← ClaudeAtlas

plot-paper-experimentslisted

Create publication-ready machine-learning experiment figures from W&B or local tabular data, especially RL training/evaluation curves, ablations, scaling plots, and benchmark bars. Use when Codex must retrieve run histories, compare methods, establish paper-facing plot semantics, render conference-quality figures, or preserve reproducible figure data and metadata.
vanzll/ai-research-accelerator · ★ 1 · AI & Automation · score 74
Install: claude install-skill vanzll/ai-research-accelerator
# Plot Paper Experiments Create honest, compact figures whose visual language is close to strong diffusion-RL papers. When installed, use `wandb-query` for W&B retrieval and `scientific-visualization` for rendering details; otherwise use exported local tables and the project's plotting stack, and disclose the missing integration. Read [references/visual-contract.md](references/visual-contract.md) before proposing or producing a figure. ## Workflow ### 0. Treat user review as template training data This skill is expected to improve through repeated author review. After every user critique or approval: 1. Separate figure-specific preference from a reusable plotting rule. 2. Apply the requested change to the current figure only after updating and reconfirming the plot contract when names or colors change. 3. When the feedback expresses a reusable preference, update this skill or its visual reference in the same turn. 4. Preserve prior approved rules unless the latest feedback explicitly supersedes them. 5. Validate the skill after every edit and state which reusable rule changed. Do not overfit one data-specific request into a universal rule. For example, a particular y-limit belongs in figure metadata; the rule that no traversed interval may be omitted belongs in this skill. ### 1. Establish the evidence 1. Identify the scientific claim, target venue, final column width, benchmark, methods, seeds, x-axis budget, and train-versus-eval semantics. 2. Read each W&B run's co