agentic-skills

Repository-agnostic engineering and creative skills and delegation roles for Claude Code, Codex, omp, and Pi. Skills use the same Markdown in every harness. Codex roles are distributed as committed TOML files; omp uses its built-in task tool with the committed Markdown roles; Pi role execution is owned by pi-subagents.

Skills

Each skill is a generic method. Read its canonical file for routing and procedure.

Brainstorming, code design, planning, implementation, artifact design, visual design, typography, and animation own their applicable review checkpoints. Implementation includes result and task-level historical review. The reviewing skill supplies the review method, reuse rules, and specialist selection. Brainstorming calls the premise-checker directly; earlier exploration does not replace that second opinion.

Artifact design supplies the common purpose, consumer, context, and ownership questions for documents and assets; its prose guidance applies to prose. The producing task skill owns specialized method, evidence, and completion, with shared artifact questions included in its applicable review. Visual design spans illustration, graphics, interfaces, and 3D work; typography owns text treatment and placement, and animation owns movement over time. Select specialists for the distinct judgments the task needs. Appearance and readability require actual presentation evidence; motion quality requires continuous playback at the intended speed and context. Report qualities left unassessed when that evidence is unavailable.

Applicable workflow instructions authorize the coordinating agent to delegate the reviews they require and the exploration or implementation they permit, without a separate operator request. Required independent review uses a suitable available reviewer unless directly applicable review can be reused; exploration and implementation delegation remain optional. Task scope, write boundaries, explicit prohibitions, harness permissions, and child-role limits still apply. agentic-subagents configures delegation after the calling workflow supplies that authority.

Delegated roles

Start each child with a fresh conversation, without forking or inheriting the parent transcript. Provide a self-contained assignment with the outcome, task boundary, and relevant settled decisions. Applicable global and repository instructions should remain available; the brief need not duplicate them. Safety, permissions, and harness constraints remain authoritative. The role and brief set the boundary, while loaded skills supply method within it.

The shared reviewing skill guides direct review, reviewer selection, and delegation briefs. Each reviewer role carries its own report-only review contract. There is no generic reviewer role. Write none explicitly when no task-specific constraints apply; this does not discard applicable instructions. Cite the repository path for a load-bearing constraint when one exists. Source restrictions, desired detail, and existing verification evidence are optional.

Children may ask the parent for material missing context through an available communication channel; otherwise they report the limitation or blocker. Clarification does not expand role authority. Explorer, premise-checker, and all reviewers are report-only roles. Their mutation limits are behavioral prompt constraints, not security boundaries; harness-provided safety and permissions still apply.

Claude Code

claude plugin marketplace add hypnotox/agentic-skills
claude plugin install agentic-skills@agentic-skills

# Local checkout
claude --plugin-dir /absolute/path/to/agentic-skills

The marketplace plugin deliberately has no manifest version, so Claude identifies Git-hosted updates by the source commit. Run claude plugin update agentic-skills@agentic-skills to check for and install a newer revision; commit identity does not schedule updates or add automatic watching.

Claude Code prefixes each role name with agentic-skills:, for example agentic-skills:agentic-implementation-reviewer.

Codex

Use a current Codex release with custom TOML agents. The full installation below comes directly from this GitHub repository and needs only Git and a shell—no npm install, generation, or build step.

Install skills and agents from the repository

Clone the whole repository into Codex's user skills directory, then link its generated agents into Codex's user agent directory. Codex discovers the nested skill folders and TOML files. In Bash on Linux or macOS:

(
  set -e
  repo="$HOME/.agents/skills/agentic-skills"
  agents="${CODEX_HOME:-$HOME/.codex}/agents/agentic-skills"
  test ! -e "$agents"
  test ! -L "$agents"
  mkdir -p "$(dirname "$repo")" "$(dirname "$agents")"
  git clone https://github.com/hypnotox/agentic-skills.git "$repo"
  ln -s "$repo/codex/agents" "$agents"
)

The destination names must be unused; the commands do not replace an existing agent directory or symlink. Keep the full checkout: some skills link to ../../agents/*.md, so copying only individual skill directories would lose their role references. For an existing checkout, link its root into ~/.agents/skills/agentic-skills and its codex/agents directory into ${CODEX_HOME:-$HOME/.codex}/agents/agentic-skills, using unused destination names. For project-only discovery, use the project's .agents/skills/agentic-skills and .codex/agents/agentic-skills instead.

Restart Codex after installation. Check /skills for the skills listed above (Codex may prefix them with agentic-skills:), and refer to agents by their declared names, such as agentic-explorer or agentic-implementation-reviewer. Give each the required brief and explicitly choose fresh/no-history context using the installed Codex delegation tool's contract; do not assume its default is fresh. The TOML files set only name, description, and developer_instructions, leaving model, reasoning, tools, and permission configuration to Codex and the user.

