What problem does it solve? Improving an AI diagram-authoring pipeline requires evidence, not guesswork. This Skill provides a structured experimentation loop for measuring and improving Archify's diagram accuracy, visual quality, and authoring speed, using real agent traces, fixed-input replays, and screenshot comparisons instead of anecdotal judgments. ## Core Features & Use Cases - Bounded experiment loop: Record a hypothesis, evidence criteria, and budget in a journal, then attribute failures to specific causes (placement, routing, diagnostics, instructions, or unsuitable rules) before changing the smallest responsible scope. - Benchmark replay and comparison: Render frozen first-draft benchmarks against base and head Archify revisions with replay.mjs, then capture side-by-side screenshots with shots.mjs to detect regressions and visual changes. - Trace collection and analysis: Gather per-diagram-type statistics (tool calls, finalize failures, diagnostic codes, lost content) from Devin CLI history or a normalized trace export using collect.mjs. - Use Case: After changing the workflow router, run a replay against the benchmark corpus to confirm no previously passing draft regressed, then screenshot the changed outputs to verify the whole diagram still reads well. ## Quick Start Ask the agent to resume the archify-tuning journal and run a bounded experiment comparing the current Archify revision against the recorded baseline using the replay and screenshot tools.
