Why this is in the vault
A behavioral-science-grounded post-mortem on AI workflow adoption that derives four diagnostic failure modes and provisional audit rules directly applicable to RDCO's own skill inventory lifecycle.
The core argument
Katie Parrott, in Every's "Working Overtime" column, dissects why some AI workflows become permanent fixtures while others collect dust on her "Island of Misfit Workflows." She examines three failed experiments — Attention Desk (a CEO communication dashboard clone), an auto-generated social media post queue, and Margot (a personal AI assistant) — and contrasts them with what stuck: her compound writing plugin.
The diagnosis maps to four failure buckets:
- Forgotten existence — the trigger never fires naturally in actual work
- Execution failure — the automation broke or required more revival effort than the tool returned
- Maintenance overhead — ongoing upkeep consumed more time than the value produced
- No recurring problem — built for a problem that didn't recur, or didn't exist at all
The behavioral science backing she cites:
- Immediate rewards predict persistence (Woolley & Fishbach) — the compound writing plugin paid off in the moment with a paragraph to push against or an argument unstuck; Attention Desk promised future calm but never delivered present relief
- Habit formation takes ~2 months median (not "21 days"), with a range of 4 to 335 days — three days was rational evidence the tool wasn't solving a real problem
- Symbolic self-completion (Wicklund & Gollwitzer, 1982) explains building tools for aspirational rather than actual selves — building a LinkedIn scheduler for a "spontaneous, shoot-from-the-hip kind of social media user"
Her derived provisional rules for new automations:
- Run manually 3 times before scheduling it
- Output must require <5 minutes of review to be usable
- 4 consecutive unused runs triggers a keep / redesign / retire decision
Her solution: "Agent Ops" — a Codex plugin that interviews her about each workflow she built or adopted and recommends one of four outcomes: keep it, redesign it, wait for the right moment, or retire it.
The full Agent Ops prompt is paywalled (Every subscriber content).
Mapping against Ray Data Co
The four-bucket failure taxonomy maps directly onto RDCO's own skill inventory. Skills in ~/.claude/skills/ vary enormously in actual daily activation — /check-board, /process-newsletter, and /morning-prep run daily and deliver immediate loop closure; skills like /hyperframes, /cert-study, /lottie, and niche investing sub-skills may go weeks unrun. There is currently no systematic lifecycle review for the skill directory. Parrott's "Agent Ops" concept is an RDCO-native opportunity: a periodic meta-audit that applies her four tests to the skills inventory and surfaces candidates for retirement, redesign, or consolidation. This is more rigorous than the current approach of organic disuse. Her "immediate reward" criterion is also a clean design heuristic for new skills at build time: if the payoff is delayed or indirect (e.g., "I'll eventually be more organized"), the skill will be abandoned — design for immediate loop closure or don't ship.
Related
- [[2026-02-09-write-with-ai-ai-hamster-wheel]] — companion diagnosis of AI tool friction and the hidden maintenance cost of workflow churn
- [[2026-04-08-ramp-ai-adoption-playbook]] — Ramp CPO's institutional playbook for making AI workflows stick across an engineering org at scale
- [[2026-06-16-every-agent-native-tool]] — Every's own example of a workflow that stuck because it delivered immediate weekly synthesis payoff