Capability · Knowledge layer

Reusable methodology bundles. Discoverable, activatable, security-scanned.

Skills are not prompt snippets. They're inspectable, versionable operating procedures with progressive disclosure, names and descriptions first, full content when activated. Five chain composition patterns, three-layer security scanning, content-hash verification on every activation.

15
Built-in categories
5
Chain patterns
3-layer
Security scan
Hashed
Content verified
packwolf.app · Skills
Live screenshot
Skills screenshot
The skills page. Built-in catalog plus marketplace, security findings per skill, version tracking, per-agent assignment.
What it actually does

The parts that make this work.

agentskills.io-compatible.

SKILL.md format with YAML frontmatter, open standard. Authored skills move with you across compatible agents.

Progressive disclosure.

Discovery mode shows ~300 tokens (names + descriptions). Activation mode shows the full ~1500-token methodology. Default context stays small.

Five chain patterns.

Sequential, parallel, conditional, feedback loop, context fork. Compose skills into multi-step procedures without writing code.

Three-layer security scanning.

Static pattern analysis (regex for exfiltration / credential access / injection), behavioral analysis (hidden instructions / Unicode tricks), LLM semantic analysis. Hash-verified re-scan on every prompt build.

Content-hash verification.

Every skill has a SHA-256 hash. Modified content fails the hash check at runtime. Update detection surfaces diffs in a changelog before you accept the change.

A/B eval framework.

Compare two skills (or two variants of one skill) on the same task. Multi-skill, multi-model comparison with metrics. Find what actually works.

How it works

The path through skills.

  1. 01

    SKILL.md goes in the registry.

    Author or import a SKILL.md. The registry parses frontmatter, runs all three security scans, and stores the content hash.

  2. 02

    Discovery prompt is small.

    When the agent system prompt builds, only skill names and short descriptions inject. The model knows skills exist without seeing their full content.

  3. 03

    Activation injects the body.

    Agent calls use_skill or operator types /skill name. The full methodology body injects into the system prompt at Priority 3.5, wrapped in [SKILL CONTENT START]/[SKILL CONTENT END] boundary markers.

  4. 04

    Hash verifies on every activation.

    The skill's content hash recomputes and compares to the stored hash. If something modified the file out-of-band, activation fails closed.

  5. 05

    Chain composition runs.

    Skills declaring chain steps execute their pattern: sequential (step by step), parallel (race), conditional (branch on output), feedback loop (iterate until done), context fork (try variations and pick best).

  6. 06

    Eval compares variants.

    Two skills, same task, multiple model providers. The eval framework scores each run and surfaces which combination actually performed better.

Common questions

Things engineers actually ask.

Tools execute actions and return structured output. Skills shape how the agent thinks. A tool reads a file; a skill is the methodology for how to analyze that file's contents. Both have their place.

Related capabilities
CapabilityToolsCapabilityWorkflowsCapabilityEvals
Source: docs/SKILLS.md

See it in your workspace.

Closed-beta cohorts are small. Tell us what you'd want this capability to handle for your team.

Request beta access