deepseek-stack

Skills that make DeepSeek models (V4.1 Flash, V4 Pro) work like a much more expensive model. Cheap models usually fail in the same few ways: they skip the plan, claim a fix works without running it, retry a broken idea over and over, and drift on long runs. Each skill here closes one of those gaps with a short loop, a hard stop condition and a demand for real evidence.

It works in any harness that reads SKILL.md files: Claude Code, Codex, Cline, OpenCode, PandaCode and others.

Install

npx skills add aicodeking/deepseek-stack

Add -g to install for every project. To install one skill only:

npx skills add aicodeking/deepseek-stack --skill ds-verify-loop

Manual install: copy skills/* into ~/.claude/skills/, ~/.codex/skills/ or your harness's skills folder.

Not sure where to start? Run /ds-help.

Skills

Core

Skill	What it does
ds-mode	Always-on working style: restate, read first, small verified steps, evidence before "done".
ds-principles	21 engineering rules of thumb, each with a trigger.
ds-help	Picks the right skill for a task.

Loops

Skill	What it does
ds-plan-loop	Plan with a check per step, critique once in a fresh pass, then build one step at a time.
ds-verify-loop	Edit, run the real check, read the real error, fix the root cause. Max 5 rounds.
ds-unstick	Same failure twice: name the shared assumption, test it, pick a different approach.
ds-context-reset	Keep a PROGRESS.md and continue in a fresh context when the run gets long.
ds-ledger	One row per decision for long or unattended runs, re-read before each step.
ds-review-loop	Author plus fresh-context reviewer until both agree. Max 4 rounds.
ds-jury	Reviews from two or three different models until they agree.
ds-best-of-n	N parallel attempts, pick the best, graft in the strong parts of the rest.
ds-swarm	Fan out parallel workers and merge their results into one report.

Build

Skill	What it does
ds-sharpen	Restate the goal, make the task deeper, rewrite the prompt.
ds-blueprint	Types, signatures and module layout first, then implement.
ds-tdd	A failing test first, then the smallest fix.
ds-blast-radius	Find what a change could break outside its diff and prove it is safe.
ds-design	Build a frontend with a clear visual direction, not the generic AI look.
ds-design-audit	Score an existing UI with evidence, fix the top problems, re-check.

Ship

Skill	What it does
ds-pr	File a short PR that leads with the problem.
ds-watch-pr	Watch CI and review bots, fix real findings.
ds-ship	From change to merged PR, unattended.
ds-ready	Same as ds-ship, but stops before the merge.
ds-adopt	Finish someone else's PR or branch and keep their credit.
ds-revive	Rebase an old PR, or decide it should be closed.
ds-pr-comments	Format comments an agent posts for you.

Understand

Skill	What it does
ds-explain	How code works and why it is built that way, in plain words.
ds-catchup	What changed while you were away, and what is waiting on you.
ds-lessons	Turn repeated agent mistakes into types, lints or tests.
ds-unslop	Remove AI tells from prose and comments.

Tips for DeepSeek

• Use thinking mode at max effort for planning and review. It matters more than any skill.
• Run big jobs at off-peak hours, when the API price is lower.
• DeepSeek is cheap, so spend tokens on quality: ds-best-of-n, ds-review-loop and ds-jury cost little and catch a lot.
• Pair a stronger planner with DeepSeek as the implementer if your harness supports separate plan and act models.

Requirements

• gh (GitHub CLI), signed in, for the Ship skills.
• python3 for ds-watch-pr/scripts/watch_pr.py.
• Subagents or a second model for ds-jury, ds-best-of-n and ds-swarm. Without them, the skills run attempts one after another in fresh contexts.

License

MIT