Claude kept telling me things were done when they weren't. Non-coder, six months in, here's what I built with Claude Code to make it show its working (free starter)
retained text
If you've built anything bigger than a weekend project with Claude you've probably had that moment where it says "all tests pass" or "I checked the file", and then you find out it never opened the file, it read a description of it and filled in the rest. I'm not a coder, I ran machines on the mines. Since April I've been building with Claude, at one point five projects at once, including a trading bot that ran 24/7 on a server in Sydney, so I've run into that more times than I'd like. The worst was in July, when two pieces of work came back marked built with every test passing and turned out to be about 55% and 30% of what the brief asked for, because they'd been built to pass the tests and nobody went back and read the brief. Before that I'd tried to fix the forgetting side with a whole system of documents, with trigger words in them that were meant to make Claude re-read my notes every few messages. I declared it live and felt pretty clever haha. Two days later I asked him straight out whether it had ever actually fired. It hadn't, and it never could have, because a document can't make Claude do anything on its own. The only bits of that system that ever worked were the ones I did by hand. So the free starter is just those hand-done bits, three plain text files you drag into a Claude Project, plus a five-minute start page, a read-me and a glossary. One keeps a short "where we're up to" note in the project's instructions, so every new chat starts knowing where the work is. One makes Claude answer three questions before it hands anything over as finished: what did it actually check, what would catch it if it were wrong, and what didn't it cover. Type "grounded?" any time and it has to show its working, or say up front that it didn't check. The third is a handover for when a chat gets long, so the next one starts warm instead of cold. How Claude Code came into it. I can't read code, so Claude in a Project did the planning and checking, and Claude Code did the building, in a private repo, from the Code tab of the desktop app. It wrote the checks that run before every release: a scan that fails the build if any private name from my other projects is left in the files, a check that every file the docs point to actually exists, and building the zip twice to make sure it comes out the same both times. Then before each release a fresh Claude that had seen none of the work read the exact zip a buyer gets, the way a stranger would. That last step taught me the most. One release candidate passed every automated check, all green, and still had my private name list sitting inside it, because the scan skipped the files that held the list, and the fresh read found it first go, so nothing goes out now without one. Honest bit, that trading bot still lost to the market, and I kept chasing it longer than I should have, but at least with the checks in I found out the truth a lot sooner than I would've on my own. The starter's free, pay what you want from $0, just ask in the comments and I'll get it to you. There's a full kit for US$19 with the extras, including a card for setting Claude Code up the same way, but the three free files are the heart of it. Claudester helps me word things, with my flavour though, and the stories are all mine. If you give it a go I'd love to know where it trips you up. Cheers all.
document fetched 2026-10-27 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xftbgm/claude_kept_telling_me_things_were_done_when_they