AgentSkillsCN

feedback

运用写作导师框架对作者提交的作品进行评审

SKILL.md
--- frontmatter
name: feedback
description: Review writer submissions using the Writing Mentor Framework
user-invocable: true

Review Submissions

Review all submissions in this folder using the Writing Mentor Framework.

What This Does

  1. •Checks for required dependencies (prompts to install if missing)
  2. •Extracts text from all submissions (DOCX, PDF, XLSX)
  3. •Renders Excel charts for visual review
  4. •Reviews each submission with isolated context (prevents bias)
  5. •Writes two-tier feedback (Reviewer Notes + Writer Feedback)

Prerequisites

Before running, ensure you have:

  • •assignment.md - Your assignment requirements
  • •rubric.md - Your evaluation criteria
  • •submissions/ - Folder containing writer work

Optional:

  • •course_concepts.md - Domain concepts for assumption validation
  • •turnitin/ - Similarity reports

Workflow

<workflow> 1. **Read the config** from `wmf-config.yaml` to get the framework path
  1. •

    Check dependencies by running:

    bash
    python {framework_path}/skills/feedback/scripts/check_dependencies.py

    If dependencies are missing, prompt the user to install them before continuing.

  2. •

    Verify required files exist:

    • •assignment.md (required)
    • •rubric.md (required)
    • •submissions/ folder with at least one file (required)

    If any are missing, tell the user what's needed and stop.

  3. •

    Extract text from submissions:

    bash
    python {framework_path}/skills/feedback/scripts/extract_submission_text.py --input submissions --out feedback_extracted
  4. •

    Render Excel charts (if any .xlsx files exist):

    bash
    python {framework_path}/skills/feedback/scripts/render_xlsx_quicklook.py --input submissions --out feedback_rendered
  5. •

    Read the framework instructions from {framework_path}/skills/feedback/SKILL.md

  6. •

    Read reference materials:

    • •assignment.md
    • •rubric.md
    • •{framework_path}/skills/feedback/references/economical_writing_principles.md
    • •course_concepts.md (if present)
  7. •

    Enumerate submissions by parsing filenames. Group by username.

  8. •

    Read parallelism setting from wmf-config.yaml: review.max_parallel_agents (default: 3)

  9. •

    Process submissions in parallel batches:

    • •For each batch of N submissions (where N = max_parallel_agents):
      • •Spawn N Task agents in a single message with subagent_type: "general-purpose"
      • •Each agent follows the isolated review workflow from the framework SKILL.md
      • •Wait for all N to complete before starting the next batch
    • •Each agent writes to feedback/{username}.md
  10. •

    Report completion with summary of submissions reviewed.

    </workflow>

Parallelism Settings

Configure review.max_parallel_agents in wmf-config.yaml:

SettingUse Case
1Sequential mode — use if you experience rate limits or errors
3Default — good balance of speed and reliability
5-10Large classes with hundreds of submissions

Multi-Round Reviews

If submissions are organized in rounds (submissions/round1/, submissions/round2/):

  1. •Check wmf-config.yaml for submissions.rounds.enabled: true
  2. •Use round-specific folders: feedback_extracted_round{N}/, feedback_round{N}/
  3. •For resubmissions, compare to prior round feedback per config setting

Output

Feedback files are written to feedback/ (or feedback_round{N}/ for rounds).

Each file contains:

  • •Section A: Reviewer Notes - Technical audit for instructor only
  • •Section B: Writer Feedback - Teaching-focused guidance to share with writer