Updated 2026-11-19 · DeepSeek V4 Hub Editorial Team

Five GitHub Agent Skill Repositories and When to Use Each (November 2026)

Say you ask a coding agent to fix an export bug. It changes the formatter, rewrites a utility, and leaves you with a larger diff than the issue warranted. A short project rule about patch scope may solve that problem. A new export feature, where several decisions need agreement before the first test, calls for more structure.

How do the five repositories compare?

Some projects offer instructions the agent loads for particular tasks. Others include project rules, coordinated workflows, or runtime hooks. The star counts are rounded figures from November 18, 2026.

GitHub star counts, November 18, 2026
RepositoryGitHub starsGood first use
mattpocock/skills323,000Clarify a design or investigate a difficult bug.
obra/superpowers317,000Take a feature through design, tests, and review.
affaan-m/ECC302,000Combine skills, hooks, and memory in an agent setup.
multica-ai/andrej-karpathy-skills233,000Narrow the scope of routine coding changes.
anthropics/skills194,000Study the structure of a reusable skill.

Which skill helps untangle one task?

The engineering skills collection is a good starting point when you can name the step that's going wrong. grill-with-docs is for settling decisions and terms with the project documentation in view. diagnosing-bugs is for a failure that needs a reproducible case before anyone changes the code.

Imagine two teams use "active subscription" differently. One team uses it to mean automatic renewal is enabled. The other uses it to mean a customer still has access during the paid period. Try the interview skill before touching the code. A useful session ends with an agreed definition and a concrete example that both teams can recognize when they review the patch.

If you want to try these workflows with a DeepSeek-backed coding agent, compare plans and usage limits on the DeepSeek Coding Plan page.

When does a full development workflow help?

Superpowers is useful when a change needs a design decision, a plan, and a way to prove that the implementation works. Its workflow covers brainstorming, planning, implementation, and review, with practices such as using a Git worktree and test-driven development. The repository documents setup paths for several hosts, including Codex and Claude Code.

Take a CSV export feature. Agree on how it handles commas, line breaks, and an empty result before implementing it. Write a test for a customer name with a comma, check the exported file, and review the diff against the agreed behavior. For a one-line label change, the same process would add more review overhead than the task warrants.

When is a broader agent setup worth it?

ECC is a collection of skills, agents, rules, hooks, and memory features. Consider it if your team repeatedly sets up the same runtime behavior. You may want a lint check after edits or a way to carry session notes into the next run. The maintainers document support by host. Check that matrix for the specific feature you need because support varies across hosts.

Try a narrow configuration first. On a test repository, enable a single check and see when it runs, what it logs, and what happens if it fails. Review hook scripts and their permissions before enabling them. A hook can run on its configured event without the agent choosing to invoke it on each occasion, so it deserves closer inspection than a paragraph of instructions.

How do you keep a small fix small?

The focused rules repository starts with one CLAUDE.md. It asks the agent to state assumptions, keep the solution simple, limit edits to the task, and give the work a verifiable finish. That is a useful baseline if your biggest problem is reviewing changes you never asked for.

For the export bug, write a test that checks whether a name containing a comma stays in one CSV field. Compare the diff to that request. If the agent changes unrelated formatters, you have a concrete review finding. To adapt the guidance to Codex or another host, put the rules in the project instruction file that host actually reads.

What should you copy when writing your own skill?

Anthropic's repository is a reference for skill authors. Start with the template, then open a skill folder to see how SKILL.md, scripts, and example resources fit together. The description tells the agent when the instructions are relevant. The body can spell out the steps and show what a finished result looks like.

For instance, a skill that checks changelogs can say when to run, where the changelog lives, and how to identify entries without an issue number. Give it a small sample file with one missing ID and one correct entry. Re-run that example after changing the instructions or switching hosts. Both versions should flag the missing ID and leave the correct entry alone.

How do you tell whether a skill helps?

Start each trial in a fresh session and a clean checkout at the same commit. Keep the model, prompt, and tool permissions fixed. Run both variants on the same task, with the rule or skill enabled in just one variant. Use the same acceptance test and repeat both variants to see how much the results vary.

If you're using a DeepSeek-backed agent, compare tool calls and token usage as well. A longer design phase may catch a missing requirement on a complex issue. On a narrow fix, it may only prolong the run. Repeat the comparison on a different task before making the workflow a project default.

What to compare across trial runs
MeasureQuestion
CorrectnessDoes the requested behavior pass a check that would fail on the starting commit?
Patch scopeCan you explain why each changed file is part of the fix?
Human effortHow many clarifications did you answer, and how much cleanup did the diff need?

FAQ

Which repository should I start with?

Start with the focused project rules if your agent routinely changes unrelated files. Try a single engineering skill if you can name the step you want to improve. Use the Anthropic template if you want to write your own skill.

Can I use the same skill in Codex, Claude Code, and DeepSeek Harness?

Instruction text can often be adapted across hosts. Check how the host discovers the skill, which tools it can use, and how it handles any scripts or hooks bundled with the skill.

Does a higher star count mean the skill will work better?

A star count indicates interest in the repository. Run the skill on a representative task and evaluate the result and review effort.