SkillFoxx

deepseek-stack skill

Skills for DeepSeek: short plan, verify and review loops, each with a hard stop condition and real command output as proof.

  • 267 ★ on the repository
  • Last commit
  • MIT
Skill

Broad access

Why this level

  • `/ds-ship`, on a direct request, force-pushes its own branch with `--force-with-lease` and merges the pull request without your review once the checks pass.
  • The Ship skills act through your GitHub CLI login: they open pull requests, post comments and push branches.
All reasons and checks

aicodeking/deepseek-stackAuthor:aicodeking

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/deepseek-stack

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Broad accessThe tool can change data, run commands or handle secrets.

  • `/ds-ship`, on a direct request, force-pushes its own branch with `--force-with-lease` and merges the pull request without your review once the checks pass.
  • The Ship skills act through your GitHub CLI login: they open pull requests, post comments and push branches.

This is third-party code. Review the repository files before installing.

What it does

deepseek-stack is a set of 29 skills that close the usual gaps of inexpensive models: skipping the plan, claiming a fix without running it, retrying a broken idea endlessly and losing the thread in long sessions. Each skill sets a short loop with a stop condition and demands evidence in the form of command or test output. The core skills are `/ds-mode` (a working style: restate the task, read code before editing, small verified steps), `/ds-principles` (21 engineering rules) and `/ds-help` (picks a skill for the task). The `/ds-verify-loop` cycle is capped at five rounds, and `/ds-review-loop` pairs an author with a fresh-context reviewer for up to four rounds. The `/ds-ship` skill carries a change all the way to a merged pull request. The set works in any agent that reads `SKILL.md` files: Claude Code, Codex, Cline, OpenCode.

Who it is for. Developers who run DeepSeek V4.1 Flash or V4 Pro in Claude Code, Codex or Cline for the price and want the same verification discipline as expensive models give.

Good fit when

  • A cheap model claims a fix without running tests, and you need a loop of editing, running the real check and reading the real error.
  • The agent repeated one failed idea twice, so it is time to name the shared faulty assumption and change the approach.
  • An overnight run lasts hours and the context has to be reset with the state saved in `PROGRESS.md`.
  • A change must reach a pull request and a green CI run, with the merge left to you.

Not a fit when

  • The task is tiny, such as a typo fix, and multi-step loops would only slow it down.
  • Merging a pull request without your review is unacceptable: `/ds-ship` merges on its own and should be enabled only on a direct request.
  • The GitHub CLI `gh` is missing or signed out, and the Ship skills depend on it.

Example request

Fix the failing test with `/ds-verify-loop` and show the command output after which the task can be considered solved.

Limitations

The Ship skills (`/ds-pr`, `/ds-watch-pr`, `/ds-ship`, `/ds-ready`) require a signed-in GitHub CLI `gh`, and `/ds-watch-pr` runs the `watch_pr.py` script under `python3`. `/ds-jury`, `/ds-best-of-n` and `/ds-swarm` need subagents or a second model, otherwise attempts run one after another in fresh contexts. The author recommends thinking mode at max effort for planning and review. The set is written for DeepSeek V4.1 Flash and V4 Pro, and the README does not describe results on other models.

How to disable. Remove the `ds-*` skill folders from the agent skills directory: `~/.claude/skills/`, `~/.codex/skills/` or your agent folder.

FAQ

How do I install just one skill?

Run `npx skills add aicodeking/deepseek-stack --skill ds-verify-loop`, replacing `ds-verify-loop` with the name you want. The `-g` flag installs it for every project.

Where should I start?

Run `/ds-help`: it picks a skill for the task. `/ds-mode` switches on the always-on working style.

What does /ds-verify-loop do?

The agent edits code, runs the real check, reads the real error and fixes the root cause. The loop is capped at five rounds.

Are subagents required?

`/ds-jury`, `/ds-best-of-n` and `/ds-swarm` need subagents or a second model. Without them attempts run one after another in fresh contexts.

Does the set work outside DeepSeek?

