A model-agnostic Agent Skills library for explicit routing, bounded context, and verifiable agent behaviour.
Small skills over standing prompt weight. Composition over monolithic instructions. Explicit routes over accidental activation.
Install · Plugin · Documentation · Quick Start · Roadmap · Capabilities · Routing · Verification
Created by Adrian Paredez, with Sole as his sovereign cognitive work partner.
Capability Fabric is an agent-improving skill layer for runtimes that support the open Agent Skills standard. It gives an agent a small set of process skills for planning, research, context budgeting, sustained execution, verification, code quality, autonomy, and skill authoring.
The library is deliberately narrow. Each skill has one responsibility, a portable
SKILL.md, and supporting references loaded only when needed.
Important
Capability Fabric changes the agent's operating method, not a single task domain. Install it once, then state the task. The runtime selects matching skills from their names and descriptions, and loads the full body only when needed.
Install into every detected agent runtime:
git clone https://github.com/AdrianParedez/capability-fabric.git
cd capability-fabric
./install/install.ps1If PowerShell blocks the script, run it without changing your policy:
pwsh -ExecutionPolicy Bypass -File ./install/install.ps1macOS and Linux:
git clone https://github.com/AdrianParedez/capability-fabric.git
cd capability-fabric
./install/install.shRemove the installation:
./install/install.ps1 -Uninstall # Windows
./install/install.sh --uninstall # macOS / LinuxInstaller options:
| Flag | Effect |
|---|---|
-All / --all |
Target every known agent location, detected or not. |
-Link / --link |
Symlink instead of copy. |
-NoDirective / --no-directive |
Install skills only. |
-Uninstall / --uninstall |
Remove installed skills and the routing block. |
Claude Code plugin install:
/plugin marketplace add AdrianParedez/capability-fabric
/plugin install capability-fabric
The Claude Code plugin metadata is experimental. The skills and installers are the supported distribution surface for the public release.
The public documentation is organized under docs/. Start with
docs/README.md.
- Research report
- Competitive analysis
- Capability matrix
- Gap analysis
- Skill taxonomy
- Dependency graph
- Activation strategy
- Context optimization
- Sources
Install the library, open an agent, and state the work in ordinary language:
research the current Agent Skills standard and write a report
plan this repository migration and keep a durable ledger
review this feature branch for behavioural regressions
create a new portable skill for release-note drafting
The routing block points the agent at skills/REGISTRY.md. The
registry defines the available capabilities and default composition paths.
Manual installation:
skills/<name>/ -> ~/.agents/skills/<name>/
skills/<name>/ -> ~/.codex/skills/<name>/
skills/<name>/ -> ~/.claude/skills/<name>/
See docs/release/ROADMAP.md for the current lane, planned work, and non-goals.
| Skill | Behaviour |
|---|---|
context-budgeting |
Controls context cost, offload points, and compaction pressure. |
progressive-research |
Finds external facts with bounded search, triage, and corroboration. |
decomposing-tasks |
Converts ambiguous work into a verifiable, risk-ordered plan. |
sustained-execution |
Keeps long runs resumable through an external task ledger. |
verifying-reasoning |
Seeks disconfirmation before conclusions are trusted. |
operating-autonomously |
Runs bounded unattended work with progress and stuck detection. |
guarding-code-quality |
Fits changes to the codebase and verifies behaviour before reporting done. |
composing-skills |
Selects the smallest matching skill set and sequences it. |
authoring-skills |
Generates spec-compliant portable skills. |
auditing-skills |
Scores and improves skills against a quality rubric. |
Skill directory shape:
skills/<name>/
SKILL.md
references.md
examples.md
templates/
checklists/
benchmarks/
Default routes:
external information progressive-research, verifying-reasoning
multi-step work decomposing-tasks, sustained-execution, guarding-code-quality
unattended execution operating-autonomously
skill creation authoring-skills
skill improvement auditing-skills
ambient control context-budgeting
Routing constraints:
| Rule | Effect |
|---|---|
| Smallest match | Load only the skills needed for the task. |
| Tiered composition | Higher-tier skills may compose lower-tier skills. |
| Acyclic graph | Skills never depend upward. |
| Portable wording | Required steps name capabilities, not runtime tools. |
| Marker-delimited install | The routing block updates in place and is never duplicated. |
capability-fabric/
.github/
ISSUE_TEMPLATE/
workflows/
CODE_OF_CONDUCT.md
CONTRIBUTING.md
pull_request_template.md
SECURITY.md
AGENTS.md
README.md
install/
install.ps1
install.sh
routing-directive.md
docs/
README.md
design/
release/
research/
skills/
REGISTRY.md
<skill>/
.claude-plugin/
GitHub Actions runs CI on push and pull request. Link Check and Release Check are
manual workflows for public documentation and release-readiness checks.
Validate changed skills where the tool is available:
skills-ref validate ./skills/<name>Installer syntax checks:
$tokens = $null
$errors = $null
$null = [System.Management.Automation.Language.Parser]::ParseFile(
(Resolve-Path 'install/install.ps1'),
[ref]$tokens,
[ref]$errors
)
$errorsbash -n install/install.sh- Experimental library.
- Runtime support depends on Agent Skills compatibility.
- The installer writes to global agent configuration files.
- Claude Code plugin metadata remains experimental.
- Benchmarks describe methods and targets, not published execution results.
- Skills improve agent process, not domain expertise by themselves.
Licensed under the Apache License, Version 2.0. See LICENSE.