Cursor Edit

Cursor Edit

Overview

This skill outsources only the act of editing files to Cursor's headless CLI (cursor-agent). Use the same contract from Claude Code, Codex, or any other AI agent.

• Calling AI agent (brain): reads and searches code, decides what to change, writes edit instructions, and verifies the result
• Cursor (hands): rewrites files according to the received instructions

The goal is to move heavy editing work to Cursor, reducing token usage in the calling AI agent and speeding up the overall task.

This only pays off for larger edits. For small changes, the cost of writing self-contained instructions and reviewing git diff can exceed the cost of direct editing. Do not force this skill onto trivial one-line or few-line edits.

When To Use

These tendencies apply:

• Not suitable for trivial edits (one to a few lines). Startup, instruction writing, and diff review often cost more than direct editing by the calling AI agent.
• Best for edits that are large, pattern-heavy, and have good sibling files to imitate. Example: adding tests that follow an existing test suite. cursor-agent can closely follow good sibling examples and often produce convention-compliant results quickly.
• Be careful with new features or non-obvious assumptions when no good sibling example exists. cursor-agent tends to pattern-match without semantic validation, so correctness is not guaranteed. In this case, the calling AI agent must separately verify semantic correctness. See "Required Verification".
• Cursor consumes the delegated tokens. The calling AI agent saves tokens, but the speed and quality advantages mainly apply under the suitable conditions above.

Operating Contract

• The calling AI agent must not use direct editing tools such as Edit, Write, or apply_patch.
• All file rewrites must go through scripts/cursor-agent-print.sh, which calls cursor-agent.
• The calling AI agent remains responsible for code reading, search, planning, writing edit instructions, git operations, build/test/lint verification, TODO management, and persistent memory updates.
• This is a soft contract. Direct editing tools are not technically blocked by hooks; the calling AI agent must follow this as discipline.

Editing Loop

1. Analyze and design — The calling AI agent reads the code and decides what to change in which files.
2. Write edit instructions — State the target files and desired changes unambiguously. Follow the guidance below.
3. Delegate — Pass the instructions to cursor-agent-print.sh so cursor-agent edits the files.
4. Verify — Review Cursor's actual changes with git diff, then build/test/lint the changed scope.
5. Retry — If the result diverges from the intent, return to step 2 with more specific instructions.

Calling cursor-agent

Call the wrapper from the target repository root. If working in a worktree, run it from the worktree root. cursor-agent reads and writes files relative to the current directory.

<cursor-edit>/scripts/cursor-agent-print.sh "<edit instructions>"

<cursor-edit> is the directory where this skill is installed.

Internally, the wrapper runs:

cursor-agent -p --force --trust --output-format text --model "${CURSOR_MODEL:-composer-2.5-fast}" "<edit instructions>"
# -p      : non-interactive print mode
# --force : apply file changes without confirmation
# --trust : trust the workspace without prompting

The default model is Composer2.5 Fast (composer-2.5-fast). Set CURSOR_MODEL to override it.

CURSOR_MODEL=gpt-5.5 <cursor-edit>/scripts/cursor-agent-print.sh "<edit instructions>"

Writing Good Edit Instructions

cursor-agent is another AI. If instructions are vague, Cursor can edit in the wrong direction. Write instructions that are self-contained.

• Name the target file paths using relative or absolute paths.
• Name existing sibling files to imitate. Example: "Read the existing foo-1.test.ts through foo-6.test.ts files and follow their conventions." This is the strongest lever for quality and convention matching. cursor-agent follows imports, naming, helpers, and local patterns much better when it has good examples.
• Describe the before and after concretely. Quoting the relevant code improves accuracy.
• Include acceptance criteria. Example: the build must pass, or function signatures must not change.
• Constrain the scope. Example: "Do not modify any files other than the specified files."
• Split large changes. Do not pack too much into a single delegation.

Good instruction example:

In server/src/domain/foo/foo.ts, add an early return at the start of Foo.validate()
that returns new EmptyNameError() when this.name.trim() === "".
Do not change the existing logic.
Do not edit any files other than the specified file.

Scope And Safety

• Delegate: project code and configuration files.
• Do directly in the calling AI agent: git operations, persistent memory updates, and TODO management.
• Do not let cursor-agent roam freely. Always name the files it is allowed to edit and keep it out of unrelated scope.
• When using a worktree: run cursor-agent-print.sh from the worktree root and include "do not modify files outside this worktree" in the instructions.

Required Verification

• After editing, always review Cursor's actual changes with git diff. Do not trust the text output from cursor-agent alone.
  • To preserve the token savings, review the diff first instead of rereading whole files.
• Confirm where files were generated. Because cursor-agent runs from the current working directory, verify that worktree edits stayed inside the worktree and did not dirty the parent repository, for example with git -C <worktree> status.
• cursor-agent does not verify semantic assumptions; it mainly pattern-matches sibling files. git diff and builds do not prove the semantic correctness of claims or logic. For new or non-obvious logic, such as whether an assertion is really valid or whether an API actually exists, the calling AI agent must verify the behavior against use cases and sibling tests. Be especially careful when there is no good sibling example. Confirm with search before claiming that something is absent or nonexistent.
• Build, test, and lint the changed scope according to the project's conventions.

Gotchas

• If cursor-agent is missing or authentication fails, report that the execution environment does not meet the prerequisite and stop.
• If cursor-agent fails or edits against the instructions, make the instructions more specific and retry.
• If retrying does not resolve the issue, report the situation to the user and ask how to proceed. Even though this is a soft contract, the calling AI agent must not silently switch back to direct editing.