brewcode%3atext-human（brewcode 3atext human 搜索）

Text Humanizer

Detect and remove AI-generated artifacts from code and documentation. Process commits, files, or folders with parallel sub-agents.

Language References

Multi-language projects: Load ALL relevant references.

Argument

First token = scope (required):

Everything after first token = custom prompt (optional). Free-form text, no quotes needed. Overrides or extends default humanization rules for this run. Passed to every sub-agent prompt.

Custom Prompt Handling

1. Parse: split args into scope (first token) and customPrompt (rest)
2. If customPrompt is present, prepend it to every sub-agent Task prompt as: CUSTOM INSTRUCTIONS (highest priority, override defaults): <customPrompt>
3. Custom prompt wins over default rules on conflict

Architecture

Orchestrator -> Detect Language -> Load Reference -> Analyze -> Classify -> Split (3-10 blocks) -> Parallel Task agents -> Aggregate

Phase 1: Scope Analysis

Commit Mode

Process all text files from commit. No extension filtering: git diff --name-only <hash>^..<hash>

File Inclusion Rules

Path Mode

find <path> -type f \( -name "*.java" -o -name "*.py" -o -name "*.ts" -o -name "*.js" -o -name "*.md" \) | grep -v -E "(target/|node_modules/|\.git/|build/|dist/|__pycache__/)"

Block Count

Single file: process directly without Task delegation.

Phase 2: File Classification

Haiku (Simple)

Sonnet (Complex)

Note: See language-specific references for detailed classification rules.

Phase 3: Block Formation

Group files by type, complexity (avoid mixed haiku/sonnet), line count balance. Keep related files together (same package/directory).

Data files block: YAML, JSON, CSV with comments → haiku for unicode fixes

Phase 4: Parallel Execution

Launch all Task calls in single message for true parallelism.

Task(subagent_type="developer", model="haiku", prompt="> **Context:** BC_PLUGIN_ROOT is available in your context (injected by pre-task.mjs hook).\n\n[CUSTOM_INSTRUCTIONS_IF_ANY]\nBlock 1: [files] [rules] Return JSON")
Task(subagent_type="developer", model="sonnet", prompt="> **Context:** BC_PLUGIN_ROOT is available in your context (injected by pre-task.mjs hook).\n\n[CUSTOM_INSTRUCTIONS_IF_ANY]\nBlock 2: [files] [rules] Return JSON")

If custom prompt was provided, prepend to EVERY sub-agent prompt (after the Context line):

> **Context:** BC_PLUGIN_ROOT is available in your context (injected by pre-task.mjs hook).

CUSTOM INSTRUCTIONS (highest priority, override defaults):
<user's custom prompt text>
---

<output_format> {"files_processed": N, "changes": [{"file": "path", "removed_comments": N, "fixed_unicode": N, "simplified_docs": N}]} </output_format>

Phase 5: Aggregation

Collect JSON results from all agents → merge statistics → generate unified report.

<humanization_rules>

AI Artifacts (Universal)

Real vs Fake Issue References

Keep project-specific ticket patterns (INTELDEV-XXXXX, JIRA-XXXXX, GH-XXX). Remove generic AI-invented patterns (BUG-001, FIX-123, ISSUE-42).

Documentation Cleanup (General Principles)

Key Rule: Never convert block docs to inline comments. Delete unnecessary docs entirely; for useful descriptions with trivial params, strip params and keep description.

Comments (Universal)

Keep WHY, remove WHAT.

Formatting (Universal)

</humanization_rules>

File Type Rules

YAML/Data Files

Check for unicode in comments. Section comments are valuable.

SQL/XML Comments

Structural comments are valuable. Analyze neighboring files first.

Custom Prompt Mode

Custom prompt overrides or extends default rules. Use for:

• Restricting scope: "only remove AI artifacts, don't touch docs"
• Adding rules: "also remove all @author tags"
• Excluding files: "skip test files"
• Style reference: "use style from src/main/Service.java as reference"

Output Format

## Humanization Report

### Execution Summary
| Metric | Value |
|--------|-------|
| Total files | N |
| Blocks | M |
| Haiku/Sonnet | X/Y |

### Block Results
[Per-block tables with file metrics]

### Totals
| Metric | Count |
|--------|-------|
| Files processed | N |
| Comments removed | X |
| Docs simplified | Y |
| Unicode fixed | Z |

Best Practices

• Load language reference first
• Process all commit files (include yaml/data files)
• Check yaml for unicode (comments may have arrows)
• Preserve valuable comments (they add context)
• Skip binaries (avoid corruption)
• Honor style reference if provided
• Keep public API docs
• Group files by complexity
• Launch agents in parallel

Error Handling

Examples

/text-human 3be67487                                              # Commit - all files
/text-human src/main/java/MyService.java                          # Single file
/text-human src/main/java/services/                               # Folder
/text-human 3be67487 don't touch docs on public records           # Commit + custom prompt
/text-human src/ only remove AI artifacts and fix unicode         # Path + custom prompt
/text-human 3be67487 also remove all @author tags                 # Commit + extra rule