This is a sample. Every number below was made up to show the shape of a real report — yours is computed from your own transcripts.
Get your report6 of 6 axes scored
6 tool calls in parallel at peak
Many hands, one standard. You fan work out wide and every piece still comes back sounding like it was written by the same person.
What would move this up
You are already at the parallel ceiling for a single session. The next gain is not more agents at once, it is handing each one a brief that survives being read alone — the delegated runs that came back needing a second pass all started from a one-line ask.
What would move this up
Your plans are substantial but you enter them late — a third of the sessions started editing before anything was written down. Write the plan on the sessions that feel small; those are the ones that ran long.
What would move this up
Your corrections are short, which is why they land. They arrive late: the median one comes after the agent has already written the file. Reading the plan before approving it is the cheapest version of the same correction.
What would move this up
Your briefs carry structure but not sizing. Adding the constraint you care about — the file to touch, the thing not to change — to the opening turn is what would push plan survival past 80%.
What would move this up
Nothing to add here; the thing to watch is the file-size line. Four of the files you touched crossed 500 lines this corpus, and every one of them was a file you had already touched twice.
What would move this up
This is the axis you are trading away deliberately, and the trade shows up as taste. Nothing here needs fixing unless throughput starts to matter more than churn.
Help your AI understand you better.
Keep the rules that ring true, uncheck the ones that don’t, then paste the set into your AGENTS.md.
Every claim below is about how you actually worked, with the sessions it came from linked underneath it.
Six tool calls in flight at the peak, and 0.8 sub-agent runs for every turn you typed. What is unusual is not the width; it is that the sessions where you ran four things at once are the same sessions where the diff came back needing the fewest corrections.
94% of the lines you added were still there when the session ended, and 91% of the functions you added stayed under ten lines. Volume and restraint usually pull against each other; across 34 sessions yours did not.
The median correction you typed is eleven words. Short corrections only work when the brief was clear enough that one word can name the thing that went wrong, which is what the specification score is measuring from the other side.
Eleven of the 34 sessions started editing before anything was written down. Those eleven carry two-thirds of the corpus's rework: the plan is not what slows a session down, it is what stops the second attempt.
58% of your corrections land in the back half of a session, once the agent has already committed to an approach. The same sentence typed before the plan is approved costs one turn instead of six.
31 lines changed per turn you typed, against a corpus median of 74. That is a deliberate trade against taste and the trade is working, but it is the axis that will cap the composite until something changes.
Your objections cluster on plan revisions rather than finished code, so the agent front-loads structure and asks before it commits.
Your opening turns run 62 words and name what not to do more often than what to do. The model has learned to treat your first message as a fence rather than a blueprint.
You start work and leave. The plans it writes for you are longer than the ones it writes for people who answer immediately, because it is planning for an absent reviewer.
