Portable AI-agent workflows for Pi, Claude Code, and Codex. This repository keeps reusable skills, Claude-style sub-agents, and shared prompt templates in one source so workflows can be authored once and used across clients.
These companion repositories are installed and versioned independently; make apply does not
install them:
- AI Tools — launchers, statusline, and cross-client runtime glue for Pi, Claude Code, and Codex.
- Pi Skill Model Router — maps this repository's portable routing metadata to exact models while keeping model, fallback, and spend policy local.
- Pi Watch Loop — owns generic bounded scheduling and protocol mechanics; watcher cadence and workflow policy remain in this repository's skills.
- Pi Session Mode — provides guarded plan/implement modes and scoped writer leases as an accidental-change guard, not a sandbox.
Optional upstream runtime:
- pi-subagents — an independently maintained Pi
delegation runtime used by
delegate-workfor lifecycle and mechanics when installed; it is not bundled with or owned by this project.
- Skills — focused workflows stored as
skills/<name>/SKILL.mdand installed into the portable~/.agents/skills/root. - Agents — Claude-style sub-agent definitions stored in
agents/and installed into~/.claude/agents/. - Prompt templates — shared Pi and Claude Code slash commands stored in
prompts/, installed into~/.agents/prompts/with Pi and Claude Code aliases.
Browse the skills catalog, grouped by topic, for the complete list.
For occasional collection maintenance, explicitly invoke /skill:skills-review in Pi
or /skills-review in Claude Code (Codex: request the skill by name). It screens a
declared source catalog and reviews a bounded content sample; it does not rewrite,
retire or install skills. See the review contract.
git clone https://github.com/flurdy/agent-skills.git
cd agent-skills
make dry-run # preview managed symlink changes
make apply # install skills, agents, prompts, and client aliases
make doctor # verify the installationThe installer manages individual symlinks rather than replacing destination roots, so unrelated user-owned skills, agents, and prompts remain untouched.
| Client | Skills | Agents | Prompt templates |
|---|---|---|---|
| Pi | Discovers ~/.agents/skills/ |
Not managed by this repository | Managed aliases in ~/.pi/agent/prompts/ |
| Codex | Discovers ~/.agents/skills/ |
Uses Codex-native agent support | Not installed; custom prompts are deprecated |
| Claude Code | Managed aliases in ~/.claude/skills/ |
Managed links in ~/.claude/agents/ |
Managed aliases in ~/.claude/commands/ |
make apply-codex remains as a compatibility alias for applying the shared skill
root and Claude compatibility aliases, without Claude-style agents or prompts.
The assembler resolves shared and optional private layers, preflights collisions, and creates managed symlinks in each client location. Later layers can override units with the same name without copying shared content.
agent-skills/
skills/<name>/SKILL.md
agents/<name>.md
prompts/<name>.md
docs/
assemble.sh
Makefile
Detailed destination variables, private layers, migration behavior, and cleanup rules live in the installation and configuration guide.
Shared skills declare portable capability and reasoning requirements in frontmatter:
model-tier: standard
effort: mediumThe supported tiers are economy, standard, and premium; reasoning effort is
low, medium, high, or xhigh. These values describe what a workflow needs,
not a fixed provider or model. Exact model IDs, fallback order, billing
classification, and spend controls remain runtime-local.
See MODEL_ROUTING.md for the policy. Pi can optionally enforce this metadata with the router described under Related repositories. Claude Code and Codex use their own runtime configuration and capabilities.
Browse the prompt catalog for the available slash commands.
Templates use $ARGUMENTS, which Pi and Claude Code both expand. make apply
installs every top-level template for both clients; run /reload in Pi
afterward. See the prompt-template setup
for destinations and migration notes.
The commands provide instructions only; none performs Git operations itself.
A sibling agent-skills-private/ repository can add private shared units plus
machine-, client-, and profile-specific overrides. The resolved order is shared →
private → machine → clients. The shared repository remains usable without any
private repository.
See Private overlays for the directory layout and commands.
- Skills catalog — available workflows and descriptions
- Installation and configuration — destinations, prompts, private layers, migrations, and cleanup
- Contributing — authoring skills and agents, metadata, permissions, validation, and tests
- Model routing — portable tier and effort policy
- Documentation index — public guides and component references
- Report issues at github.com/flurdy/agent-skills/issues.
- Pull requests are welcome at github.com/flurdy/agent-skills/pulls.
Created by flurdy.
MIT License. See LICENSE.