Introduction

In DeepSeek Harness or other agent development scenarios, AI often performs well in code generation but lacks engineering discipline. Common failures include claiming completion without actually writing to disk, overwriting stale state during collaboration, and repeating failures without evidence. luxueliu-agent-discipline-skills addresses these problems.

What Is It

It is a set of engineering brakes and handoff procedures for agents, containing four single-file skills. Developed by luxueliu, it is compatible with DeepSeek Harness / Claude Code / Codex. Each skill is a single SKILL.md file with zero dependencies, ready to use once installed.

Core Features

  • Pre-delivery self-check: forces a disk read-back, compares the change manifest, scans for encoding and escaping issues, and intercepts dry runs presented as real executions.
  • Course correction: stops momentum, restates the latest goal, classifies the drift type, and applies the minimal safe correction.
  • Systematic debugging: evidence-driven, minimal failing sample, single-hypothesis validation, and regression proof.
  • Multi-agent relay: re-reads state before writing to prevent stale state, marks gaps first when inheriting partial context, and distinguishes “executed” from “achieved”.

Installation and Activation

Use the following installation command:

dsh plugin --profile web add luxueliu/luxueliu-agent-discipline-skills

After installation, the four skills are copied into ~/.dsh/skills/. After restarting the session, they are automatically loaded based on their descriptions, or can be invoked by name.

Typical Usage

  • Trigger agent-course-correction: say “You are drifting” or “Stop for a second.”
  • Trigger deepcode-review: after writing files, before claiming completion.
  • Trigger systematic-debugging: ask “Find out why it failed” or “The same fix failed a second time.”
  • Trigger deepcode-collab: for multi-agent task pickup and handoff.

Use Cases and Notes

These disciplines also apply to human team code review and handoffs.

Note:

  • Some checks in deepcode-review (Windows output encoding, path validation) are illustrated for Windows; on other platforms, use equivalent checks.
  • Cross-references between skills are self-contained within this repository; if a skill is removed in isolation, related cross-references may break.

Conclusion

This skill suite provides concrete engineering disciplines that help agents avoid common mistakes. See the repository for more details: