Academic Taste Skill — Development skill for Claude Code

academic-taste

A [Claude Code](https://claude.com/claude-code) skill that distills the **academic taste** of a specific research field from its top-venue papers, then uses that taste to help you frame a research idea you have already chosen.

It is meant to run *after* topic selection, and pairs with [research-ideation](https://github.com/chendahe666/research-ideation-skill).

Core idea

**Taste is discrimination, and discrimination is learned by contrast.** Reading ten thousand average papers teaches you the average. What teaches taste is seeing, within one field, why one paper became an oral or a best paper, another an ordinary poster, and a third a reject.

So instead of reading broadly, the skill:

• samples papers by tier: award, oral, spotlight, highly cited, "poster that was later highly cited", ordinary poster, rejected;
• reads each introduction blind, with shuffled order and hidden tiers, so strong papers are not rationalized after the fact;
• turns every observation into a contrastive, counted rule, such as "high tier 11/16 vs. low tier 3/16", graded by strength.

Taste is distilled on three layers:

Layer	Question
Idea taste	What problems and insights does this field reward? Where do ideas come from?
Story taste	How is the introduction told: opening hook, how the gap is built, where the insight appears, Figure 1, contributions, how strongly claims are worded
Evidence taste	What evidence does the introduction promise, and what do reviewers accept?

One caveat is built in: **an introduction is a post-hoc story, not necessarily how the idea was actually born.** The skill cross-checks stated origins against author talks, blog posts, Test-of-Time talks and cited precursors, and labels its confidence.

Workflow

1. Scope. Narrow the field to one sub-area and one community, pick the target venues, and default to the last 3 years.
2. Build the corpus. Take a tier-stratified sample from OpenReview. Ad