edit-prompts

name	edit-prompts
description	Read, edit, and diff .claude prompt assets (SKILL.md, agent definitions) in the supervised repo.
user_invocable	true
arguments	[{"name":"action","description":"Action: list, read <name>, edit <name>, diff <name>, history","required":true}]

/edit-prompts — Prompt Asset Editor

This skill is the only way to read or modify .claude/ files in the supervised repo. Direct Read/Edit/Write calls to files under the supervised repo's .claude/ directory are forbidden — use this skill instead.

Why

Every prompt asset change must be:

1. Logged — recorded in .supervisor/prompt-edits.jsonl with timestamp, sha1, and diff stats
2. Diffed — the full unified diff is shown before and after edits
3. Traceable — so the researcher can correlate prompt changes with metric changes across runs

Available Assets

Use pixi run researcher-dot-claude-list to discover available assets. The asset names are derived from harness.toml configuration (skill name + agent names).

You can also use paths relative to .claude/ (e.g., rules/some-rule.md).

Actions

List all assets

pixi run researcher-dot-claude-list
pixi run researcher-dot-claude-list --json

Read an asset

pixi run researcher-dot-claude-read skill
pixi run researcher-dot-claude-read <agent-name>

Or use the CLI directly and read the output.

Edit an asset

To edit, you must:

1. Read the current content with researcher-dot-claude-read
2. Prepare the new content
3. Pipe it to researcher-dot-claude-edit:

echo "new content" | pixi run researcher-dot-claude-edit skill

In practice as Claude Code, the workflow is:

1. Bash("eval \"$(pixi shell-hook)\" && PYTHONPATH=src python -m supervisor_harness.cli prompt-read skill") — read current content
2. Make your changes to the content
3. Bash("eval \"$(pixi shell-hook)\" && PYTHONPATH=src python -m supervisor_harness.cli prompt-edit skill <<'PROMPT_EOF'\n<new content>\nPROMPT_EOF") — write and get diff

Diff without writing

echo "proposed content" | pixi run researcher-dot-claude-diff skill

View edit history

pixi run researcher-dot-claude-history
pixi run researcher-dot-claude-history --json

Workflow for Prompt Changes

1. Always snapshot first: pixi run researcher-snapshot --label before-edit
2. Read the asset you want to change
3. Edit with the new content — review the diff output
4. Snapshot after: pixi run researcher-snapshot --label after-edit
5. Restart the loop: pixi run researcher-loop