AgentSkillsCN

mem-editing

内部技能，用于修改说明文件。可通过 /mem-edit 命令或其他需要编辑说明的命令调用。无准入限制——直接与调用者协同完成流程。

SKILL.md
--- frontmatter
name: mem-editing
description: Internal skill for instruction file changes. Called by /mem-edit command or other commands needing instruction edits. No gate - flows with caller.
argument-hint: [files or description]

Modify Instructions

Edit instruction files with full analysis workflow. Use for any instruction change: optimization, bug fixes, adding rules, refactoring.

Target: $ARGUMENTS (files or description of change)

Task Tracking

FIRST: Create one TaskCreate per row below BEFORE any other work. Mark in-progress/completed as you proceed:

#SubjectDescription
1Ensure scope identifiedSkip if $ARGUMENTS is file path, else ask
2Gather contextRead primary files, linked files, grep for related. For each related file, add sub-task "Read: [file]"
3Analyze and clarifyCheck conflicts, redundancy. For each ambiguity about user intent, use AskUserQuestion to clarify.
4Report findingsFiles affected, before/after, rationale. If multiple valid approaches, use AskUserQuestion to get direction.
5Ensure jj changeCreate new change if working copy clean, skip if already have uncommitted changes
6Apply changesEdit files, verify consistency across all affected
7CommitMessage describing changes

Instructions

STOP rushing. Invest thinking tokens now to save iteration tokens later.

  1. •

    Ensure scope identified:

    • •If target is a file path → that's the primary file
    • •If target is a description → identify which file(s) need changes
  2. •

    Gather context - Read primary file(s) and all linked files (@docs references):

    • •Search for related files
    • •Grep for key concepts/terms in other instruction files
    • •Check commands, skills, docs that reference the same concepts
    • •For each related file found, add sub-task "Read: [file]"
  3. •

    Analyze and clarify thoroughly (ultra, deeply, freakingly, super ultrathink!):

    • •Speak your mind LOUDLY. Don't just use a thinking block, but tell me everything you have in mind.
    • •Think through each instruction as a fresh agent - what could be misinterpreted?
    • •Check for redundancy and conflicts ACROSS files
    • •For each ambiguity about user intent, use AskUserQuestion to clarify
    • •Identify all changes needed (may span multiple files)
    • •Apply best practices from supporting docs
  4. •

    Report findings:

    • •Files affected
    • •Before/after for each change
    • •Rationale
  5. •

    Ensure jj change - Check jj status. If working copy is clean, create new change. If already have uncommitted changes, use current change.

  6. •

    Apply changes - Preserve all salient information, verify consistency across all affected files.

  7. •

    Commit - Commit changes with descriptive message.

What to Check

When reviewing instructions, look for:

  • •Ambiguity - what could a fresh agent misinterpret?
  • •Cross-file conflicts - do related files have contradicting rules?
  • •Redundancy - is this duplicated elsewhere?
  • •Missing context - does this assume knowledge not provided?

Supporting Files

  • •@core.md: Core principles - self-verification, minimal info, writing style, examples, anti-patterns
  • •@commands-skills.md: Slash command and skill structure, task tracking guidelines
  • •@instructions.md: CLAUDE.md, memory files, structured prompting (XML tags)
  • •@agents.md: Agent structure and patterns