Skill Grader / report card

microsoft/debugpy

Graded B on context cost and craft. 189 tokens of this skill ride in the system prompt of every message, whether or not it fires (light footprint). Last graded Sep 9, 2026; re-graded weekly from HEAD.

View on GitHub ↗Share the grade Grade another skill

B overall · 3.25/4

github.com/microsoft/debugpy@HEAD · 7 skills · 189t resident (light) · 559t per activation · 2,953t bundle

Resident footprint

189t of description text ride in the system prompt of every message across 7 skills (paper's observed per-skill range: 50–280t).

Description honesty

Defensive padding detected: 18 comma-separated trigger keywords; 90 words. Padding is individually rational and collectively ruinous — it dilutes every co-installed skill.

Body size

Largest body is 303 words (~559t), loaded on every trigger. Corpus median 921, p90 2207.

Progressive disclosure

Single-file skill small enough to need no reference layer.

Factoring

Scoped to a coherent procedure set.

CLI leverage

Prompt-only. Every rule stated in prose pays attention rent; consider a companion CLI for deterministic steps.

B overall (3.25/4). 7 skill(s), 189t resident on every message (light), 559t loaded per activation. Weakest axis: cli leverage (C).

Put the grade in your README

The badge shows the current letter grade and links back to this page. It re-renders from the latest grade, so it never goes stale when you improve the skill.

Markdown

[![Skill Grader: B](https://seoagent.com/skill-grader/badge/microsoft/debugpy.svg)](https://seoagent.com/skill-grader/microsoft/debugpy)

HTML

<a href="https://seoagent.com/skill-grader/microsoft/debugpy"><img src="https://seoagent.com/skill-grader/badge/microsoft/debugpy.svg" alt="Skill Grader: B" height="20"></a>

How this was graded

Six deterministic axes from “@skills: Attention Is All You Have” (Yin et al., 2026) and the open seo-skill-bench footprint scorer: resident footprint, description honesty, body size, progressive disclosure, factoring, and CLI leverage. Tokens are approximated as chars/4. No LLM, no signup, same result every run. Read the full methodology.