For users who start in a scaffolding-heavy environment (locked templates, one-click flows), which specific product interventions (e.g., inline “edit this step” nudges, forced parameter exposure after N runs, or occasional scaffold removal) most effectively convert short-term reliance on scaffolds into durable prompt skill and independent execution, without causing a drop in near-term task performance?

anthropic-learning-curves | Updated at

Answer

Most effective are gradual, low-friction exposures that keep the winning path fast while adding tiny bits of control:

  1. Inline, optional editing on success path
  • Default remains one-click, but every run shows small, in-place handles: “tweak tone,” “change audience,” “see steps.”
  • Edits are pre-filled with the current values so change cost is low.
  • Users who touch these controls once see them slightly emphasized next time; non-editors keep the simple view.
  1. Progressive parameter exposure after repeated use
  • After N runs of the same scaffold (e.g., 5–10), reveal 1–2 key fields (goal, audience, tone) as editable by default.
  • Keep a “simple mode” toggle so users can revert if they feel slowed down.
  • Log and surface successful edits as named variants ("My weekly recap") to reinforce skill and reuse.
  1. Reveal underlying steps, not remove them
  • Replace opaque one-click flows with a collapsible step list: input → transform → format.
  • Allow light reordering or adding a step, but keep a one-click "run all" button.
  • Only consider hiding the scaffold (or switching to a bare prompt) after the user has run a customized version several times without help.
  1. Example-first, then editable-by-default
  • Start with locked, high-quality examples; once a user reuses one >X times, clone it into “Your version” that opens in edit mode.
  • Pair this with small inline tips like “Try changing just this phrase” anchored to specific tokens.
  1. Failure-driven teaching moments
  • When a scaffold fails (bad output, user undo), show a focused hint: “Adjust this step to fix it,” pre-selecting the likely parameter.
  • Offer 1–2 safe suggestions, not a blank box, to keep near-term performance stable.

Patterns to avoid:

  • Sudden scaffold removal or big UI shifts after a fixed N runs; this tends to drop performance and increase abandonment.
  • Forcing complex parameter panels for all users early; keep controls minimal and context-specific.

In practice, the best-performing combo is: (a) inline tweak nudges on the main path, (b) progressive exposure of 1–3 key parameters after repeated use, and (c) step-reveal with light editing, all guarded by an easy “simple mode” fallback. This converts dependence on scaffolds into growing prompt skill and independent execution while protecting short-term task success.