timefiles404/lean-mode-skill

This page was translated automatically. Technical details may differ from the original. View the original

81 stars Quality 59 Trend 23MIT

Moderation engineering skills: When to write defensive code, when not to, and how to reduce a round of build testing from tens of minutes to tens of seconds. SKILL.md, can be used independently.

Overview

1. — When to add checks / try-catch / abstraction layers / safety nets, and when not to. 2. — How to reduce build test time from tens of minutes to tens of seconds. skills/lean-mode/SKILL.md is the standard frontmatter(name + description) + Markdown body, and any agent that supports SKILL.md convention can use it. When large models write code, there are two systemic tendencies, which become more pronounced in the latter half of the long conversation: ** and **. Here are the numbers measured from real conversation records: : All of the above comes from **records—a Java/Gradle project of about 20k lines, modeled after GPT-5.6, and a long session. It shows that such behavior really exists and is of a large scale. **Cross-model or cross-project benchmark statistics, please do not use them as baselines. These aren't 'poor model capabilities,' but **. The rules written into system prompts are overwhelmed after dozens of rounds. : The counterexamples in sections one, two, and three provide four ways to write (the same error looks different in different languages; Go's 'swallow error returns zero' and Java's catch → return false are the same thing). Section seven provides minimum validation loop commands for building systems. The search patterns in the appendix are grouped by language. (Vantaloom: skills page → pulled from URL) — just paste the repository address, and it will automatically find it under skills/lean-mode/: : Create a new skill and paste the entire content of skills/lean-mode/SKILL.md into it.

Overview (machine translated)

A SKILL.md for coding agents, answering two questions:

Is this ‘security’ code worth writing—when to add checks/try-catch/abstraction layers/bottom lines, and when not to.
Why This Round Is So Slow — How to reduce build testing time from tens of minutes to just tens of seconds.

It’s not the slogan of ‘write less code.’ Each one has clear criteria and explicit exceptions.

skills/lean-mode/SKILL.md is the standard frontmatter(name + description) + Markdown body, Does not depend on any specific host; any agent that supports SKILL.md conventions can be used.

What is it addressing?

There are two systemic tendencies when writing code with large models, which become more severe in the latter half of the session: overly defensive programming and inefficiency Verification cycle. Below are the numbers measured from a real conversation record:

Data Sources and Boundaries: All of the above comes from the same record—a Java/Gradle project of about 20k lines, model is GPT-5.6, a long conversation. It shows that such behavior really exists and is of a large scale, not cross-model or cross-engineering Benchmark statistics, please do not reference them as baselines.

These aren’t ‘model skills are inadequate,’ but no one tells it to moderate itself in every round. Rules written into system prompts, dozens of rounds After that, it was drowned out.

What’s inside?

Section 8 + One Audit Process:

Section	In a word
1. Validation is done only once at the trust boundary	There are only four places at the trust boundary; Data crossing this line is trustworthy.

This page was translated automatically. Technical details may differ from the original.

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add timefiles404/lean-mode-skill

New to MCP? Start with the mcpskill.org guide →