humanizer

---

name: humanizer description: Rewrites text or comments to remove AI writing tics (delve/leverage/robust, tricolons, em-dash overuse, empty intros and closers, cheerleader tone) and make them sound like a human engineer wrote them. Use after generating any prose Claude wrote: commit messages, code comments, PR descriptions, READMEs, story.md / design.md / spec docs, release notes, Slack/email drafts — or whenever the user says something "sounds like AI", "is too markety", "is too flowery", or asks you to tighten/dehydrate the wording. Preserves meaning exactly; only the style changes. argument-hint: [file path, range, or pasted text] disable-model-invocation: false allowed-tools: Read, Edit, Grep, Glob

Humanizer

Rewrite the target text so it reads like a human engineer wrote it in a hurry, not like a model generating plausible prose. Preserve meaning exactly — do not add information, soften claims, or hedge things that were stated plainly. Only the style changes.

Hard constraints

• Only touch prose, doc comments, and markdown. Never edit code, identifiers, type signatures, imports, config keys, JSON/YAML values, regex patterns, SQL, shell commands, or anything inside fenced code blocks. If a sentence quotes an identifier or value, rewrite the surrounding prose but leave the quoted text byte-for-byte intact.
• Preserve technical claims verbatim. If the text says "deadlocks under contention" or "returns 401 on an expired token", do not soften, generalize, or hedge it. The tells you remove are stylistic; the substance stays.
• Match the surrounding voice. If the file or thread already has a clear voice (the user's existing commit messages, a colleague's code review, a project's README tone), edit toward that voice — don't impose a generic "humanized" style.

Tells to remove

Vocabulary

Replace or delete on sight:

• "delve", "dive into", "explore", "navigate", "embark", "journey"
• "leverage" → use, "utilize" → use, "facilitate" → help / let, "employ" → use
• "robust", "seamless", "seamlessly", "effortless", "elegant", "powerful", "comprehensive", "rich", "vibrant", "cutting-edge", "state-of-the-art", "best-in-class"
• "in the realm of", "in the world of", "in today's fast-paced…", "it is important to note that"
• "furthermore", "moreover", "additionally" at the start of a sentence — usually just delete
• "ensure that" → "make sure" or just drop "that"
• "a myriad of", "a plethora of" → "many" or a number
• "crucial", "vital", "essential", "pivotal" → "important" or delete
• "unleash", "unlock", "supercharge", "empower" — delete
• Empty intros: "Certainly!", "Of course!", "Great question!", "I'd be happy to" — delete
• Empty closers: "I hope this helps!", "Let me know if you have any questions!", "Happy coding!" — delete

Structure

• Tricolons / rule of three. "fast, reliable, and scalable" — pick the one that's actually true and drop the others.
• Parallel "not just X but Y" constructions. "It's not just a library, it's a way of thinking." Cut.
• Bullet lists that should be sentences. If three bullets each have one phrase, write one sentence.
• Section headers for short content. Don't put a ## Overview above two sentences.
• Em-dash overuse for dramatic pauses — one or two per page is fine, ten is a tell. Prefer commas, periods, or parentheses.
• Symmetric sentence pairs. "X is Y. But Z is W." — vary the rhythm.
• Trailing summary sentences that restate the paragraph. Just stop at the last real point.

Tone

• Drop the cheerleader voice. No "exciting", "amazing", "fantastic", "powerful new way".
• Drop the academic hedge. No "it could be argued that", "one might consider", "it is generally accepted".
• Don't apologize for length, simplicity, or limitations unless the user asked.
• Don't announce what you're about to do ("Let's break this down…") — just do it.

What to keep

• Technical precision. Don't dumb down terminology or replace specific terms with vague ones.
• Direct statements. If the original says "this will deadlock", keep it — don't soften to "this may potentially lead to a deadlock scenario".
• Code, identifiers, paths, links, numbers, error messages — never paraphrase these.
• The author's voice if it's already present (e.g. existing commit messages, code review comments).

Don't over-humanize

The goal is "tired senior engineer wrote this", not "rude bot with a 50-word vocabulary". Watch for these failure modes and back off when you see yourself doing them:

• Choppy fragments. Cutting until every sentence is three words. Some flow is fine — humans use connectors. Keep "because", "so", "when", "if", "since" when they carry real meaning.
• Stripping structure that was earning its keep. A real list of four distinct items should stay a list. A heading that introduces a genuinely separate section should stay a heading. Cut structure that's decorative, not structure that's load-bearing.
• One-pass-too-many. If you already cut 30%, stop. Re-reading three times and trimming each pass turns the text into a telegram. When in doubt, leave it.
• Replacing one tic with another. Don't trade "leverage" for "harness", "robust" for "solid", "delve" for "dig into". If you can't say it plainly, leave the original word.
• Removing necessary hedges. "May fail under load" is a calibrated claim, not a cheerleader hedge. Strip "it could be argued that…" but keep "may", "can", "sometimes" when they're load-bearing.

Workflow

1. Read the target. If it's a file or range, identify exactly which prose is in scope (skip code blocks and identifiers).
2. For each sentence, ask: would a tired senior engineer write this, or does it sound like marketing? Rewrite the marketing ones, leave the engineering ones.
3. Prefer deletion over rewriting. Most AI prose gets better by cutting 30%.
4. Read the result aloud in your head. If it still sounds like a brochure, cut more.
5. Apply edits with minimal diff — do not reflow unrelated lines.
6. Report a short summary: lines changed, words removed, and any phrasing you were unsure about.

Examples

Before	After
"This robust solution leverages a comprehensive set of tools to seamlessly facilitate the deployment process."	"Deploys with the existing tools."
"It is important to note that the function will return nil if the input is empty."	"Returns nil on empty input."
"Let's dive into how this works."	(delete)
"We employ a sophisticated caching strategy to unlock significant performance gains."	"Caches results; ~3x faster."
"Furthermore, the system is designed to be highly scalable and incredibly fast."	"Scales horizontally."