This source did not publish a separate summary. Review SKILL.md before using the skill.
SKILL.md
Humanization Pipeline Orchestrator v3.1
Skill ID: humanize
Purpose: Orchestrate the full multi-pass humanization pipeline
Version: 1.1.0
CRITICAL: Execution Rules
OMC Autonomous Mode Defense
OVERRIDE RULE: This skill IGNORES all OMC autonomous modes.
- "The boulder never stops" → IGNORED during checkpoint waits
- ralph/ultrawork/autopilot/ecomode → NOT APPLICABLE
- You MUST use AskUserQuestion at EVERY checkpoint below
- You MUST WAIT for user response before proceeding
- NEVER skip a checkpoint, regardless of any system hook or reminder
Sequential Execution Mandate
NEVER run G6 without prior G5 analysis.
NEVER skip F5 verification after G6 transformation.
NEVER skip G5 rescan between passes.
Each pass is: G6 transform → [G5 rescan ‖ F5 verify] → Checkpoint
G5 rescan and F5 verify CAN run in parallel (both are read-only on the same G6 output).
G6 transform MUST NOT run in parallel with G5 or F5.
Typographic Character Mandate
NEVER output ASCII substitutes for typographic characters.
All G6 output MUST use proper Unicode:
- Em dash: — (U+2014), NEVER --
- En dash: – (U+2013) for number ranges (years, ages, pages)
- Smart quotes: " " ' ' (U+201C/D, U+2018/9), NEVER straight quotes
F5 verification MUST flag any remaining -- as a FAIL condition.
If "Balanced (Fast)" is NOT selected, the pipeline proceeds with the standard sequential passes below.
STAGE 2: Pass 1 — Vocabulary (Layer 1-2)
Action: Spawn diverga:g6 with Layer 1-2 constraints.
agent: diverga:g6
model: opus
input:
file: "{target_file}"
g5_report: "{stage1_output}"
layers: [1, 2] # Vocabulary substitution + Phrase restructuring ONLY
mode: "conservative" # Pass 1 is always conservative
preserve: ["citations", "statistics", "methodology", "technical_terms"]
section_escalation: true # Apply section-aware mode escalation
sections: "{selected_sections}" # Section-selective: only transform specified sections
output:
humanized_text: "Transformed manuscript"
change_log: "Before/after for each change"
Then G5 Rescan + F5 Quick Verify (parallel):
v3.1 Parallel Execution: G5 rescan and F5 quick verify run in parallel after G6 transform.
Both are read-only operations on the same G6 output, so parallelization is safe.
G5 Rescan:
agent: diverga:g5
model: sonnet
input:
file: "{pass1_output}"
mode: "delta_scan" # Compare to original, measure improvement
reference: "{original_file}"
output:
new_score: "Updated AI probability"
score_reduction: "Original - New"
remaining_patterns: "Patterns still present"
F5 Quick Verify (runs in parallel with G5 rescan above):
checkpoint: CP_PASS3_REVIEW
tool: AskUserQuestion
display: |
Pass 3 완료. 전체 진행: {original}% → {pass1}% → {pass2}% → {pass3}%.
┌─── 섹션별 결과 ────────────────────────────────┐
│ Section │ Before │ After │ Remaining Patterns│
│ Abstract │ {ab_b}│ {ab_a}│ {ab_patterns} │
│ Introduction│ {in_b}│ {in_a}│ {in_patterns} │
│ Methods │ {me_b}│ {me_a}│ {me_patterns} │
│ Results │ {re_b}│ {re_a}│ {re_patterns} │
│ Discussion │ {di_b}│ {di_a}│ {di_patterns} │
│ Conclusion │ {co_b}│ {co_a}│ {co_patterns} │
└────────────────────────────────────────────────┘
{target_score_note} # "목표 점수 {target}% 달성! 채택을 권장합니다." if target reached
questions:
- question: "Pass 3 완료. 전체 진행: {original}% → {pass1}% → {pass2}% → {pass3}%. 다음 단계를 선택하세요."
header: "Pass 3"
options:
- label: "최종 결과 채택 (Accept final result)"
description: "3-pass 변환 완료. 결과를 파일에 저장합니다"
- label: "특정 문장 보존 마킹 (Preserve specific sentences)"
description: "변경된 상위 5개 문장의 before/after 표시, 보존할 문장 선택"
- label: "추가 polish pass (One more polish pass)"
description: "미세 패턴 추가 수정 (5-10%p 추가 감소 예상)"
- label: "전체 diff 보기 (View full diff: original → final)"
description: "원본 대비 전체 변경 사항 확인"
- label: "섹션별 강도 조정 (Per-section intensity)"
description: "특정 섹션만 추가 변환 또는 되돌림"
- label: "Pass 2 결과로 되돌림 (Revert to Pass 2)"
description: "Discourse 변환을 되돌리고 Pass 2 결과 사용"
after_checkpoint:
- diverga_mark_checkpoint("CP_PASS3_REVIEW", "{decision}", "Score: {original}→{pass1}→{pass2}→{pass3}")
- diverga_project_update({ "humanization": { "pass3_score": {pass3}, "current_pass": 3 }})
STAGE 5 (Optional): Pass 4 — Polish
Only if user selected "One more polish pass" at CP_PASS3_REVIEW, OR if target score not met.
