Bulk Edit Skill

Process bulk edits from a comments/feedback file safely using sequential agents.

Usage

code

/bulk-edit <comments-file> [--target <file>]

Examples:

code

/bulk-edit specs/comments-18-jan-mary-gdocs.md --target main.tex
/bulk-edit specs/reviewer-feedback.md

How It Works

For each uncompleted thread in the comments file:

•
Classify complexity of the edit:
- •Simple (haiku): typos, wording tweaks, small clarifications, dropping sentences
- •Medium (sonnet): rewriting paragraphs, adding explanations, moving content
- •Complex (sonnet/opus): structural changes, narrative flow, multi-part edits
•
Spawn a fresh agent with appropriate model:
- •Agent reads ONLY that specific thread
- •Agent applies the edit to target file
- •Agent marks thread as completed with ✅ prefix
- •Agent returns 1-line summary
•
Agent is discarded (context freed)
•
Continue to next uncompleted thread

Execution Instructions

When this skill is invoked, follow this workflow:

Step 1: Parse the comments file

Read the comments file and identify all threads. A thread is completed if it starts with ✅.

python

# Pseudocode for identifying threads
threads = []
for section in comments_file.split("## Thread"):
    if section.strip():
        thread_num = extract_thread_number(section)
        completed = section.strip().startswith("✅") or "## Thread" line contains "✅"
        threads.append({"num": thread_num, "content": section, "completed": completed})

Step 2: For each uncompleted thread

For each thread where completed == False:

A. Classify complexity:

Complexity	Indicators	Model
Simple	"typo", "nit", "drop this", "clearer:", single word change	haiku
Medium	"rewrite", "clarify", "explain", paragraph-level change	sonnet
Complex	"restructure", "move to", multiple parts, narrative changes	sonnet

B. Spawn agent:

code

Task tool:
  subagent_type: "general-purpose"
  model: <haiku|sonnet based on complexity>
  prompt: |
    Apply this reviewer comment to {target_file}:

    ---
    {thread_content}
    ---

    Instructions:
    1. Read {target_file}
    2. Find the relevant section (use the quoted/highlighted text as anchor)
    3. Apply the edit as described in the comment
    4. If there's a reply with more context, incorporate that too
    5. After editing, mark this thread in {comments_file} with appropriate status:
       - ✅ = fully addressed
       - ➖ = N/A (comment is praise, informational, or no edit needed)
       - ⏳ = partially addressed (some aspects done, others need more work)
       - ⚠️ = needs manual attention (unclear, conflicting, or requires user decision)
       - 6. Return a 1-line summary: "Thread {N} [status]: <what you did or why>"

    Examples:
    - "Thread 5 ✅: Changed 'performs worse than random' per suggestion"
    - "Thread 12 ➖: Comment was praise ('I love this!'), no edit needed"
    - "Thread 23 ⏳: Added explanation but TODO for Yulong to verify numbers"
    - "Thread 34 ⚠️: Unclear what 'restructure this' means - needs user input"

C. Wait for agent to complete, then continue to next thread.

Step 3: Summary

After all threads processed, report:

code

Processed: 66 threads
  ✅ Addressed: 45
  ➖ N/A: 10
  ⏳ Partial: 5
  ⚠️ Needs attention: 6

Progress Tracking

Progress is tracked IN the comments file itself:

Marker	Meaning	Action
`## 📋 Thread 1`	Not started	Will be processed
`## ⏳ Thread 1`	Partially addressed	Will be processed (needs more work)
`## ✅ Thread 1`	Addressed/completed	Skipped
`## ➖ Thread 1`	N/A / not applicable	Skipped (comment doesn't require edit)
`## ⚠️ Thread 1`	Needs manual attention	Skipped (agent couldn't resolve)

When to use each:

•✅ Addressed: Edit fully applied as requested
•➖ N/A: Comment is informational, praise, or already resolved
•⏳ Partial: Some aspects addressed, others remain
•⚠️ Needs attention: Unclear request, conflicting info, or requires user decision

Resumable: Re-run /bulk-edit - only 📋 and ⏳ threads are processed.

Safety Rules

•One agent at a time - never parallel agents on same file
•Fresh agent per edit - prevents context bloat
•Mark progress in source - enables resume
•Unclear edits get ⚠️ - don't guess

Model Selection Guide

Edit Type	Examples	Model
Typo/nit	"typo", "s/foo/bar"	haiku
Drop content	"consider dropping", "remove this"	haiku
Wording	"clearer:", "maybe rephrase"	haiku
Add clarification	"explain why", "add context"	sonnet
Rewrite paragraph	"restructure this", "rewrite to..."	sonnet
Move content	"move to methods section"	sonnet
Structural	"split into subsections", "reorder"	sonnet
Narrative	"the flow is off", "argument unclear"	sonnet
Multi-part	Thread with multiple sub-comments	sonnet

Example Session

code

User: /bulk-edit specs/comments-mary.md --target main.tex

Claude: Processing 66 threads from specs/comments-mary.md → main.tex

Thread 1 (simple): Spawning haiku agent...
  → Thread 1: Fixed typo "sandbaggs" → "sandbags"

Thread 2 (medium): Spawning sonnet agent...
  → Thread 2: Rewrote abstract opening for clarity

Thread 3 (simple): Spawning haiku agent...
  → Thread 3: Dropped redundant sentence in intro

...

Summary:
  ✅ Addressed: 52/66
  ➖ N/A: 8 (praise, informational)
  ⏳ Partial: 3 (threads 23, 41, 55 - need verification)
  ⚠️ Needs attention: 3 (threads 34, 51, 58 - unclear)

Re-run to process ⏳ threads, or address ⚠️ manually.

Notes

•Agents use model parameter in Task tool to select haiku vs sonnet
•Each agent gets ~full context window for its single edit
•Orchestrator context grows ~150 tokens per edit (sustainable for 100+ edits)
•Comments file becomes the source of truth for progress tracking.