Skills Field Notes

Claude Code Skills tested, not listed

Every review answers three questions: what it solves, who should skip it, and where it breaks. No directories. No star-count rankings.

Start with install
2skills run against a baseline
15skills from a single install
≈10sto install fifteen (5 runs, 8.6–12.4s)

How much of a skill the agent sees

Before you finish this review, the agent has only read the first block below. The body loads when it decides to use the skill.

Two fields in the frontmatter are the whole calling card:

name: superpowers
description: >
  A complete software development methodology and skill
  composition system, ported from obra/superpowers.
  Provides structured, repeatable workflow: brainstorming,
  design, planning, TDD, subagent-driven dev, code review,
  finishing branch. Use for coding, building projects,
  refactoring, bug fixing, or any programming tasks.

Read from the local install at .dsh/skills/superpowers-skill/SKILL.md, 2026-10-19.

The body is a seven-phase workflow with no skippable step: brainstorm, design, plan, TDD, parallel implementation, review, finish. Each phase has an artefact and a checkpoint. That is the value and that is the cost.

The body stays in context for a few turns after invocation. The full comparison is in the Superpowers review.

Separating what the agent reads from what you read is the only reliable way to judge whether a skill will fire, and whether it deserves the space.

Three questions, by search intent

  • What a skill actually is

    The definition and the boundaries A folder with a SKILL.md that says when to use it and how, plus optional scripts. Not a plugin, no network required. It is a way to freeze a piece of judgement so the agent repeats it.

  • Which ones earn their place

    Pick by task Twelve repos, two of them installed and measured against a baseline. The rest are judged from the repository alone, and every entry says which is which.

  • Why it stopped working

    The error list Almost never a broken skill. Wrong directory, an open session that never reloaded, or frontmatter that parses to nothing and kills automatic triggering.

Run recently

Install one in five minutes

The cross-agent installer is skills from vercel-labs. It supports Claude Code, Codex, Cursor, Windsurf and 40+ others.

npx skills add obra/superpowers

Restart the session afterwards — the skill list is read at startup. All three blocks below are real output from this machine on 2026-10-19.

succeeded · 12.4s (this run)
o  Source: https://github.com/obra/superpowers.git
•  Cloning repository…
o  Repository cloned
o  Found 15 skills
•  Installing all 15 skills
—  Done!  Review skills before use;
   they run with full agent permissions.

Excerpt — the per-skill receipts are omitted. Fifteen directories in 12.4 seconds. Do not skip the last line: skills run with full agent permissions.

.claude/skills/
brainstorming/SKILL.md
test-driven-development/SKILL.md
systematic-debugging/SKILL.md
writing-plans/SKILL.md
subagent-driven-development/SKILL.md
requesting-code-review/SKILL.md
using-git-worktrees/SKILL.md
…15 directories total

Real directories with real SKILL.md files, not symlinks. Count them to know whether the install was complete.

run again · 10.0s
o  Found 15 skills
•  Installing all 15 skills
   overwrites: Claude Code
—  Done!  Review skills before use

A second install reports no conflict and overwrites in place. If you edited a SKILL.md locally, do not run it again.

Decide two things before installing: the scope (project or user) and the agent. The installer asks, or take --agent.

The full error table and path reference live in the install guide.

When the install fails

The first failure here was not a firewall. A git proxy pointed at a dead port, and fixing that one setting ended it.

failed after 21s
failed to clone repository
fatal: unable to access
  'https://github.com/obra/superpowers.git/':
  Failed to connect to github.com port 443
  via 127.0.0.1 after 2055 ms:
  Could not connect to server
—  Installation failed

via 127.0.0.1 in the error means the request never left the machine.

Two fixes, pick one:

# which port is git actually using?
git config --global --get http.proxy
# point it at the live port, or drop it
git config --global http.proxy 127.0.0.1:7897
git config --global --unset http.proxy

If the message is Recv failure: Connection was reset, the direct connection is being reset — a network problem, not the installer. Re-run once you have a working egress.

Asked often

Installed, but nothing happens?

Three places, in order of how often they are the cause. First, the location: project scope is .claude/skills/ and user scope is ~/.claude/skills/ — installing into the wrong directory is the same as not installing. Second, the session: the skill list is read when a session starts, so an already-open session will not see it. Third, the frontmatter: if the YAML fails to parse, the body still loads and every field is empty, so automatic triggering silently stops working and only a typed /skill-name reaches it.

Skills or MCP?

Decide by what is missing. A skill supplies a procedure or a judgement call and needs no network. MCP supplies access — files, databases, external services. If the agent needs data it cannot see, MCP is the only way in. If it needs to work differently, a skill is lighter. They are not alternatives.

How much context does one skill take?

One line by default. Progressive loading means only the name and description from the frontmatter sit in context each turn; the body is read when the skill is actually invoked. Install dozens and the descriptions alone start to compete for budget — the docs say the description list scales with the context window and the least-used entries are dropped first when it overflows.

Why does the install fail?

Usually not the firewall — a git proxy pointing at a port that is no longer listening. On this machine git was configured for 127.0.0.1:7890 while the live proxy was 7897, and the error read Failed to connect to github.com port 443 via 127.0.0.1. Pointing http.proxy at the live port installed 15 skills in 12.4 seconds with the same command.

Judgement, not a directory

Bookmarking a skill takes a second. Deciding whether it survives a real task takes a day. This site only does the second thing.

Pick by task