. I worked across product, CX, sales, marketing and RevOps at the same time. In any given week I might sit in a customer call, write a spec, debug an implementation and review a roadmap request. The work was scattered by design, and scattered work doesn’t add up to a story on its own. It just disappears.
So I built Pulseboard over a weekend. I scoped it with GPT, wrote the spec with Claude, built the UI on Lovable and put it on Supabase. It logged my work by category and team, showed where my effort went, and used AI to draft a performance summary from the log.
It taught me one real thing. Within two weeks of logging, I could see I spent far more time on implementation support and cross-team coordination than I’d estimated. The work that felt most visible, writing specs and sitting with customers, was a smaller share of my week than I thought. That gap between what you think you do and what you actually do is still the most useful thing the product has shown me.
When I first wrote this page, I listed what I’d do differently. The AI summary was impressive before it was useful. I should have started from the output, not the input. And I should have thought harder about the name. In November 2026 I went back and did all three.
Reviewing My Own Product Like I Had to Launch It
I came back to Pulseboard with three questions. Is it clear what this is? Would a stranger’s first five minutes work? Would someone trust it with a year of their work?
The answers were uncomfortable.
The product didn’t know what it was. This page called it a reflection tool. The app’s own landing page called it a performance review generator for product managers, engineers and designers. Neither was quite true.
Onboarding assumed my job. You had to name at least one team before you could start, reviews ran on calendar quarters, and the examples were all product and engineering. A freelancer, a nurse or anyone on an April fiscal year was either blocked or mis-served on the second screen.
The trust story had holes. The AI endpoint didn’t check who was calling, so anyone holding the site’s public key could use it as a free model on my credits. There was no way to delete an account, and no notice that your work notes went to a third-party model.
And there were bugs that quietly damaged the record, which is the one thing a work log can’t do. Cancelling an edit and then adding a new entry overwrote the old one. Dates were read as UTC, so anyone west of London would see every entry a day early. Lists silently stopped at a thousand entries. In November, the weekly view opened on week three of January.
Decisions That Weren’t Obvious
From Pulseboard to Winledger
Version one, 2025: one weekend
Version two, Nov 2026: audit, decide, rebuild
01. It’s a brag sheet, and its competitor is a spreadsheet
Almost everyone I know in a corporate job keeps an Excel file of their wins and digs it out at review time or when they start job hunting. That’s the job this product does, and the reflection framing was a nicer story than the one people actually live.
Naming the competitor changed the roadmap. Logging a win has to be faster than adding a row, so the main screen opens on a single line: type what you did, press Enter, done. And the first-run experience became “bring the sheet you already keep”. The import takes any columns and any layout. It guesses which column is which, works out whether 03/04 means March or April from the rest of the sheet, reads months and quarters, previews the result and explains every row it skips. For people without a sheet, there’s a start-fresh path: pick your role for suggested categories, then list what you remember from each recent quarter, one line per win, no exact dates needed.
02. I removed the AI
This was the hardest call, because the AI summary was the feature I was proudest of. Two things decided it. The prompt asked the model to “quantify achievements”, which invites invented numbers in a document people hand to their employer. And a brag sheet is full of confidential work detail that people had no reason to expect would leave the app.
The summary is now built only from what you wrote: pick a period and get plain text grouped by category, with the impact and praise attached, ready to paste into a self-review. Nothing is rewritten, summarised or invented, and no AI provider receives anyone’s entries. Knowing where a model doesn’t earn its place is part of the job.
03. Fix the record before adding anything
A work log that loses or shifts entries is worse than none, because it still looks complete. Every data bug came before any new feature. Dates now go through one module that treats them as calendar days, and the tests run in Los Angeles, Kolkata and Auckland time, because date bugs only show up away from UTC. The fetch pages past a thousand entries. Saves resolve as success or failure, and the form only closes on success.
04. Own the stack
The weekend version ran on a Supabase project that Lovable created and controlled. I moved everything onto infrastructure I own. The database schema was rewritten from scratch, and its row-level security was checked against a local Postgres with two test users: neither could read or write the other’s entries or files. On top of that sit account deletion, plain-language privacy and terms pages, and Google sign-in. Hosting is on Vercel.
05. A new name
On the original page I wrote that names are cheaper to change before launch than after. So I changed it. I ran a brand pass with one hard constraint, that the name’s address had to actually be free, and Winledger came out of it: a running record you add to and read back, which is exactly what the spreadsheet was trying to be.
How Version Two Was Built
The weekend version tested what a PM can ship with no engineering team. Version two tested something else: what a PM can ship with an AI coding agent as the engineering team.
I used Claude Code for the whole rebuild. It started with four parallel reviews of the codebase, each from one angle: a product audit, a daily user’s journey, whether onboarding fit anyone but me, and multi-user security. I merged those into one prioritised plan. Then three tracks of work ran in parallel on separate git branches (brand and landing page, the core app, and trust and onboarding), and I merged and tested them.
My job was the part an agent can’t own. I made the positioning call, decided the AI goes, set the order (fix the record first), approved each step that touched real accounts, and checked the results. The agent was fast at reading code and writing it. It still needed someone to decide what the product was for.
What I’d Do Differently
There is one user, and it’s me. Everything above is a claim about what should work. I don’t yet have evidence that anyone else returns weekly, which is the only metric that matters for a log. The spreadsheet’s real failure is that people stop updating it in April, and software alone doesn’t fix a habit. Reminders and a weekly nudge are the obvious next step, and I haven’t built them.
The signal-versus-noise flag still isn’t built. Version one’s biggest lesson was that a summary should reflect significance, not volume. A way to mark the wins that matter is still on the list, two versions later.
Some edges are visible. Google’s consent screen names the backend’s address instead of Winledger’s, because a fix needs a paid custom domain. Sign-up emails come from Supabase’s built-in mailer, which is rate limited and not meant for a real launch.
What This Is Evidence Of
The interesting work in version two wasn’t the code. It was deciding what the product was (a brag sheet, not a reflection system), who it competes with (the spreadsheet people already keep), what to cut (the feature I was proudest of), and what has to be true before anything new gets built (the record can’t lose your work).
It’s also a working answer to a question I get asked a lot: what changes when a PM has an AI agent that can write the code? The building gets cheap. The judgement doesn’t.
Winledger is a React and TypeScript web app on Supabase (Postgres, auth, storage), hosted on Vercel and built with Claude Code. Version one, Pulseboard, was built on Lovable in 2025. Winledger is live and free: bring the spreadsheet you already keep, or log this week’s wins one line at a time.