README

Personal Agent Skills

Portable personal agent skills for coding agents
Evidence-based review, planning, documentation, and research workflows

Portable agent skills maintained by Terabytesoftw. Each skill is self-contained under skills/ and follows the Agent Skills directory format.

Getting started

Composer

Publish terabytesoftw/skill on Packagist before using the normal installation below. Create skills.json first when the consumer needs a reproducible target and an exact Composer trust list:

{
  "target": ".agents/skills",
  "aliases": [".claude/skills"],
  "auto-sync": true,
  "vendor-sources": false,
  "dependencies": {
    "composer": {
      "trusted-replace": true,
      "trusted": ["terabytesoftw/skill"]
    }
  }
}

composer config allow-plugins.llm/skills true
composer require --dev llm/skills:1.15.0 terabytesoftw/skill:^0.1@dev
composer skills:show

^0.1@dev is a development-compatible constraint, not a stable release or an exact pin. The package maps dev-main to 0.1.x-dev; the consuming project's lock file records the selected commit. Review changes before updating.

The package publishes extra.skills.source: skills for the llm/skills Composer plugin. composer skills:show verifies the donor and synchronization state; it does not prove that a coding client discovered or loaded a skill. Confirm the desired skill in the client from the consuming project or package root.

For local package development, use a Composer path repository from a disposable consumer and require terabytesoftw/skill:dev-main. Keep that path override out of shared project metadata unless the team intentionally uses the same checkout layout.

Skills CLI

As an optional client-independent route, select and install skills from the repository:

npx skills@latest add terabytesoftw/skill

This route requires the external skills CLI and its supported Node.js runtime. Composer remains the primary distribution metadata for this repository.

Use a skill

Ask the client to use the desired skill and provide its required inputs. For an adversarial review, specify:

provider and model for each reviewer;
effort for each reviewer, preserving an explicit value or using default when none is specified;
maximum review rounds;
a cost ceiling when the active host can enforce or report one; and
manual, sequential, or stop as the fallback.

The skill verifies active client capabilities before dispatch. Model selection, cost limits, external calls, and source-code transmission are never inferred from installation.

Skills

The catalog contains seven skills:

Skill	Purpose
adversarial-review	Critical review with independent contexts or disclosed fallback
analyze-first	Inspect a non-trivial problem before implementation
assumption-tracking	Expose and validate material assumptions
conventional-commits	Prepare commit text and one changelog entry per PR
phpdoc	Preserve useful PHPDoc and static-analysis types
plan-then-execute	Plan multi-step or high-risk work before implementation
research-protocol	Verify version-specific behavior from primary sources

CATALOG.md records the inventory, inclusion decisions, exclusions, and provenance boundaries.

Usage examples

Copyable prompts and illustrative outputs provide two practical scenarios for every skill. Each example states its minimum context, expected artifact, verification criteria, and limits. The separate cross-client evaluation records actual Claude Code, Codex, and GitHub Copilot runs without turning observed output into the expected answer.

Testing skills

This catalog intentionally does not duplicate skills for creating, migrating, configuring, or running PHP tests. Testo maintains those workflows in its official skill catalog, publishes the directory through extra.skills.source, and documents automatic project installation through the suggested llm/skills Composer plugin. The local phpdoc skill only preserves useful documentation and type information when test code is already in scope.

Maintenance

Composer contains distribution metadata only. CI invokes Composer validation, repository linters, workflow security checks, and evaluation patch validation directly.

Behavioral cases live under evals/adversarial-review/. They are maintainer evaluations, not product testing skills or unit tests, and never contact external providers automatically.

Built with AI

License

Original repository work is available under the BSD 3-Clause License. The adversarial-review skill retains the notices and license texts required by the review material identified in its provenance reference.

terabytesoftw / skill

Package info

Security

Statistics