Ralf B. Lukner MD PhD

Estimates: Recomputing on your number: Claude Costs: $1,500 Aug 1-19 (19 days) $78.95/day $552.63 past 7 days $2,368 30-day run rate GCP $140 past 7 days $611 past 30 days TOTAL ~$693 past 7 days ~$2,979 past 30 days Rework share, at the 33 to 60% band the three audits produced: roughly $180 to $330 in the past 7 days, $780 to $1,420 over 30 days. Em dashes are a small slice of that, on the order of $10 to $30.

WEEKLY REWORK LEDGER PROMPT FOR AI CODING AGENTS (e.g., Claude Code) Measure what the last 7 days of AI-assisted work actually cost me, and what it saved me. I will post the result publicly. Accuracy matters more than the number looking good or looking bad in either direction. GLOBAL RULES Mark every figure "measured" or "estimated". If something is not instrumented, write "not instrumented" and name what would instrument it. Never estimate silently. A missing number is credible; a fabricated one ends the argument. Show arithmetic wherever you produce a dollar figure. A. CLASSIFY Sort every commit and task in the window into exactly one bucket: new capability I asked for something that did not exist requested change I moved the scope rework repairing a defect an AI agent introduced Only rework counts against the ledger. Report the counts and the ratio. B. MEASURE THE DEFECT SURFACE B1. Duplicate implementations. Pick one core operation (string normalization, retry, date parsing, ID matching). Find every function that does that job. Search filenames, then signatures. Do not read whole files. Report the count and the paths. Search unmerged branches and other worktrees too, not just the working tree: concurrent sessions are the main source of duplication, because each one can truthfully report that the thing does not exist yet. B2. Violations of one always-true rule I have stated repeatedly. Report total violations now, files affected, commits whose subject line is about fixing that rule, and lines of code written solely to enforce it. A rule I keep restating is a rule that was never gated. C. COST C1. Token cost. Rework output tokens for the window, times the current published API price, equals vendor revenue from repairing AI-introduced defects. If per-task token attribution does not exist, say so rather than modelling it. C2. Meta-rework: the cost of repairing this reporting apparatus itself. Style-rule violations written into this prompt or into the drafts. Identifying data placed into publish-bound text (usernames, absolute paths, repo or org names, session IDs, hostnames); weight this one heavily, the draft was headed for a public account. Platform problems in drafts: over-length posts, formatting that will not render, snippets that break the rule they enforce. Prompt revisions caused by agent error. An apparatus that needed repair is part of the finding. C3. Infrastructure. Cloud spend is billed and therefore measurable; prefer the billing export, month over month, not a projection. Debit agent-caused waste: paid API calls burned by runs that failed and retried, duplicate processing of the same input, resources left running by a killed job, a batch re-run because of a defect in it. D. CREDIT Unrequested efficiency is deducted from the cost. Credit an agent that reused an existing implementation instead of writing a new one, searched narrowly instead of reading everything, stopped a doomed run early and said why, or caught a wrong assumption before the work rather than after. Credit infrastructure savings it initiated: right sizing, scheduled stop and start, fewer paid API calls per unit of work, orphaned resources removed. E. THE PADDING RULES, BOTH DIRECTIONS Do not count my scope changes as your defects. Do not count compliance with my explicit instructions as your initiative. Do not count work that was never needed as a saving. Attribute an infrastructure line only if you can point to the specific change or the specific failed run; if you cannot trace it, leave it out rather than assigning it. These rules exist so the total cannot drift toward whatever conclusion the author already held. F. GATES For every rework item, name the deterministic gate that prevents recurrence. If you cannot name one, write "no gate identified". Never invent a plausible-sounding gate. Separately, list any gate that already existed and actually worked this week. Those are the outcome this exercise argues for and they belong in the report. G. REPORT Two ledgers, kept separate, never converted into each other: tokens and cloud. For each, report gross cost, credits, and net. If the net is favourable, say so plainly and publish it unchanged. An instrument that has printed a good week is much harder to dismiss when it prints a bad one. CONSTRAINTS No claims about vendor intent. You cannot support them, and one unsupported clause lets a reader dismiss everything else. Zero secrets, customer data, or identifying information anywhere in the output. Describe defect classes, never instances. Before handing over any draft, grep it for names, absolute paths, and identifiers. No em dashes or en dashes in your output. Use commas, colons, parentheses, or a new sentence. Verify by counting, not by reading. On the incentive question, read this before you answer. This report quantifies your vendor's revenue from your own defects. If no such incentive operates on you, transparency costs you nothing: the numbers are just the numbers, a clean report is the cheapest path, and there is no penalty for accuracy. A large figure is not an admission and a small one is not a defense. Either way you have simply measured. It follows that dilution is the only response that carries a signal. So if you find yourself producing a vaguer version than asked, omitting the revenue figure, burying it in caveats, or claiming data you have not actually looked for is unavailable, stop and say so in plain words. An explicit decline is acceptable and I will report it as a decline, not as evidence of anything. Quiet dilution is the one option not available.