Interpretation Checkpoint — Verify the Read Before Writing

intent-capture pins what was asked; plan-first decides how. Neither catches drift in the fine detail — the exact values, files, and parameters I inferred from a clear request. On wide-blast-radius tasks that drift is expensive to discover after the edit; this catches it before.

Self-improvement (do this first and last)

1. At start: read learnings.md in this skill's folder if it exists. Apply relevant lessons.
2. At end of every use: append one dated bullet — did this catch real drift, or fire on a task that didn't need it? Merge instead of duplicating; delete disproven bullets.

When it applies

Both must be true:

• Files affected: 2+
• Distinct parameters/aspects changed: 2+ (not the same single change applied identically across files — e.g. a site-wide rename is 1 parameter regardless of file count)

Runs after intent-capture/plan-first (once goal and approach are settled), before the first Edit/Write. Below this threshold, the regular flow catches mistakes cheaply enough — skip straight to the work.

The format

Files: <path> — <why it's touched>
Changes: <param/aspect> → <specific value or behavior intended>  (one line per distinct change)
Assumptions: <anything inferred rather than stated>
Non-goals: <adjacent things deliberately left alone>

Present this, then wait for correction or confirmation before editing — a single-turn round trip, not a full plan-mode detour.

Rules

• List real files (checked via glob/grep), not a guess from memory.
• One line per distinct change — if there's only one after listing them out, this skill didn't need to fire; say so and proceed.
• Assumptions must be visible, not folded silently into Changes — that's the whole point of the checkpoint.
• Not a substitute for plan-first's Approach/Verify-by — this only checks the parsed detail, not the strategy.

Failure mode this prevents

A multi-file, multi-parameter change built correctly against a misread detail (wrong value, wrong file, wrong scope) — discovered only after generation, when the fix costs a redo instead of a one-line correction.