Engineering notes / Muhammad Kashif
ChatGPT Sites and Work: Shipping Pages Without Losing the Brief
ChatGPT Sites and Work features make it suspiciously easy to ship a page or a slide deck before the brief has finished arguing with itself. The same notebook week shouted about tokenmaxxing — companies pushing every empl…
ChatGPT Sites and Work: Shipping Pages Without Losing the Brief
By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan
ChatGPT Sites and Work features make it suspiciously easy to ship a page or a slide deck before the brief has finished arguing with itself. The same notebook week shouted about tokenmaxxing — companies pushing every employee into AI until bills explode into crores a month or a year's budget burns in weeks. I want the shipping speed. I refuse the blank-cheque context window.
Ship the page, keep the brief
Sites-style flows and Work surfaces can turn a paragraph into a published-looking page before stakeholders agree on the offer. That speed is useful for internal prototypes. It is hostile to client work when "looks live" gets mistaken for "is approved."
My rule: watermark drafts, block custom domains until the brief checklist is green, and keep the source brief linked from the page footer in staging. Pretty without agreement is just a confusing meeting waiting to happen.
Tokenmaxxing as a management failure
Forcing every employee to use AI for everything detonates token bills. Jensen Huang arguing for huge per-engineer token budgets is a strategy take, not a licence for unmetered chaos. I meter by team, set defaults to cheaper models, and require a reason to escalate — like defaulting a printer to black-and-white.
Crore-scale burn in weeks is usually unbounded context paste plus frontier models on trivia. That is not transformation. That is a fire drill with invoices.
Routing, caching, lean context
Simple tasks to cheap models; hard tasks — a two-hundred-page contract — to top models. Cache handbooks instead of re-reading them ten times. Prefer Projects over resending long chats. Run open models locally with LM Studio and Gemma when privacy or cost demands it; a Mac Studio as an inference rig is a real option for some teams.
These fixes are boring. Boring is how finance stops hating the AI program.
Same prompt, three models, judged honestly
The notes compared GPT-5.6, Grok 4.5, and Claude Fable 5 on results, time, and cost — games, Minecraft clones, ads from mood boards. I copy the method, not the winners. Winners rotate. The comparison spreadsheet stays.
When Sites generates a page, I note which model and which brief version produced it. Otherwise A/B tests become mythology.
Creative toys with a cost column
GTA-style scenes, Minecraft clones, thirty-second product ads from HTML and three.js are excellent skill builders. They also burn tokens if everyone on a fifty-person team "just tries." I schedule creative labs with a budget ceiling and a show-and-tell, not an open tab forever.
Learning budgets are explicit. Production budgets are explicit. Mixing them is how both get cut.
Work surfaces for reports, not for secrets
Work-for-data and slide generation belong on classified-appropriate data. I ban pasting payroll and customer dumps into exploratory Sites experiments. Separate workspaces, separate memory, separate fools-rush-in energy.
If a page needs live data, it graduates to a repo with auth — not a forever-draft that quietly became production.
What I sell
Brief-gated publishing checklists, token routing policies, and a monthly cost review that names the top ten prompts by spend. Leaders who want Sites speed without metering get a demo. Leaders who want metering get a program.
Closing gate
No custom domain until brief sign-off. No frontier model default. No unbounded Projects that hoard secrets. Ship ugly drafts early; ship polished pages only when the brief and the bill both make sense.
Brief watermark that actually ships
Staging Sites pages show "DRAFT — brief unsigned" in the header until the checklist is green. I have lost arguments with stakeholders who thought pretty meant approved. Watermarks are ugly on purpose. Token routing policies do not fix social confusion; watermarks do.
When someone asks to remove the watermark early for a "quick share," I share a screenshot instead of the live URL.
Top-ten prompt spend review
Monthly finance-friendly report: ten prompts by spend, owner, whether a cheaper model would have sufficed. Tokenmaxxing dies in daylight. Sites and Work features stay, metered.
Learning budget vs production budget
Two line items. Creative Minecraft clones and GTA-style labs burn the learning budget. Client Sites pages burn production. Crossing the streams is how both get cut by finance after one scary invoice.