skillnet

---
name: skillnet
description: >-
  Search, download, create, evaluate, analyze and route reusable agent skills with SkillNet.
  Use when asked to find or reuse a skill, turn a repo/document/trace into a skill,
  assess skill quality, build a local skill graph, select skills for a task,
  or fill a clearly identified capability gap with a reusable skill.
  中文：搜索、下载、创建、评估、分析与路由技能。不用于普通代码修改或仅阅读文档。
metadata:
  version: "0.1.1"
  requirements: "Python 3.10+, skillnet-ai 0.1.1+, network; model API for create/evaluate/analyze; graph extra and embedding API for analyze/route; Claude or Codex SDK for route"
---

# SkillNet

Find a useful skill, bring its complete resources into the workspace, create and
evaluate reusable skills, or analyze and route a local library through the SkillNet SDK.

## Choose the workflow

- **Analyze and route:** build a local skill graph, then select skills for a task.
  Read [routing.md](references/routing.md) for separate model and SDK configuration.
- **Find:** search and return relevant candidates with URLs and tradeoffs.
- **Reuse or install:** inspect relevant skills already available, then search →
  select → download → inspect → apply when the user requested its use.
- **Create:** choose one source → create through SkillNet → check structure → evaluate.
- **Evaluate:** evaluate the requested local skill or GitHub skill URL and explain
  the findings in terms of what the user wants to do.

Keep the user's original task in focus. Ordinary coding and document reading do
not require a SkillNet search. An explicit creation request is sufficient reason
to create a skill; it need not meet an additional complexity threshold.

## Runtime and configuration

Use the installed `skillnet` CLI (version 0.1.1+). Prefer `--json` for reliable paths,
URLs and error information. If the command is missing, outdated, or its interpreter
is uncertain, read [setup.md](references/setup.md). `python -m skillnet_ai` is an
alternative when using the Python environment that actually contains the SDK.

Search and public download need no model API key. Before creating or evaluating,
reuse the user's existing configuration. If it is missing, read setup.md and help
configure **API key + model API base URL + model name** once. `skillnet doctor --json`
shows configuration sources and missing fields without revealing keys. It makes
no network request unless explicitly asked.

Agent login and model API authentication are separate. Use an endpoint compatible
with Chat Completions; do not extract credentials from an agent's login files.

Resources linked here are relative to the directory containing this SKILL.md,
not the shell's working directory. Keep generated outputs in the user's workspace.
For persistent skill installation and host discovery, read
[platforms.md](references/platforms.md).

## Find and use a skill

Start with a focused keyword and a small result set:

```text
skillnet search "pdf" --limit 5 --json
```

If those candidates are irrelevant, make at most one semantic query with a short
description of the missing capability:

```text
skillnet search "extract tables from scanned financial reports" --mode vector --limit 5 --json
```

Choose by task fit, dependencies and reported limitations. Stars and stored
assessments are useful signals, not proof that the skill will work. With no useful
candidate, report that briefly and continue the user's task.

Download the selected `skill_url` into a task-local folder, or the active host's
skill directory when the user wants it installed:

```text
skillnet download "<skill_url>" --target-dir "./downloaded_skills" --json
```

Use the returned `data.path`; do not reconstruct paths from names or table output.
Download includes scripts, references and other files. Read SKILL.md, then inspect
only the resources and operations relevant to the task. Use the skill within the
user's authorization and the host's permissions. Loading or reading it does not
require a second approval when the user already asked for its use.

An existing destination is preserved by default. Use `--overwrite` only when an
update/replacement is intended. Do not silently delete an existing skill.

## Create and evaluate

Tell the user that creation includes a model-based evaluation, then run:

```text
skillnet create --prompt "A skill for checking CSV headers and reporting missing columns" --output-dir "./generated_skills" --evaluate --json
```

Use `--no-evaluate` when the user explicitly wants to skip evaluation. For repository,
document or trajectory input, read [workflow-patterns.md](references/workflow-patterns.md).
The CLI uses the real SDK, checks generated structure before evaluating, and returns
paths and per-skill results. Keep SDK errors visible instead of substituting
host-written output and calling it a SkillNet result.

Evaluate an existing skill with:

```text
skillnet evaluate "./generated_skills/example" --json
```

Explain the five dimensions and actionable weaknesses. Model evaluation does not
execute the skill by default and does not prove the user's task succeeds. When
appropriate, exercise the skill on a small representative input and inspect the
actual output. Avoid automatic regeneration loops or unrequested batch evaluations.

## Results and recovery

A successful command has `ok: true`. A failed command exits nonzero and includes an
error and, where available, partial `data`. If creation succeeded but evaluation
failed, preserve `data.paths` and retry **evaluation**, not creation.

Report what was found/created, its real path, any validation or quality issues,
and whether it was exercised on the user's task. Distinguish downloaded files,
valid structure, completed model evaluation, and verified task behavior.

Read [api-reference.md](references/api-reference.md) for flags, output fields,
provider compatibility and error recovery. Read
[security-privacy.md](references/security-privacy.md) when handling private inputs,
credentials or unfamiliar third-party operations.