GAIK agent skills

All GAIK agent skills live here, packaged as one portable Agent Plugins v1 plugin, gaik-toolkit. Add the repository as a plugin marketplace and every skill below is available in Claude, Claude Code, Codex, Copilot and VS Code. Each skill is also an ordinary Agent Skill that can be copied or uploaded on its own.

Skills

Every skill is a directory under skills/ with a SKILL.md (the workflow the agent follows) and a README.md (what the skill does, for people).

For developers building with the gaik package. These carry decisions and failure modes that are expensive to rediscover — which parser preserves a table's structure, why an extraction request is refused, why a hybrid search quietly became vector-only — rather than restating the API reference.

Ready-to-run use cases. No code needed. brief-to-slides runs in any agent that can run Python and bash; the other four are written for Claude Desktop and need a local MCP server, which each skill's README sets up.

report-writing has Claude write the report itself in Claude Desktop. generate-structured-reports runs gaik's normalize, curate and synthesize pipeline instead, with quotes checked against the sources, and needs an agent that can run Python.

Long detail sits in references/ (or reference/) and is loaded only when the task needs it. The Claude Desktop skills also carry their setup files beside SKILL.md: setup.bat and transcription-MCP/. Their sample recordings, purchase orders, price list and example output are not part of the plugin; they are in the repository under implementation_layer/no-code-assets/samples/, and each skill's README links its own.

Data, services and credentials

The plugin contains instructions, reference text, scripts and two PowerPoint templates. It declares no MCP server, hook or command, and it sends nothing anywhere by itself. A skill reaches an outside service only when the agent follows it, and only the services below. Each one receives the user's own content, under that provider's terms.

Installs. Skills that run code install packages from public registries: gaik and its dependencies from PyPI (gaik==0.8.5 with the extra the task needs), mcp, python-dotenv, python-pptx, pillow, cairosvg, lxml and playwright from PyPI, Playwright's Chromium build, the font packages and lucide-static from npm for brief-to-slides, and @modelcontextprotocol/server-filesystem from npm for the Claude Desktop MCP setup. Those commands name exact versions, and the packages' own dependencies resolve when the user installs them. generate-structured-reports installs gaik[report-writer], and a provider's gaik extra when it is missing, without a version pin. After the user agrees, it also downloads Pandoc from its official release through pypandoc. Its SKILL.md describes both.

Credentials. No credential is stored in this plugin. The skills use whatever API key the user has set for gaik in their environment or in a local .env file, and they send it only to the provider it belongs to. The setup.bat scripts ask the user for the key and write it to a .env file on the user's machine. That file stays there and is never part of the plugin or the repository.

Directory review notes

• skills/brief-to-slides/assets/template.pptx (391 KiB) and blank-template.pptx are the decks the skill builds on, so they stay in the plugin and their review hold is accepted. Without them the skill cannot keep a template's layouts, fonts and theme.
• The skills ask for no userConfig value. The keys are gaik's ordinary environment variables, which gaik also reads outside the agent, and Codex has no userConfig.

Installing

One package serves every client. The repository root carries a marketplace file, .claude-plugin/marketplace.json, that Claude, Claude Code, Codex, Copilot CLI and VS Code all read.

Claude Code 2.1.278 and Codex 0.149.1 were tested end to end (September 2026), and the Claude Desktop route on claude.ai in October 2026. The other rows follow those clients' documentation. Claude's Sync automatically option, on by default, keeps the plugin up to date as this repository changes.

Layout

The root also holds the plugin icon that the Claude and OpenAI directory listings show.

agent-skills/   (implementation_layer/no-code-assets/)
├── plugin.json          # Agent Plugins v1 manifest
├── .claude-plugin/
│   └── plugin.json      # Claude Code's own manifest; also the directory listing fields
├── LICENSE
└── skills/
    ├── parsing-documents/
    ├── extracting-structured-data/
    ├── searching-documents/
    │   ├── SKILL.md
    │   ├── README.md
    │   └── references/
    ├── generate-structured-reports/   # also scripts/: the stage helper
    ├── brief-to-slides/        # also assets/ and scripts/
    ├── incident-report-writing/
    ├── construction-diary-creation/
    ├── purchase-order-processing/
    └── report-writing/
        ├── SKILL.md
        ├── README.md           # setup guide
        ├── reference/
        ├── setup.bat           # installs the MCP servers for Claude Desktop
        └── transcription-MCP/

