Recommendations
What to change, with the evidence behind it.
Loading…
Loading…
What to change, with the evidence behind it.
Loading…
Findings
6
Worth acting on
6
High impact
5
Need a proof first
0
71% of everything you are charged for comes from this one step — 44 call(s), averaging 65 seconds each.
It is the most expensive, the slowest and the busiest step at once. One improvement here counts three times over, and work on any other step barely moves the total.
Do thisFix 'module.content' first. Nothing else is close.
Start to finish took 92 minutes, but the AI calls themselves only add up to 70 minutes. Running the calls that do not depend on each other at the same time would finish in roughly 1 minutes instead.
This saves time, not money — the same calls, the same tokens, the same bill. You can never go faster than the single slowest call, which is why the target is not zero.
Do thisStart with 'module.content'. It makes 44 separate calls, more than any other step, and they most likely do not need each other's answers.
598,720 tokens sent, 57,392 received, across 44 call(s).
A gap this wide usually means a long standing instruction, background text that is pasted in but never used, or examples that could be shortened. You pay for all of it on every single call.
Do thisCheck how much of the 'module.content' prompt is identical every time — that part can usually be cached or cut.
129,506 tokens sent, 10,466 received, across 12 call(s).
A gap this wide usually means a long standing instruction, background text that is pasted in but never used, or examples that could be shortened. You pay for all of it on every single call.
Do thisCheck how much of the 'outline.generate' prompt is identical every time — that part can usually be cached or cut.
90% of everything you are charged for is the prompt. You send 836,125 tokens to get 88,171 back.
Everyone reaches for a cheaper model first. When the prompt is this large, making it shorter takes more off the bill than switching models would — and it is safe, because the model and the answers it gives do not change.
Do thisLook at the standing instructions you send with every call before considering a cheaper model.
68,414 tokens sent, 12,995 received, across 11 call(s).
A gap this wide usually means a long standing instruction, background text that is pasted in but never used, or examples that could be shortened. You pay for all of it on every single call.
Do thisCheck how much of the 'assessment.generate' prompt is identical every time — that part can usually be cached or cut.
Computed from observed usage at 2026-08-19T18:04:50+00:00. Findings marked No quality risk change cost without changing what a model returns, so they need no equivalence proof. Anything that swaps a model stays gated behind one.