What Are Agent Commands in the AI Job Search Framework?

Agent commands in the AI Job Search Framework are markdown-driven workflow specifications stored in the .claude/commands/ directory that interpret slash commands like /apply or /setup as deterministic, reproducible job application pipelines.

The MadsLorentzen/ai-job-search repository implements a declarative command architecture where natural language instructions become executable code. Each agent command is a standalone markdown document that Claude Code parses sequentially, orchestrating file system operations, external API calls, and skill invocations to automate the entire job search lifecycle.

How Agent Commands Work

Agent commands function as the primary interface between the user and the framework's automation capabilities. They transform simple text triggers into complex, multi-step workflows while maintaining strict data boundaries.

File Structure and Invocation

Agent commands reside as .md files within the hidden .claude/commands/ directory. When a user types a slash command such as /apply, /setup, or /rank, Claude Code locates the corresponding file—for example, .claude/commands/apply.md or .claude/commands/setup.md—and executes the declared workflow step by step. The public command catalog is documented in README.md between lines 147-162, which maps each slash command to its underlying markdown specification.

Execution Flow and Data Handling

Each markdown file contains a step-by-step recipe that may invoke other commands or call specialized skills. For instance, the /apply command internally calls /rank to score job postings before generating application materials. Commands operate on a data-only model: they read from and write to the framework's own data files including documents/, job_search_tracker.csv, and outcome.md, but never modify external resources without explicit user consent. The permission allowlist in .claude/settings.json enforces this boundary by guaranteeing only documented commands can be invoked by the model.

Core Command Categories

The framework organizes functionality into four distinct categories, all implemented as discrete markdown command files in .claude/commands/.

Core Workflow Commands

These commands bootstrap the system and execute the primary job application pipeline:

• /setup – Initializes candidate profiles from the documents/ directory or via interactive interviews, as defined in .claude/commands/setup.md.
• /scrape – Invokes portal-specific CLI tools located under .agents/skills/*/cli to fetch job postings from supported Danish job boards.
• /apply – Runs the full draft-review-revise pipeline, reading job_search_tracker.csv and outputting finalized PDFs.

Post-Apply Utilities

Manage interview preparation and result tracking:

• /interview – Prepares interview packs from processed applications.
• /outcome – Updates outcome.md and synchronizes status back to job_search_tracker.csv.
• /gmail-sync – Auto-detects application status from Gmail threads.
• /notion-sync – Generates a read-only Notion view of the tracker without modifying external data.

Enhancement and Analysis

Optimize application quality through scoring and enrichment:

• /rank – Batch-scores scraped postings against the fit framework defined in .claude/commands/rank.md.
• /expand – Enriches existing profile documents with additional context.
• /upskill – Generates learning plans based on identified skill gaps.
• /html-report – Renders an offline dashboard from tracker data for local analysis.

Customization and Maintenance

Extend framework capabilities for specific user needs:

• /add-template – Registers custom LaTeX or Typst templates by specifying source paths, compile commands, and font dependencies.
• /add-portal – Scaffolds new search skills for unsupported job boards under .agents/skills/.
• /reset – Optionally wipes personal data from the repository while preserving command and skill definitions.

Integration with Skills

Agent commands do not encapsulate business logic directly. Instead, they invoke skills—modular capabilities defined in .claude/skills/—to perform specialized tasks. The job-application-assistant skill, documented in .claude/skills/job-application-assistant/SKILL.md, provides the core evaluation framework that /apply and /rank utilize to score job fit against candidate profiles. This architectural separation allows commands to remain purely declarative while skills encapsulate complex, reusable logic.

Practical Usage Examples

Execute these commands within Claude Code to operate the framework:

# Initialize your candidate profile from existing documents

/setup

# Scrape supported portals and generate a ranked shortlist

/scrape
/rank

# Apply to a specific job posting with full automation

/apply https://jobindex.dk/job/1234567

# Register a custom CV template and scaffold a new portal skill

/add-template
/add-portal

# Synchronize outcomes to Notion (read-only operation)

/notion-sync

Summary

• Agent commands are markdown specifications stored in .claude/commands/ that define executable workflows for the AI Job Search Framework.
• Execution follows a data-only model, restricting modifications to repository-local files like job_search_tracker.csv and outcome.md.
• The command ecosystem spans four categories: core workflow, post-apply utilities, enhancement tools, and customization generators.
• Security is enforced through .claude/settings.json, which maintains an allowlist of permitted slash commands.
• Skills located in .claude/skills/ provide the underlying logic, while commands in .claude/commands/ provide the orchestration layer.

Frequently Asked Questions

Where are agent commands stored in the repository?

Agent commands reside as individual markdown files within the .claude/commands/ directory. According to the source code structure, each filename corresponds directly to its slash command invocation (e.g., setup.md for /setup and apply.md for /apply).

How do I add a new agent command to the framework?

Create a new .md file in .claude/commands/ containing the step-by-step workflow instructions in markdown format. The continuous integration pipeline automatically lints the syntax and validates the new command against the permission allowlist defined in .claude/settings.json.

What is the difference between an agent command and a skill?

Agent commands are high-level orchestrators stored in .claude/commands/ that manage workflows and coordinate data flow between files. Skills are modular capabilities located in .claude/skills/ that encapsulate specific domain logic, such as the job-application-assistant skill used by /apply and /rank to evaluate job fit.

Can agent commands modify files outside the repository?

No. The framework implements a strict data-only philosophy where commands read and write only within the repository's own file structure, specifically targeting documents/, job_search_tracker.csv, and outcome.md. External systems like Gmail or Notion are accessed read-only or via explicit user-approved integrations, ensuring the permission model in .claude/settings.json cannot be bypassed.

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