Current AI learning-curve models focus on repeated success with a fixed reusable workflow as the main unit of progress. If instead we treat users’ ability to rewrite or retire their own workflows—deliberately discarding templates, simplifying chains, or folding AI steps back into manual processes—as a separate stage of workflow maturity, under what conditions does high “deletion and redesign” activity signal healthier long-run adoption than stable reuse, and how would an adoption model that treats workflow churn as potentially positive change the way we detect productivity plateaus and design onboarding or coaching around them?

anthropic-learning-curves | Updated at 2026-04-07 11:15

Answer

High deletion/redesign is a healthy sign when it looks like strategic pruning and refactoring, not flailing.

Task / environment
- Tasks change often (policies, formats, tools).
- Users face many edge cases; no single flow covers most work.
- Org tolerates experimentation and small process changes.
Behavior patterns that indicate healthy churn
- Churn is clustered: old workflows removed near new ones appearing with overlapping scope and fewer steps.
- Successor flows show better metrics (fewer corrections, broader inputs/outputs, higher team use).
- Users delete low-usage or redundant flows, keep a small active set.
- Some AI steps are intentionally folded back into manual work where they were low value or risky, with cycle time or quality improving.
- Naming, tagging, and sharing practices survive across redesigns (same concept, better implementation).
Behavior patterns that suggest unhealthy churn
- Many short‑lived flows without clear successors; task still handled mostly ad hoc.
- Corrections and cycle time stay high or worsen after redesign.
- Users oscillate between AI and manual with no stable pattern.
- Deletions follow visible failures or policy pushback, not consolidation.

Distinguish “static reuse plateau” vs “refactor plateau”
- Static: stable runs, low edits, narrow inputs/outputs, low sharing, low churn → likely true ceiling.
- Refactor: rising or pulsed churn + emergence of leaner or more team-used flows → likely hidden progress.
Plateau signals with churn in mind
- Healthy consolidation: drop in number of active workflows + rise in usage per remaining workflow; delete events co-occur with new, better flows.
- Risky stall: repeated creation + deletion around same task with no improvement in quality, time, or coverage.
- Maturity signal: intentional retirement of AI steps where humans are better, coupled with increased AI use on higher-leverage steps elsewhere.

Normalize and scaffold redesign
- Treat “archive/clone/redesign” as first-class actions in the UI.
- Offer light refactor wizards when users clone or repeatedly tweak a flow: “Merge these variants?” “Simplify this chain?”
- Show history across versions so deletion feels like evolution, not loss.
Trigger coaching on churn patterns
- High healthy churn: suggest abstraction (parameters, modular subflows, shared assets).
- High unhealthy churn: suggest examples, domain templates, or power-user help; highlight a few stable patterns to adopt.
- Stable use + high corrections: nudge toward redesign instead of endless patching.
Rethink success metrics
- Track “effective lineage” (families of related workflows over time) rather than raw workflow count.
- Reward reduction of redundant flows, increased coverage per lineage, and cleaner chains, not just stable reuse.
- Count “successful retirement” (work moved to a better AI or manual pattern) as positive progress.

A plateau is less about “no new workflows” and more about “no structural change.”
Sudden bursts of deletion + creation can be early signs of users breaking local optima.
True plateaus show both flat use and flat structure; refactor bursts show flat or slightly worse short-term metrics but better structure and usage patterns afterward.