Update the checkout and restart Codex:

git -C "$HOME/.agents/skills/agentic-skills" pull --ff-only

The directory link follows added, changed, and removed agent files without regeneration or relinking. To uninstall, remove only the agent-directory symlink and move the checkout out of the skills directory (or remove it if no longer needed).

Native repository plugin: skills only

Alternatively, install the skills through Codex's plugin marketplace. This repository includes a root plugin.json and marketplace catalog:

codex plugin marketplace add hypnotox/agentic-skills
codex plugin add agentic-skills@agentic-skills

# Local checkout instead of the GitHub marketplace source
codex plugin marketplace add /absolute/path/to/agentic-skills
codex plugin add agentic-skills@agentic-skills

Choose one skills installation route to avoid duplicate discovery. The plugin carries the role files but does not register custom Codex agents. If using the plugin and wanting named agents too, keep an ordinary checkout outside the skills directory and link only its codex/agents directory into Codex's agent directory as above. Refresh the marketplace with codex plugin marketplace upgrade agentic-skills; manage the installed plugin with Codex's plugin commands. This repository is not a listing in OpenAI's public plugin directory.

omp (Oh My Pi)

Use a current omp release with plugin skill and task-agent discovery. Install through its native marketplace commands; omp reuses the repository's Claude-compatible catalog:

omp plugin marketplace add hypnotox/agentic-skills
omp plugin install agentic-skills@agentic-skills

# Local checkout, without installation
omp --plugin-dir /absolute/path/to/agentic-skills

# Persistent local checkout link instead (Bun required for uninstall)
omp plugin link /absolute/path/to/agentic-skills

Choose one route. The marketplace installation and local link both expose the committed skills and all twelve Markdown roles in agents, including all nine reviewers. The explicit package.json omp manifest is empty because omp discovers these conventional directories without extension modules or path mappings. No pi-subagents, Pi-specific overrides, generation, or build step is needed.

The marketplace and --plugin-dir routes do not require Bun for this instruction-only package. Local-link removal goes through omp's npm/link plugin manager and requires bun on PATH. To persistently install a local checkout without that prerequisite, use omp plugin marketplace add /absolute/path/to/agentic-skills followed by the same marketplace install command.

Restart omp after CLI installation or linking. After an in-session /marketplace install, use /reload-plugins to refresh skills and task agents. Skill commands use /skill:agentic-reviewing (and the other skill names above); roles use their unprefixed frontmatter names, such as agentic-implementation-reviewer, not Claude Code's agentic-skills: prefix.

If omp also discovers a Claude installation of this package, it deduplicates identical skills instead of importing them twice. Differing revisions remain available under namespaced skill names; keep both installations aligned or select a single discovery source if you want only one revision exposed.

Launch reviewers with omp's native task tool and the required review brief. With its default batch schema:

{
  "context": "Outcome: the change must preserve the public API. Review surface: src/ and the affected tests. Settled decisions: no API changes. Applicable task-specific constraints: none.",
  "tasks": [
    {
      "agent": "agentic-implementation-reviewer",
      "task": "Review the change for correctness and contract consistency. Report findings with precise evidence; do not edit or delegate.",
      "solutionSpace": "Independent correctness review; no known defect or prescribed fix."
    }
  ]
}

The calling workflow must authorize delegation. omp starts task children without the parent transcript and carries discovered skills and applicable context files into the child; do not pass Pi's context: "fresh" field. Here context is shared assignment text, not a history mode. Supply each child a complete brief. Role and brief boundaries remain authoritative; report-only roles' mutation and no-further-delegation limits remain prompt constraints, not a sandbox.

For other specialist reviews, replace agent with the appropriate name from Delegated roles. Use the active tool schema if task.batch is disabled. The shared roles do not pin models, thinking, or tools; omp's configuration and supported per-item launch overrides remain authoritative. Follow agentic-subagents before choosing overrides. Background results are delivered automatically; use wait only when blocked and agent://<id> for full reports. Use write to agent://<id> for clarification when omp exposes peer messaging.

For marketplace updates:

omp plugin marketplace update agentic-skills
omp plugin upgrade agentic-skills@agentic-skills

Upgrade this plugin explicitly: omp's upgrade-all version comparison does not detect new revisions of this versionless catalog entry. For a local link, update the checkout instead. Restart or reload after updates. Uninstall with omp plugin uninstall agentic-skills@agentic-skills for a marketplace install, or omp plugin uninstall agentic-skills for a local link; the latter removes the link, not the source checkout. Use --scope project on marketplace installation, upgrade, and removal for project-only use; plugin link is user-scoped.

