Eight problems, eight prompts.
Each row is a thing teams actually complain about, what Loupe does with it, and the exact rewrite that closed it.
Rework loops
One task, nine prompts. Nobody sees it happen.
Loupe groups turns into tasks and flags loops over four.
fix the rate limiter
Replace the fixed-window limiter in src/mw/rate_limit.go with a token bucket: 100 rps, burst 20. Tests must pass.
Context blindness
31% of prompts name no file, branch or test.
A context score per prompt, plus a nudge in the editor.
the login is broken
Login fails on Safari 18 after src/auth/session.ts:88. Repro in e2e/auth.spec.ts.
Secrets in prompts
Keys and prod URLs pasted straight into chat.
Redaction on the laptop before anything is uploaded.
connect to postgres://admin:hunter2@prod-db...
connect to [redacted · database url]
Ramp-up time
You cannot coach what you cannot see.
Week-one prompt quality vs team median, with two tips.
how does this repo work
Summarise the request path from cmd/api/main.go to the first handler. Cite files.
Silent spend
Agent bills grow, nobody can attribute them.
Cost per repo, team and merged PR.
12-turn loop on infra · $3.40
2-turn playbook run · $0.72
Blind spots
A hook gets removed and telemetry quietly stops.
Fleet health, tamper alerts, cloud-agent enrolment.
4 laptops silent for 9 days
MDM re-push, alert in Slack, 0 silent
Freestyle prompting
Everyone reinvents the same prompt badly.
Winning prompts become playbooks, pasted from the CLI.
14 versions of "write tests"
1 playbook · 84% merge rate
Audit questions
"Show me what AI touched this release."
Prompt → commit → PR trail, exportable.
a week of git archaeology
one filtered export