humanizer

name: humanizer description: | Strip AI-tells from text and rewrite it in a personal voice profile so it reads as natural human writing before sending. Four-stage pipeline: deterministic pre-pass (free) + LLM rewrite via tool-use + platform post-processing. Voice profiles are JSON files with real writing samples that teach the LLM to match the person's actual cadence.

Triggers on: "humanize this", "make this sound less AI", "make this less robotic", "rewrite in my voice", "paste-ready", "clean this up so it sounds like me", "rewrite for LinkedIn", "make this sound human", "remove the AI from this", "strip the AI", "make this sound like Debanjan", "humanize before I send this".

Also triggers when the user pastes a block of AI-generated text and says "fix this" or "send this to [person/platform]" — it almost certainly needs humanizing first. user_invocable: true

Humanizer

When invoked, run the humanizer script and return the cleaned text to the user.

Step 1 — Capture the text

The user will either:

• Paste the text inline in chat ("humanize this: [text]")
• Mention a file path ("humanize the draft in draft.md")
• Have already sent the AI-generated text as context

If the text is not obvious, ask: "Paste the text you want humanized."

Step 2 — Single text or batch?

Single text (default): paste or pipe one block of text and run the single-input commands in Step 4.

Batch mode: use --batch when you have more than ~5 texts to humanize in one session, or when you need structured CSV output (text_id, original, humanized, tier, cost, error). Batch fans out across parallel workers and isolates per-row failures — a single bad row does not abort the run.

Batch mode

Trigger: user has a list of AI-drafted texts (cold emails, LinkedIn posts, tweet thread, etc.) and wants all of them humanized in one shot.

Input CSV format:

text_id,text_to_humanize,voice_name
row1,"Certainly! I'd be happy to delve into this...",debanjan
row2,"Absolutely! Let's leverage synergies...",debanjan
row3,"It's worth noting that this is a game-changer...",

• text_id — any unique string identifier
• text_to_humanize — the AI text to rewrite
• voice_name — optional; falls back to --voice CLI default if blank

Run:

py execution/content/humanizer.py --batch input.csv --max-workers 4 --tier default --out output.csv

Output CSV columns: text_id, original, humanized, tier, cost_usd, error

Failed rows have a non-empty error column; all other rows complete normally. After the run, a summary line is printed to stdout:

BATCH SUMMARY: 10 total, 9 succeeded, 1 failed, $0.00312 total cost

Batch + dry-run (no API calls, no cost):

py execution/content/humanizer.py --batch input.csv --dry-run --out preview.csv

Batch flags:

All other single-input flags (--voice, --platform, --tier, --max-length, --dry-run, etc.) apply to every row in the batch. Per-row voice_name column overrides --voice for that row.

Step 2b — Pick the voice

Default to --voice debanjan unless the user specifies otherwise.

If the user says "in my voice" or "sound like me" and they are Debanjan, use debanjan.

To add a new voice later: copy execution/content/voices/_template.json to voices/{name}.json and fill in real writing samples. See directives/content/humanizer.md for the "Adding a new voice" section.

Step 3 — Pick the platform

Ask or infer from context:

Step 4 — Run with --show-diff first (recommended)

Always run with --show-diff first on the first call so the user can see what changed:

py execution/content/humanizer.py --text "Certainly! I'd be happy to help you delve into..." --voice debanjan --platform linkedin --show-diff

The before/after diff prints to stderr. The humanized text prints to stdout.

For a quick cost check before the real call:

py execution/content/humanizer.py --text "..." --dry-run --show-diff

Step 5 — Hand back the cleaned text

After the script finishes, copy the stdout output and present it to the user. If --show-diff was used, also mention the key changes (e.g., "Stripped 'Certainly!', replaced 'delve', shortened 3 sentences").

Common commands

# Basic humanize (stdin)
echo "Certainly! I'd be happy to..." | py execution/content/humanizer.py

# Direct text, show diff
py execution/content/humanizer.py --text "..." --show-diff

# LinkedIn-specific cleanup
py execution/content/humanizer.py --text "..." --platform linkedin --show-diff

# Tweet (hard 280-char cap)
py execution/content/humanizer.py --text "..." --platform tweet

# Dry run (no LLM, see pre-pass + cost estimate)
py execution/content/humanizer.py --text "..." --dry-run

# Premium tier for nuanced rewrite
py execution/content/humanizer.py --text "..." --tier premium --show-diff

# Free Gemini path
py execution/content/humanizer.py --text "..." --tier gemini --show-diff

# From file
py execution/content/humanizer.py --file draft.md --platform email --show-diff

When NOT to use this skill

• The user wants to write something from scratch (use Claude to draft, then humanize)
• The text is already short and direct (e.g., a single sentence reply) -- skip humanizing, it adds latency for minimal gain
• The user explicitly says they like the AI's phrasing and just wants to send it as-is

Choosing single vs batch