agentlik

agentlik

Claude Code skills, grouped into seven packs so each can be enabled independently. 27 skills covering reasoning, foresight, learning, writing, development workflow, local-model delegation, and personal coaching.

Each pack is a Claude Code plugin. You install the packs you want and leave the rest off, so you only pay context for what you use.

Packs

Auto-triggered skills put their description in Claude's context every session, and Claude can invoke them on its own. Manual only skills cost nothing until you type /name — Claude cannot reach them. See Skill visibility below.

Two packs have external requirements: local-llm needs a llama.cpp server running, and dev's docket skill needs the docket CLI installed.

Install

Requires Claude Code. Both methods below install at user scope, so the packs are available in every project.

As a marketplace (recommended)

claude plugin marketplace add swayson/agentlik
claude plugin install reasoning@agentlik
claude plugin install writing@agentlik

Install only the packs you want. To update later:

claude plugin marketplace update agentlik
claude plugin update reasoning@agentlik

Note that claude plugin install copies the pack into ~/.claude/plugins/cache/. Edits you make to a clone will not take effect until you bump the pack's version in .claude-plugin/plugin.json and re-run update. If you intend to edit the skills, use the method below instead.

As live-editable clones

This makes each pack load directly from your clone, so edits apply in the next session with no install or update step.

git clone https://github.com/swayson/agentlik.git

Then link each pack you want into ~/.claude/skills/. Anything placed there loads automatically as <name>@skills-dir.

macOS / Linux:

ln -s /path/to/agentlik/packs/reasoning ~/.claude/skills/reasoning

Windows (PowerShell — mklink through Git Bash mangles the paths):

New-Item -ItemType Junction -Path "$HOME/.claude/skills/reasoning" -Target "C:/path/to/agentlik/packs/reasoning"

Verify with claude plugin list; each linked pack should report loaded.

To remove a link on Windows, use [System.IO.Directory]::Delete("$HOME/.claude/skills/reasoning"). Do not use rm -rf on a junction — depending on the tool it may follow the link and delete your clone.

Enabling packs per project

Keep the packs you rarely need switched off globally and turn them on in the repos that want them:

claude plugin disable futures@skills-dir --scope user
cd ~/some-research-project
claude plugin enable futures@skills-dir --scope project

The second command writes .claude/settings.json in that project. Settings precedence is user < project < local < flag < policy, so a project-level true overrides a user-level false. Use --scope local for machine-specific choices you do not want to commit.

Check what a pack costs before enabling it:

claude plugin details reasoning@agentlik

This reports always-on tokens (paid every session) and on-invoke tokens (paid each time a skill fires), broken down per skill.

Skill visibility

A skill with disable-model-invocation: true in its frontmatter is hidden from the model: its description is never loaded, and Claude cannot invoke it. You reach it by typing /name. Skills that benefit from firing automatically — simplify after writing code, causal-reasoning-engine on a "why did this happen" question — leave the field off and pay the always-on cost.

Set visibility in the SKILL.md itself, not in the skillOverrides setting. The frontmatter field travels with the skill and is version-controlled.

Writing and editing skills

writing/skills/write-great-skills documents the conventions these packs follow — invocation choice, the information hierarchy, progressive disclosure, and the failure modes to prune for. Read it before adding a skill. packs/reasoning/skills/luck is a worked example of progressive disclosure: a compact SKILL.md with four reference files behind context pointers.

To add a pack, create packs/<name>/.claude-plugin/plugin.json and packs/<name>/skills/<skill>/SKILL.md, then add the pack to plugins in .claude-plugin/marketplace.json.

Other contents

prompts/ holds standalone prompt files, and claude-settings/ a reference settings.json. Neither is loaded by the plugin system.

Licence

Apache 2.0. See LICENSE.