Six Loops That Make GPT-5 Worth the Upgrade
GPT-5's reasoning is sharper but you have to coax it. These six loops are the difference between a confident wrong answer and a verified right one.
The GPT-5 reasoning dial
GPT-5 ships with a "reasoning effort" parameter most users never touch. The default is medium. The loops below all assume high - set it once at the top of your conversation and forget it.
1. Always-on Chain-of-Verification
On any factual question, fire /cov immediately after the first answer. GPT-5 will produce a list of claims, an independent verification of each, and a corrected rewrite. We've seen this catch 4 out of 5 hallucinations on biographies, dates, and statistics.
2. Skeptic Loop for synthesis
When GPT-5 summarizes a long document, it omits weak evidence. /skeptic surfaces the three weakest claims and the contradicting evidence. This single loop turns a polished summary into a defensible one.
3. Compare Three Options for code architecture
Ask GPT-5 to "build a notifications service" and you get one design. Add /compare and you get three: a queue-based design, an event-stream design, and a simple cron + table design. Each comes with the situation it's clearly best in.
4. Red Team for product plans
GPT-5 writes great launch plans and terrible risk sections. /redteam fixes this. Ten failure modes, scored, with the top three to mitigate before ship.
5. Persona Shift for the exec readout
Your engineering plan is too dense for the VP. /persona "VP of Sales who needs to know what changes for her team" rewrites it without losing substance.
6. Strict JSON for pipelines
When you're calling GPT-5 from code, end every prompt with /json and a schema. GPT-5 obeys schemas better than 4.x but still needs the explicit nudge.
The meta-loop
The single best habit: never ship the first answer. Fire one loop, always. The 20-second cost is trivial; the upgrade in answer quality is not.