How the clients read it

• Agent Plugins v1 clients (Codex, Copilot, VS Code, Cursor, Kiro) read plugin.json at this directory's root and discover skills in skills/.
• Claude Code does not implement Agent Plugins v1. It reads its own manifest, .claude-plugin/plugin.json, and discovers the same skills/. The Agent Plugins migration guide keeps a client's working files beside the portable manifest and removes them only once that client is tested without them — Claude Code still needs this one, and it is what names the skills gaik-toolkit:… even when the directory is loaded directly.
• Codex finds the marketplace through .claude-plugin/marketplace.json, which it accepts as a legacy-compatible location; its native one is .agents/plugins/marketplace.json. One file serves everything today — add the native one only if Codex stops reading the legacy path.

The format is deliberately small, and two of its rules shape any extension:

Component locations are fixed and cannot be redirected. plugin.json names and describes the plugin; it cannot relocate skills/, and it cannot declare a component inline. A client finds skills by looking in skills/ for immediate child directories containing a SKILL.md — it does not search deeper, so a skill nested two levels down is invisible.

The manifest schema is closed. Allowed fields are $schema, name, version, description, author, homepage, repository, license, keywords and extensions. Client-specific data goes under extensions, keyed by a reverse-domain namespace such as com.openai, or into a top-level directory named exactly for that namespace. A client ignores namespaces it does not implement, so the package stays portable.

Releasing a change

Bump version in both manifests — plugin.json and .claude-plugin/plugin.json — with every change under this directory. Each client caches an installed plugin under the version it read, so an edit that keeps the old version never reaches anyone who already installed it. The test below fails if the two versions disagree. Then run uv run python implementation_layer/toolkit_demo_app/scripts/build_no_code_catalog.py: the demo app's skills page serves a generated copy of these skills and shows the plugin version.

The skills quote gaik's API. implementation_layer/unit_tests/test_agent_plugin.py checks the manifest, the marketplace entry, every skill's frontmatter and links, and that the gaik names the skills rely on still exist — so a rename in gaik fails CI until the skill is updated too.

Extending

A skill. Create skills/<skill-name>/SKILL.md with name (matching the directory) and a third-person description that says when to use the skill, not only what it does — that text is the whole basis on which an agent decides to open it. Put anything long in references/ and say in SKILL.md when to read it. No registration step.

An MCP server. Declare it twice, because the two families read different files — tested: Claude Code 2.1.278 loads only .mcp.json, Codex 0.149.1 only mcp.json, so the same server in both files does not collide:

An invalid mcp.json disables only the MCP servers; the skills still load.

Client-only features go where that client looks, never into the manifest's top level. Agent Plugins clients use their extension namespace — extensions.com.openai for Codex apps and hooks, a com.github.copilot/ directory for Copilot agents and commands. Claude Code has no namespace; it reads agents/, commands/ and hooks/hooks.json at the plugin root.

Another plugin. Put it in its own directory beside this one, with both manifests, and add an entry to .claude-plugin/marketplace.json whose source is its relative path.

Distribution

.claude-plugin/marketplace.json makes this repository a self-hosted marketplace: anyone can add it by name, as in the table above. It does not list the plugin in any client's curated directory. Those take separate submissions — Anthropic's claude-plugins-official, which Claude Code's Discover view browses, accepts third-party plugins through its plugin directory submission form.

Verifying

claude plugin validate .
claude plugin validate ./implementation_layer/no-code-assets/agent-skills
uvx --from skills-ref agentskills validate implementation_layer/no-code-assets/agent-skills/skills/searching-documents
uv run pytest implementation_layer/unit_tests/test_agent_plugin.py

To try a change in Claude Code without installing it, run claude --plugin-dir ./implementation_layer/no-code-assets/agent-skills; it loads the plugin for that session only.

License

MIT, same as the toolkit.