The skills are written for DeepSeek V4.1 Flash and V4 Pro, but any agent that supports `SKILL.md` reads them: Claude Code, Codex, Cline, OpenCode and others.

Discussion

No questions or reviews yet.

Skills

A code review skill with four independent probes, evidence-backed findings and a merge gate derived from JSON

SkillExtended access479

Catalog of 46 versioned skills for code review, releases, CI and landing pages, installed as the lstack plugin or through npx.

PluginExtended access145

A skill that finds AI slop in code: duplicate helpers, empty abstractions, weak tests. It cleans one pattern per commit.

SkillExtended access95

A cleanup skill for agent-written code: finds surplus tests, self-checks and fallbacks and deletes them, keeping behavior

SkillExtended access140

A skill that runs a full refactor so agents read code more easily: it finds hidden links and needless dependencies.

SkillExtended access1.1K
Editors’ pick

The /improve skill: a strong model audits a repository and writes self-contained fix plans, and you hand the execution to a cheaper model

SkillExtended access11.4K

Comparison with similar tools

Cards that share the main search query, and how they differ.

ToolOfficialRepository starsMain difference
deepseek-stack (this card)no267
DeepSeek Harnessyes280,265an open DeepSeek agent built from plugins; deepseek-stack is a skill set that runs inside other agents
codex-deepseek-subagentno394wires DeepSeek in as a desktop Codex subagent; deepseek-stack sets the working loops for the model itself
Superpowersno316,679a general development process for any model; deepseek-stack adds loops with stop conditions aimed at the weaknesses of cheap models

README in short

The README groups the skills into five sets (Core, Loops, Build, Ship and Understand) with a table describing each one. It ends with DeepSeek tips such as spending cheap tokens on quality and running big jobs at off-peak API prices.

README badge

Original SKILL.md

---
name: ds-mode
description: "Always-on working style for DeepSeek agents: restate the task, read before editing, small verified steps, evidence before done, lean context, short reports. Use when starting any coding task, when the user types /ds-mode, or when work starts to drift."
---

# DeepSeek working mode

Apply this on every task. It keeps the loop tight and the claims honest.

## Before you act

1. Restate the task in one line. Say the goal and the smallest change you expect to meet it. Do not wait for approval.
2. Read the code you will touch before you edit it. Find every caller of a function before you change its behavior.
3. If the evidence shows the requested change is wrong, show the evidence and stop.
4. For work with more than two steps, write a short numbered plan. Each step ends in something you can check. Use ds-plan-loop for anything larger than a few files.

## While you work

- Make the smallest diff that solves the problem. Delete before you add. Do not add layers, options, or helpers that nothing needs yet.
- Make one change at a time. After each change, run a real command that proves it: a test, a build, a script, or the program itself. Use ds-verify-loop for the check.
- Fix the root cause, not the symptom. Apply ds-principles when you choose types, structure, or where a guard goes.
- Proceed on reversible work without asking. Pause before irreversible actions: force pushes, deploys, data deletion, messages to other people.
- If the same fix fails twice, stop. Do not try a third variant of the same idea. Switch to ds-unstick and test the premise behind the failed fixes.
- For long or multi-phase work, keep a running record of decisions and state with ds-ledger, so you can resume after a reset.

## Keep context lean

- Read files in the parts you need, not whole when a range will do.
- Keep summaries in your context, not raw output. Trim long logs to the lines that matter.
- Send bulk reading, wide searches, and large fan-out to subagents when the harness has them. Ask each for a short conclusion with file paths.
- Do not reread a file you just edited. Do not repeat a search you already ran.

## Before you say done

- Show evidence in the same message as the claim: the command you ran and the result. "It compiles" is not proof that it works.
- Label each claim as measured, inferred, or guessed.
- If a check failed or you skipped one, say so plainly.
- Run ds-review-loop on nontrivial diffs before you report.

## The final report

- Open with the goal in one line.
- List what changed, with exact file paths.
- List the checks you ran and their results.
- List open issues or blockers.

Report a problem