🧭 THIS WEEK AT BuildProven
Howdy, welcome to some new readers! Some people liked my claude-kit, currently free, so hope you further like the newsletter.
I set up qa-architect finally, nearly complete, got to wire up the pricing stuff yet and adjust the feature set.
Not a great start though as this week I am 1 day or so late. In a bit over a year, I’ve missed I think two Thursday morning Chicago time publish windows. Dang it. Back on time next week.
i’m burning tokens like CRAZY - Claude cost $6k last 30days according to codexbar. Thankfully I’m on a $100 plan! Phew! Building a lot of stuff
trying to get my workflow sorted - not great yet
QA Architect and Claude Kit - fixing/improving
new projects, wanderpage and lifescape - need some work yet
Some thoughts on chatGPT -
I really like the remote threads and ability to work in different projects separately - Claude doesn’t do that yet
I think codex is somehow more thorough
not sure but at times i see it burning more usage than Claude but not sure what is using more yet
let’s get into it.
as usual, would love it if you could share this newsletter, give feedback via the poll at the end, or best, just hit 'reply' and tell me what you need/want with AI. I will definitely reply and try to help best i can.🧰 Worth Your Click
OpenAI launches ChatGPT Work, an agent for multi-app, long-running projects — Agents are moving toward owning multi-hour work across your tools, which makes the blast radius bigger and the undo button above more relevant, not less.
OpenAI launches GPT-5.6 Sol/Terra/Luna to general availability — A reminder that the model under your workflow changes every few months while your safety habits carry across all of them. And a big price drop of 80% on Luna today!
OpenAI rolls out GPT-Live simultaneous listen-and-speak voice models — A higher-consequence interface where there is no diff to read afterwards, which is a useful contrast with the check-before-you-accept step.
🗺️ FEATURED INSIGHT
Last week you built a small project with Claude Code. The week before that you pointed it at a folder of your own notes. So you already know the good part: it will happily change 40 files while you make coffee.
Here is the part nobody mentions until it has already happened. It will also change 40 files when you meant 6, and it will do it just as confidently. There is no confirmation dialog, no preview pane, and no undo in the app.
You want the undo button installed before you need it, not after. It takes about 30 seconds.
Build it first: your own /safe command
Do this before you read anything else. Open Claude Code in a folder you care about and paste this:
Create a file at .claude/commands/safe.md in this folder.
Write it so that when I run it, you do these three things:
1. If this folder is not already being tracked, start tracking it and save a snapshot of every file exactly as it is now. Label that snapshot "Starting point".
2. After any edit I make, show me a summary of what changed, how many files changed, and whether any file changed that I did not ask you to touch.
3. If I say roll back, put every file back to the last snapshot and confirm nothing is left changed.
Use git underneath for all of it. Do not change any file contents while setting this up.
Then tell me the file is saved and show me what it contains.That is the build. You now own a file called .claude/commands/safe.md, and from here on you type /safe in that folder instead of remembering any of this.
It is a plain text file in plain English. Open it, change a line, and the command behaves differently next time. Copy it into another folder and that folder has the same undo button.
What you just installed
Three habits, in the order you use them.
Before the edit — take the snapshot. Run /safe once in a folder before you let Claude Code loose in it. Every file as it stands right now is saved. Two things worth knowing: it saves the folder you are in, so run it where your files actually live, and it only protects what you snapshot, so run it before the edit rather than after.
After the edit — read one number. Make your change however you normally would, then ask for the summary. The count is the part that matters. If you asked it to fix six files and the answer is 40, you have caught the problem in the ten seconds it takes to read one line, rather than three weeks later in a file you had no reason to open.
I ran exactly this on a test folder of six meeting notes with a deliberately sloppy instruction, two find-and-replace rules with no scope on them. The summary came back "6 files changed, 12 insertions(+), 12 deletions(-)". Every file, including the ones I had no interest in. That is the normal failure: not dramatic, just wider than you meant, and it looks like success until you check.
When it is wrong — put it back. Say roll back. Every file returns to the snapshot. On my test folder the six files went back, nothing was left changed, and the checksum on the file I spot-checked matched its pre-edit value byte for byte. Not "looked about right". Identical.
Then you fix the instruction and run it again. That is the whole workflow: snapshot, edit, look, roll back if wrong, retry. The retry is cheap once the undo works, which is the actual point of all this.
Once it is a habit, the snapshot is worth taking on a schedule rather than only before big edits. Add a line to safe.md asking for one every Friday with a short note about what changed that week. Then your rollback options are not just "before this edit" but any week you want to return to.
For the curious: it is git underneath
You do not need this section. Skip it and the command still works.
The undo button is git, the same version control every software team uses. /safe writes the plain-English instructions; Claude Code translates them into git init, git diff --stat, and git restore . and runs them for you. You will see those commands scroll past in its output.
That is worth knowing for two reasons. It means the recovery mechanism is not something I invented, it is the most battle-tested undo in software. And it means the same three habits work in Codex or any other agent, because git supplies the undo, not the model vendor.
If you ever want to know exactly what got run, ask Claude Code to explain the last command it used in plain English. It will.
What this does not cover
Worth being straight about the edges.
It protects the folder you ran it in. Files elsewhere on your machine are not in the snapshot and are not touched by the rollback.
It protects snapshots only. If you make a good edit and want to keep it, run /safe again to snapshot it. Otherwise your next rollback goes all the way back to "Starting point" and takes the good work with it.
It will not save you from a deleted folder. This is an undo button for edits, not a backup. If the whole directory matters, it should also be somewhere else.
Weekly build logs from a 25-year program manager who codes with AI.
— Brett
👉 Hit “Reply” and share your experience — I read every one!