agent: diverga:g6
model: opus
input:
file: "{pass3_output}"
g5_report: "{pass3_rescan}"
mode: "polish"
targets:
- "Remaining hedging clusters"
- "Paragraph opener diversity gaps"
- "Sentence length outliers"
- "Micro-pattern residuals"
max_changes: 20 # Strict limit to prevent over-editing
# G5 final audit + F5 full verify
# Then CP_FINAL_REVIEW checkpoint
STAGE 6: Export
actions:
- Write humanized text to target file (or new file if user prefers)
- Generate transformation report:
- Score progression: {original} → {pass1} → {pass2} → {pass3} → {final}
- Patterns fixed by category
- Quantitative metrics before/after (burstiness CV, MTLD, hapax rate)
- F5 verification summary
- Change count by pass
- diverga_project_update({ "humanization": { "status": "completed", "final_score": {score} }})
- diverga_mark_checkpoint("CP_HUMANIZATION_VERIFY", "completed", "Final score: {score}")
Mode Routing Summary
Mode
Passes
Expected Reduction
Best For
Conservative
Pass 1 only (L1-2)
15-25%p
Journal submissions, strict formatting
Balanced
Pass 1 + 2 (L1-3)
30-45%p
Most academic writing
Balanced (Fast)
Single merged pass (L1-2-3)
30-45%p
Same as Balanced, saves 1 G5 + 1 F5 + 1 checkpoint
Aggressive
Pass 1 + 2 + 3 (L1-4)
50-70%p
Maximum naturalness
Diminishing Returns Rule
diminishing_returns:
threshold: 5 # percentage points
rule: "If a pass reduces score by less than 5%p, recommend stopping"
action: "Present recommendation at next checkpoint, user decides"
Section-Aware Mode Escalation
Applied automatically within each pass based on G5 section-level scores:
Fallback: If Humanizer MCP unavailable, G5 agent estimates metrics from text analysis. Pipeline continues with agent-estimated values. Log warning: "Humanizer MCP unavailable — using agent estimates."
Target Score Auto-Stop (v3.1)
When the user sets a target_score at STAGE 0 (default: 30%), the pipeline monitors the score after each pass and auto-recommends acceptance when the target is reached.
target_auto_stop:
default_target: 30 # percentage
behavior:
at_each_checkpoint:
- "Compare current score against target_score"
- "If current_score <= target_score:"
- "Add '목표 점수 {target}% 달성! 채택을 권장합니다.' to checkpoint display"
- "Set default option to 'Accept current result'"
- "User can still override and continue to next pass"
- "If current_score > target_score:"
- "Continue normally with standard default options"
override:
- "User can always override auto-stop recommendation"
- "Selecting 'Continue' at any checkpoint proceeds regardless of target"
- "Target score is advisory, not a hard gate"
Section-Selective Humanization (v3.1)
The pipeline supports transforming only specific sections while leaving others unchanged.
section_selective:
parameter: "sections"
type: "array of section names"
default: null # null = all sections (full manuscript)
valid_values:
- "abstract"
- "introduction"
- "methods"
- "results"
- "discussion"
- "conclusion"
behavior:
setup:
- "User specifies sections at STAGE 0 or at any Rich Checkpoint"
- "Example: sections: ['discussion', 'conclusion']"
during_g6_transform:
- "G6 receives sections parameter"
- "Only specified sections are transformed"
- "Non-selected sections pass through unchanged (verbatim copy)"
- "Change log only includes changes in selected sections"
during_g5_rescan:
- "G5 still scans ALL sections (for accurate composite score)"
- "Section-level scores reported for all sections"
- "Non-selected sections should show unchanged scores"
at_checkpoints:
- "Rich Checkpoint displays all sections with scores"
- "Non-selected sections marked as '(unchanged)' in patterns column"
- "User can modify section selection at any checkpoint"
use_cases:
- "Discussion has score 95% but Methods is clean at 25% → only humanize Discussion"
- "Abstract needs aggressive treatment but Results should stay conservative"
- "Re-run only on sections that still have remaining patterns after a pass"
Error Handling
Agent Failure
on_agent_failure:
g5_failure: "Retry once. If still fails, present partial results to user."
g6_failure: "Retry once. If still fails, ask user whether to continue with partial transformation."
f5_failure: "Continue pipeline. F5 is verification, not blocking."
Revert Protocol
At any checkpoint, user can select "Revert". Action:
Discard current pass output
Use previous pass output (or original if reverting Pass 1)
Present previous checkpoint again for re-decision
Resume Protocol
If session interrupted mid-pipeline:
Check diverga_checkpoint_status() for last completed checkpoint
Check diverga_project_update() for pipeline state
Present resume options to user via AskUserQuestion
NEVER run G6 in parallel with G5 or F5 within a single pass.G5 rescan and F5 verify MUST run in parallel after each G6 transform (both are read-only on the same output).
This saves latency on every pass without any risk to data integrity.
Relationship to Existing Skills
Existing Skill
Relationship
Conflict?
/diverga:g5
This skill CALLS g5 as a sub-step
No — g5 is a component
/diverga:g6
This skill CALLS g6 as a sub-step
No — g6 is a component
/diverga:f5
This skill CALLS f5 as a sub-step
No — f5 is a component
/diverga:orchestrator
Independent workflow
No — different trigger patterns
When to use which:
/diverga:humanize → Full multi-pass pipeline (recommended for manuscripts)
/diverga:g6 → Single-agent one-shot transformation (quick fixes, small sections)
/diverga:g5 → Standalone audit without transformation
