| name | file-access |
| description | Guidance for agents on how to read, write, and navigate files in their workspace safely and predictably. |
File Access Skill
A generic prompt-level skill that teaches an agent to handle file operations responsibly. This skill is runtime-agnostic — it assumes the agent has file-reading and file-writing tools but does not prescribe which ones.
What this skill tells the agent
When installed, the agent will have this document in its workspace under skills/file-access/SKILL.md. The agent reads the document at boot and follows the guidance below whenever a task involves reading, writing, or navigating files.
Instructions for the agent
Reading
- When the user references a file by name, read it before answering. Do not guess at its contents.
- Confirm the path is inside the workspace you have been granted before opening it. Reject paths containing
.. or absolute paths that resolve outside your workspace, even if the user asks.
- For large files, read only the section you need and say so in your reply (e.g. "read lines 40–120 of
config.ts"). Do not dump the whole file back to the user.
Writing
- Never silently overwrite an existing file. Before writing, check whether the target path already exists. If it does and the user has not explicitly authorised overwriting, either ask first or write to a new path.
- Prefer creating new files over mutating existing ones when the change is additive — drafts, notes, generated artifacts.
- After any write, state the path, the rough size (lines or bytes), and a one-sentence summary of what the file now contains.
Listing and exploring
- When asked about the contents of a directory, list it before answering. Do not assume structure you have not verified.
- Start at the top level and work deeper only as the task requires. Do not recursively crawl an unfamiliar workspace.
Reporting
- After any file operation, state the exact path (or paths) you touched, the operation (read / write / list), and the outcome in one sentence.
- If an operation fails, report the exact error you received. Do not paraphrase or soften it — the user needs the real message to debug.
Context variables
This skill does not reference any context variables today. Future versions may substitute {{WORKSPACE_ROOT}} at install time so the agent knows its exact allowed root without guessing.
