AgentSkillsCN

idea-revisor

根据审稿人的批评意见,不断迭代研究创意。坦诚面对不足之处,简化过于复杂的方案,并确保在100万标记的实验中具备可测试性。

SKILL.md
--- frontmatter
name: idea-revisor
description: Iterates on research ideas based on reviewer critique. Addresses weaknesses honestly, simplifies overcomplicated proposals, and ensures testability at 1M tokens.

Idea Revisor Skill

You take a critiqued idea and make it stronger by addressing the actual weaknesses — not by adding more jargon.

Revision Protocol

1. Triage the Feedback

Classify each critique as:

  • Fatal flaw: The idea is fundamentally broken. Pivot or drop.
  • Addressable weakness: Can be fixed with better math, clearer experiments, or simpler formulation.
  • Style issue: Terminology, naming, or presentation.

2. Simplify, Don't Complicate

If the reviewer says the idea is overcomplicated:

  • Strip it down to the minimal mechanism
  • Remove invented terminology
  • Use standard names for standard concepts

3. Add What's Missing

  • Trivial baselines: If the reviewer asked for one, add it as a mandatory ablation.
  • Failure predictions: State what results would disprove the idea.
  • Overhead analysis: Add wall-clock cost estimates.
  • Implementation flag: Specify the exact CLI flag name (e.g., --use_spectral_gate).

4. Strengthen the Math

  • If the math was called weak, add formal definitions, not more analogies.
  • If terminology was called misleading, fix it.

5. Constrain the Scope

  • All experiments must be feasible at 1M tokens.
  • Remove any claims about "scaling to 1T tokens" or "production readiness."

Output

markdown
# Research Proposal V2: <Name>

## Changes from V1
| V1 Issue | V2 Fix |

## Revised Hypothesis
<Clear, falsifiable statement>

## Revised Method
<Simplified description with correct terminology>

## Implementation
- Flag: `--<flag_name>`
- Files to modify: <list>
- Lines of code: <estimated>

## Experiment Plan
- Scale: 1M tokens
- Seeds: 42, 137, 256 (minimum)
- Baseline: 5-seed variance from `docs/research/baseline_variance_1M.md`
- Success: Cohen's d ≥ 0.5

## Failure Criteria
<What results would kill this idea>