humanize-redo

/humanize-redo — second polishing / partial re-run

Based on cwd, the most recent _workspace/{run_id}/is found and humanize-korean , then the skill's strict polishing (Phase B) is called again.

User instructions

$ARGUMENTS

Globwith _workspace/YYYY-MM-DD-*/final.md(or 01_input.txt) to match and identify the latest run_id . If none, guide with "There is no previous run. Start with /humanize" and exit.
Parse user instructions:
- Category specification("translationese only", "idioms only", "emoji only") → re-polish only findings in that category
- Paragraph specification("only this paragraph", "only the second paragraph") → only findings in that range
- Intensity adjustment("lower intensity" / "conservatively" → S1 only, "raise intensity" → S1+S2+S3)
- Rollback request("revert this change") → process that edit as a content-fidelity-auditor rollback
- No instructions / "do a second polishing" → round 2 for all residual findings
korean-style-rewriter re-call input: the existing 02_detection.json or 05_naturalness_review.json's residual findings + user instructions are passed as target_filter.
Artifacts are version-separated as 03_rewrite_v2.md(or v3). Backed up as the previous final.md는 final_prev.md.
strict Phase C parallel verification → Phase D final output (change comparison table + new grades).

Maximum round 3. If unresolved beyond that, recommend human review with hold_and_report.