← Về thư mục
📄 / / proc / 45 / cwd / root / .hermes / skills / software-development / kanban-content-batching / SKILL.md

name: kanban-content-batching description: Batch-based Kanban workflows for large shared content edits, taxonomy/tagging, and final human review gates. category: software-development

Kanban content batching

Use this skill when a shared document or knowledge base needs large-scale cleanup, tagging, grouping, or phased rewriting.

When to use

Trigger this skill when the work has at least one of these traits:

  1. Many sources or sections.
  2. Multiple phases that can overlap.
  3. A risk of double work, task sprawl, or missed verification.
  4. A need to update shared master files while preserving auditability.
  5. A final human review gate after all automation is complete.

Core workflow

1. Split the work into phase tasks

Create a parent task for the whole job. Then create child tasks for distinct phases such as:

  1. taxonomy or rules design
  2. batch repair or tagging for file A
  3. batch repair or tagging for file B
  4. final reordering or consolidation

Keep each child task narrow. Do not mix taxonomy design with tagging or reordering.

2. Batch small and verify each batch

For content-heavy edits, process small batches instead of one large sweep. A good default is 3 to 5 items per batch.

For each batch:

  1. read or parse the relevant slice
  2. update only that slice
  3. verify required fields or formats are present
  4. write only when the batch passes verification

If a batch fails, stop the batch. Do not write partial results unless the workflow explicitly allows it.

3. Separate deterministic work from LLM work

Prefer local parsing, regex, and structured validation for large master files. Use LLM calls only for the narrow steps that need reasoning or translation. If a provider returns rate limits or timeouts, reduce batch size or switch that step to local processing.

4. Preserve auditability

Write clear task comments or summaries after each phase. Use the parent task to track the overall state. Use children for the actual work items. Do not create extra fix tasks unless the user explicitly wants a separate audit trail.

5. Hold final review until the end

If the user says they only want one final review, keep the parent task blocked at the end. Continue through all planned subtasks without asking for approval between batches. Only pause for review after all planned work is done.

Tagging and grouping pattern

For taxonomy-heavy work, use this order:

  1. Define the group taxonomy.
  2. Apply tags to each source or section.
  3. Reorder the file by group.
  4. Verify the final structure.
  5. Block the parent task for human review.

Never combine taxonomy definition and file rewriting in the same child task.

Manual Gates and Handoffs

For high-stakes research or content pipelines where the user must approve each phase:

  1. Manual Gate Rule: In the task body, explicitly state that after completion (writing file, adding comment), the worker MUST stop and wait for user approval. It must NOT advance the status of the next task even if it is the owner.
  2. Critique Handoff: When a phase (like a Critique task) finds issues, do NOT set the task to blocked (which stops the pipeline). Instead:
  3. Record the critique findings as a Task Comment.
  4. Set the task to done.
  5. The next phase (Synthesis/Reporting) is then responsible for reading the previous task's comments and incorporating the feedback.
  6. Master Report Stewardship: Prohibit "overwrite" behavior. Every synthesis phase must read the existing master file and append/integrate new findings to maintain research inheritance.

Pitfalls

  1. A worker that exits cleanly without kanban_complete or kanban_block is a protocol violation. Check logs and rerun with explicit terminal state handling.
  2. Do not let one failed batch block the whole project from the beginning. Resume from the next unconfirmed slice when safe.
  3. Do not auto-complete a content task when the user asked for final manual review.
  4. If rate limits appear, shrink batch size before increasing prompt size.
  5. Overwrite Trap: Losing data from previous batches by writing a fresh file instead of reading and extending the master report.
  6. Blocked Pipeline: Using status='blocked' on a critique task which prevents subsequent phases from even entering a 'ready' state. Use done + comments instead.

Verification checklist

Before moving to the next phase, confirm:

  1. the batch was written
  2. required fields are present
  3. no placeholder or corrupted text remains
  4. the task tree still maps cleanly to the phase plan
  5. the parent task has a clear final review state

This skill may use session-specific notes or reusable scripts in:

  1. references/content-batching-patterns.md
  2. scripts/verify_master_batch.py