Hiroshi automatically generates five built-in skills in ~/.hiroshi/skills/ the first time it starts. These skills are ready to use immediately — the agent will call them whenever they are relevant to a task you describe.
Built-in skill scripts are stored at ~/.hiroshi/skills/<name>/ on your filesystem. You can open, read, and edit them at any time to customise their behaviour.
git_manager
The git_manager skill orchestrates Git repository state, runs diffs, commits code changes, and manages branches. Hiroshi calls it whenever you ask about version control operations in your workspace.
Schema
{ "action": "string", "branch": "string", "commit_message": "string" }
| Action | What it does |
|---|
status | Runs git status and returns the output |
diff | Runs git diff and returns unstaged changes |
commit | Stages all changes with git add . then commits with the provided commit_message |
branch | Creates and checks out a new branch named by branch |
Commit all changes with message 'fix: update config'
git_manager with { "action": "commit", "commit_message": "fix: update config" } and report the result.
web_search
The web_search skill queries search engines and returns matching page titles, snippets, and links. Hiroshi calls it when you ask for up-to-date information it does not already know.
Schema
Example
Tell Hiroshi:
Search for 'Rust async best practices'
web_search with { "query": "Rust async best practices" } and summarise the top results for you.
browser_automation
The browser_automation skill provides headless browser scraping and viewport screenshot capabilities. Use it when you need Hiroshi to retrieve or inspect the content of a URL.
Schema
{ "url": "string", "action": "string" }
Fetch the HTML from https://example.com
browser_automation with { "url": "https://example.com", "action": "fetch" } and report back on the page content.
file_janitor
The file_janitor skill tracks file hashes, lists duplicate items, structures chaotic workspace directories, and updates README file indices. It walks the workspace tree, computes an MD5 hash for every file, and returns the full mapping so the agent can identify duplicates.
Schema
Available actions
| Action | What it does |
|---|
scan | Recursively walks . (the workspace), hashes every file, and returns a JSON object mapping paths to MD5 digests |
Scan the workspace for duplicate files
file_janitor with { "action": "scan" } and analyse the results to surface any files that share an identical hash.
task_sync
The task_sync skill syncs project milestones, deadlines, and logs task completion states inside task markdown files. Use it to track progress directly from conversation.
Schema
{ "task_name": "string", "status": "string" }
| Value | Meaning |
|---|
pending | Task is outstanding — written as [ ] in TODO.md |
done | Task is complete — written as [x] in TODO.md |
Mark 'Implement login page' as done
task_sync with { "task_name": "Implement login page", "status": "done" } and append the checked entry to TODO.md. 
