I let Computer ship and break a production site this weekend - lessons learned
Ran Computer (hosted) and Personal Computer (local Mac) side-by-side Fri–Sun. Burned ~2,100 credits across two runs. Shipped a production site, broke it, rebuilt it. Honest field report below.
For context: I used to lean heavily on Perplexity, then moved most of my work to Claude Code and ChatGPT, plus some Gemini/Grok. I picked up Pro again mainly to see whether Computer can actually automate the stuff I currently juggle across a few tools on my Mac.
1. Stack Pro before Max if you can.
On my account, Pro Computer credits and Max Computer credits behaved as separate pools. I started on Pro, burned through the 4,000 Pro bonus credits, then upgraded to Max and got the full Max Computer bonus on top (35,000 in my case). That netted me ~39,000 total vs ~35,000 - if I’d gone straight to Max.
This seems promo and rtiming-dependent, but if you’re eligible, it’s a nice way to squeeze extra runway out of Computer before committing to Max.
2. Filesystem-first beats URL-first (in my tests).
Early runs, I pointed Computer at live URLs for content that already existed as markdown on my Mac. The token/credit cost came out around 3–5x higher, and the model drifted more on formatting. When I switched to local `.md` paths on Personal Computer, outputs were tighter and the same work used roughly a third of the credits.
For me, it was: Markdown > HTML > transcripts as input, both for quality and cost.
3. One artifact per run.
I tried to do “site update + new page + backend function” in one big brief. That single run cost 1,087 credits over 59 minutes. Doing the same work as three scoped runs (one per artifact) would have been cheaper and much easier to revert piece by piece.
Today, Computer feel sharper at one thing well than at being a giant multi-tool orchestrating everything in one shot.
The miss: no git gate. I let Personal Computer deploy straight to Vercel with no GitHub repo in the chain. It worked. Then a follow-up run drifted hard on brand and design, and I had nothing clean to revert to. I ended up rebuilding the git history from production and spent hours on cleanup.
Rule I’m living by now: local clone → GitHub → Computer opens a PR → human reviews/merges → Vercel auto-deploys from main. Never let Computer push straight to prod.
Two asks for the Perplexity team:
1. A first-class “lock file” primitive.
Right now, I paste a `BRAND-LOCK.md` into every prompt with “DO NOT MODIFY.” A true immutable-input concept (with diff/highlight if Computer even touches it) would prevent the kind of slow brand drift that cost me Saturday.
2. A live credit meter in the Computer UI during a run.
Today, I only see the final cost in the thread/usage view; there’s no way to see “this run is already at 600 credits” while it’s executing. A running counter with an easy “abort” button would let me kill bad runs before they quietly burn 1K+ credits.
I’m planning to keep running Computer on real work for the next couple of months and will share a short field report each weekend. If you’ve found different ways to save credits or avoid drift, I’d love to compare notes.