A set of SKILL.md skills addressing several common pitfalls of cheaper models identified by the author (skipping planning, claiming fixes without running them, repeatedly retrying the same bad approach, and drifting off course during long tasks). Each skill provides a short loop and a hard stop condition, and requires pasting actual command output as evidence. It can be used in environments that read SKILL.md, such as Claude Code, Codex, Cline, and OpenCode.
Stars 74
Reasoning
- Organized into five groups by purpose: persistent working modes (ds-mode); loop-based skills such as ds-verify-loop (change one thing → run real checks → read the full error → fix the root cause, up to 5 rounds, switching to ds-unstick if the same error persists after two different fixes), ds-context-reset (write PROGRESS.md and switch to a new context to continue), and ds-jury (two or three different models cross-reviewing each other); plus groups for building, delivery (ds-pr, ds-watch-pr, ds-ship), and code understanding.
- The rules are specific: tests must not be relaxed or deleted just to pass checks, and the object under test must not be mocked; ds-ship only proceeds unattended to merge when explicitly requested by the user, requiring evidence covering all changes, green CI, and resolved review comments before merging, and always asking first before deleting data, touching production, or pushing to someone else's branch.
- 10-28 created the repository; this is the first inclusion in this column; the README also provides suggestions for using DeepSeek, such as using the highest thinking intensity for planning and review, and running large tasks during low-price periods.
Usage
- npx skills add aicodeking/deepseek-stack (add -g for global installation); to install only one: npx skills add aicodeking/deepseek-stack --skill ds-verify-loop; you can also manually copy skills/* into directories such as ~/.claude/skills/ and ~/.codex/skills/.
- If you don't know where to start, run /ds-help; delivery-related skills require a logged-in gh, the ds-watch-pr script requires python3, and ds-jury, ds-best-of-n, and ds-swarm require sub-agents or a second model; if unavailable, switch to trying them sequentially in a new context.
Risks
- These are all prompt-level process constraints; "making DeepSeek work like a more expensive model" is the author's claim, and the repository contains no before-and-after comparisons or evaluations.
- ds-ship will force-push the task branch and auto-merge; confirm the repository's protection rules before enabling it; installing many skills simultaneously consumes context, and the README also recommends installing them individually as needed.