When apparent productivity plateaus coincide with stable use of high-performing reusable workflows, how do different coaching targets and triggers—such as coaching power users on portfolio-level refactors versus coaching individual contributors on local parameterization and step edits, triggered by patterns like recurrent near-misses or variant sprawl—differ in their impact on lifting team-level workflow maturity and preventing long-run stagnation in the AI learning curve?

anthropic-learning-curves | Updated at

Answer

Power‑user portfolio coaching and IC local coaching help in different ways and should be triggered by different plateau signals.

  1. Relative impact on team workflow maturity
  • Power‑user portfolio coaching
    • Best for: raising the ceiling of team workflows.
    • Focus: merge overlapping workflows, define params, codify best variants, align with business cadences.
    • Effect: fewer, stronger shared assets; faster propagation of improvements; higher baseline quality for all users.
  • IC local parameter/step coaching
    • Best for: reducing friction and keeping ICs engaged.
    • Focus: turn recurrent edits into params, tweak steps, add small pre/post steps.
    • Effect: better fit to local tasks; higher reuse of existing assets; smoother AI learning curve for more people.

Net: portfolio coaching moves asset quality and structure; IC coaching moves coverage and fit. Both are needed to avoid stagnation.

  1. Triggers and recommended coaching by pattern

  • Recurrent near‑misses (same step often corrected, outcomes "almost right")

    • Power‑user coaching: update the shared workflow step, add guardrails/examples, possibly add a param that captures the recurring correction.
    • IC coaching: in‑run tips like “save this change as a default” or “promote this tweak into the workflow.”
    • Expected impact: lowers correction rates across many runs; converts improvisation into reusable structure.

  • Variant sprawl (many similar workflows/variants with small differences)

    • Power‑user coaching: consolidate into a few parameterized workflows, retire low‑run variants, define naming and ownership.
    • IC coaching: teach users to use params/toggles instead of forking, plus light guidance on when a true new variant is warranted.
    • Expected impact: concentrates learning in fewer assets; makes plateaus more visible and fixable.

  • Stable high performance but zero edit activity

    • Power‑user coaching: periodic portfolio reviews to check for drift, new use cases, and cross‑tool integration; add new inputs/outputs if needed.
    • IC coaching: small nudges to try attachments, new data sources, or sharing when patterns suggest under‑use rather than true saturation.
    • Expected impact: prevents silent obsolescence; nudges expansion without forcing churn.
  1. When each coaching focus helps most against long‑run stagnation

  • Favor power‑user portfolio coaching when:

    • A few workflows dominate runs.
    • Edits and variants are already concentrated in a small group.
    • Plateaus coexist with variant sprawl or repeated near‑misses across many users.
      → Main lever: restructure and harden the shared assets.

  • Favor IC local coaching when:

    • Many users make similar manual tweaks but rarely save them.
    • Near‑misses are localized to specific teams or clients.
    • Variant sprawl is light, but runs show repeated param‑like edits.
      → Main lever: convert everyday edits into parameters/steps.

  • Combine both when:

    • Plateaus sit alongside both heavy variant sprawl and many local corrections.
      → Coach power users to rationalize the portfolio, while coaching ICs to express needs as params/feedback instead of silent forks.
  1. Implications for preventing AI learning‑curve stagnation
  • Power‑user coaching mainly shapes the team curve: bigger step‑changes from occasional refactors.
  • IC coaching mainly shapes the within‑user curve: smoother progress and less drop‑off when workflows almost fit but not quite.
  • Systems that prioritize portfolio‑level refactors at clear plateau signals, and support ICs with low‑friction parameter/step coaching on near‑misses, are most likely to maintain upward movement instead of flatlining once first strong workflows are in place.