How the Skill Installation Script Operates and Which Hosts It Configures

The skill installation script (bin/install.sh) validates the Agent Workbench pack, copies skill artifacts including AGENTS.md, and configures repositories to support Claude Code, Codex, and other AI hosts through an idempotent three-phase process.

The rohitg00/ai-engineering-from-scratch repository ships with a portable skill installation script designed to drop the Agent Workbench curriculum into any target repository. This bash utility automates the deployment of curriculum packs while establishing host-specific invocation contracts for multiple AI coding environments.

How the Skill Installation Script Operates

Located at bin/install.sh in the Agent Workbench pack, the script executes three distinct phases to safely deploy skill assets.

Phase 1: Validation

The script first declares a list of required pack items: AGENTS.md, VERSION, docs/, schemas/, and scripts/. It verifies each item exists under the pack root ($PACK_ROOT) before proceeding.

Critically, the script checks whether AGENTS.md already exists in the target repository. If detected, the installation aborts unless the --force flag is supplied. This protects existing host contracts from accidental overwrites while guaranteeing a complete pack structure.

Phase 2: Installation

During the core installation phase, the script performs the following atomic operations:

• Copies the canonical AGENTS.md into the target repository: cp "$PACK_ROOT/AGENTS.md" "$TARGET/AGENTS.md"
• Recursively copies the docs/, schemas/, and scripts/ directories using mkdir -p and cp -r
• Writes the pack version into a hidden marker file (.workbench-version)

This phase installs the full suite of skill assets—documentation, schema definitions, and helper scripts—in a single, reproducible step.

Phase 3: Feedback

Upon completion, the script emits a concise success message displaying the installed version. It suggests next manual steps including configuring task_board.json, running acceptance commands, and executing scripts/init_agent.py to bootstrap the agent environment.

Hosts Configured by the Script

The critical file installed by the script is AGENTS.md, which defines the host-invocation contract for all portable skills. This markdown file specifies how each skill can be invoked across different execution hosts.

Claude Code

For Claude Code environments, the script configures slash-command style invocations. The host contract enables syntax like:

Use /learn-agent-skills to start the Agent Skills path.

This routing allows Claude Code to recognize and execute skills using native slash command patterns.

Codex

For Codex hosts, the installation configures natural-language style invocations:

Use learn-agent-skills to start or resume the Agent Skills path.

The script ensures the target repository recognizes this invocation pattern without requiring slash prefixes.

Other Compatible Hosts

For generic AI coding assistants, the script configures standard natural language routing:

Use learn-agent-skills to start or resume the Agent Skills path.

These host definitions enable any downstream lesson to rely on a unified, host-agnostic skill name (learn-agent-skills) while concrete invocation syntax gets automatically resolved by the host-specific contract baked into AGENTS.md.

Running the Skill Installation Script

Execute the installer from the root of the skill pack:

# Navigate to the pack directory

cd phases/14-agent-engineering/42-agent-workbench-capstone/outputs/agent-workbench-pack

# Standard installation (aborts if AGENTS.md exists)

./bin/install.sh

# Force overwrite existing installation

./bin/install.sh --force

Verifying Host Contracts

After installation, confirm the host configurations are properly deployed:

# Inspect the installed AGENTS.md for host definitions

cat AGENTS.md | grep -A2 "## Invocation belongs to the host"

This output displays entries for Claude Code, Codex, and generic natural-language hosts, confirming multi-environment compatibility.

Key Files in the Installation Workflow

The skill installation script interacts with several critical components:

• bin/install.sh – The idempotent installer script that validates packs, copies assets, and writes version markers
• AGENTS.md – The host-invocation contract defining how portable skills are called on Claude Code, Codex, and other hosts
• VERSION – The pack version file displayed during installation feedback
• .workbench-version – Hidden version marker written to the target repository
• scripts/init_agent.py – Helper script mentioned in post-installation hints for bootstrapping the agent environment

Summary

• The skill installation script operates through three phases: validation of required files, atomic installation of assets, and user feedback with next steps.
• The script is idempotent by default, refusing to overwrite existing AGENTS.md files unless the --force flag is provided.
• It configures three host types: Claude Code (slash commands), Codex (natural language), and other compatible hosts (generic natural language).
• Core artifacts installed include AGENTS.md, documentation directories, schema definitions, and the .workbench-version marker.

Frequently Asked Questions

What makes the skill installation script idempotent?

The script checks for the presence of AGENTS.md in the target directory before executing any copy operations. If the file exists, the installation aborts immediately unless you explicitly pass the --force flag. This prevents accidental overwrites and ensures repeated runs are safe—subsequent executions simply refuse to modify existing installations without explicit user consent.

Which AI hosts does the installation script configure?

The script configures Claude Code, Codex, and other compatible AI hosts by deploying the AGENTS.md host-invocation contract. Claude Code receives slash-command routing (/learn-agent-skills), while Codex and generic hosts use natural-language invocation patterns (learn-agent-skills). This multi-host configuration allows the same skill pack to operate across different AI coding assistants without manual reconfiguration.

How do I force an overwrite of an existing skill installation?

Append the --force flag when executing the script:

./bin/install.sh --force

This bypasses the existing file detection and replaces AGENTS.md, the docs/, schemas/, and scripts/ directories, and updates the .workbench-version marker. Use this flag when upgrading to a new version of the Agent Workbench pack or when recovering from a corrupted installation.

What files are required for the installation script to proceed?

The script validates the presence of five required pack items before installation: AGENTS.md, VERSION, docs/, schemas/, and scripts/. If any of these components are missing from the pack root ($PACK_ROOT), the script aborts during the validation phase. This ensures only complete, valid skill packs get deployed to target repositories.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}