SHED HAPPENSOne-page PDFMaking AI work accountable.
Charles “Chuck” McFarland built an AI-assisted work system to help build Shed Happens, a wildlife data company that ships an app.
He brought construction-management discipline to AI: written work orders, evidence, separate inspection and owner approval. He directs the work, defines measurement rules and reviews labels; Claude and Codex help implement it.
Work with a record.
Each job carries its scope, tests, evidence and review into the next session. Production changes, external releases, new spending and public communications require specific founder approval.
- Scope
- Build
- Check
- Remember
Two examples, with receipts.
Approval stays specific
A recorded test found that one approval could clear two unrelated requests. The repair bound each decision to its own request. Saved fixtures show unrelated work remaining blocked.
A test can say no
In a new-scan test on 26 antler sides from 18 cards, a candidate increased independent tip matches from 51.2% to 63.4%, but produced more stray guesses. The improvement was uncertain, and the candidate failed its acceptance rule. It was not adopted.
October 27 status: an audit reported training/test overlap in the earlier model lineage. Chuck stopped the retrain and ordered a clean reset. Earlier training is treated as unproven; no validated new model is claimed.
Measure the resources, too.
A local-log report estimates $12,633 of Claude API-equivalent usage in September and $4,071 for October 1–27, against a stated $200 monthly subscription. These are estimated usage values, not invoices, realized savings or proven productivity gains. Codex is separate.
The contribution is a practical way to bound AI work, inspect its claims and keep the evidence—including when the answer is “this did not pass.”