Pi

Install pi-subagents as the sole subagent provider, then install this package:

pi install npm:pi-subagents
pi install git:github.com/hypnotox/agentic-skills

# Local checkout
pi install /absolute/path/to/agentic-skills

This package declares its committed agents directory for native discovery. Use the role names with the installed subagent API. Pass context: "fresh" explicitly when launching these roles so a global context default cannot fork the parent conversation:

{
  agent: "agentic-explorer",
  context: "fresh",
  task: "Question: ...\nEvidence boundary: ...\nApplicable task-specific constraints: none"
}

Use /subagents-guide tool-reference for the installed version's complete invocation contract.

Merge these Pi-specific settings into your existing user or project settings while preserving unrelated values:

{
  "subagents": {
    "agentOverrides": {
      "agentic-premise-checker": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-explorer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-implementer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-implementation-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-code-design-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-plan-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-instruction-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-artifact-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-visual-design-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-typography-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-animation-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]},
      "agentic-retrospective-reviewer": {"systemPromptMode": "append", "inheritProjectContext": true, "inheritGlobalContext": true, "inheritSkills": true, "defaultContext": "fresh", "excludeTools": ["handoff_session"]}
    }
  }
}

These overrides append the specialist prompt to Pi's base prompt, load project and global instruction files, and expose discovered skills without inheriting the parent conversation. They keep Pi-specific runtime fields out of the cross-harness role prompts. Remove the obsolete agentic-reviewer override when updating an existing installation.

Use Pi's native contact_supervisor and parent reply mechanism for material clarification when available. Roles do not delegate further, and handoff_session remains excluded. Supported per-launch model and thinking choices may follow agentic-subagents; effective Pi and pi-subagents configuration remains authoritative. This package adds no runtime routing.

For delegation authorized by an applicable workflow, if subagent is inactive and subagents_enable is exposed, call subagents_enable({}) first. It activates tools without launching work; subagent becomes available on the next model request. Use this supported activation before treating delegation as unavailable. If subagent is already exposed, proceed directly.

Before execution, call subagent({ action: "list", capabilities: true }) and select only an executable agent. Before passing a model override, call subagent({ action: "models" }) and use an exact provider/id from its available-model output. Use native background execution when a role needs ordinary installed extensions; foreground SDK children do not automatically load ambient extensions. Restart Pi after changing installed packages or these settings, and use /subagents-guide agents for further inspection.

Development

omp, Pi, Claude Code, and Codex use the committed skills and self-contained agent files directly. Installation and execution require no generation or build step; there are no install or packaging hooks that generate instructions.

Agent sources and rendering

The reviewer Markdown files in agents and every TOML file in codex/agents are generated. Maintain their sources instead:

• templates/reviewers/template.md owns the common role boundary and composition.
• templates/reviewers/roles owns each reviewer's frontmatter, introduction, and ## Focus section. Each source filename determines its output filename in agents/.
• skills/agentic-reviewing/SKILL.md owns the generic method. Its complete ## Review and ## Report sections are included in every reviewer; the selection and delegation guidance stays in the skill.

scripts/generate-agents.ts fills the template's four slots: introduction, review, focus, and report. Each slot must occur exactly once. No template processing happens in any harness. Explorer, premise-checker, and implementer remain handwritten in agents/.

The same generator renders every role's Markdown metadata and body into a self-contained Codex TOML file. It composes reviewers from their owning sources in memory, not from potentially stale generated Markdown. Codex filenames follow the frontmatter name; duplicate or invalid names fail before outputs are changed. Build notices are excluded from developer_instructions, and no harness settings are added to the shared role prompts. omp discovers the shared Markdown agents directly; it has no separate generated format.

After changing reviewer sources or handwritten roles, run npm run generate and commit the source and generated changes together. Generation also removes marked generated agents whose source was removed, leaving unowned files alone. The generator and templates are development tooling, excluded from the npm package.

Checks

npm run check requires the current Node release, npm, and Claude Code. It checks generated files without rewriting them, then runs typechecking, tests, and harness validation. npm run generate:check runs only the non-writing drift check and fails for missing, changed, or obsolete generated agents in either format. Tests parse the Codex TOML, compare it with the canonical Markdown, and inspect npm's actual package file list, including the hidden marketplace. Codex itself is not required for these development checks.

npm install
npm run generate # After changing reviewer sources
npm run check
npm pack --dry-run

License and provenance

AGPL-3.0-only. Adapted from hypnotox/agentic-workflows; see NOTICE.

The methods expand principles from agentic-doctrine into self-contained skills and focused reviewer criteria. This README owns the provenance reference; installed skills and roles carry the guidance they need.