AutoGineer
Generated banner for Write Like Meng On X
Generated logo for Write Like Meng On X

Write Like Meng On X

Meng ToCommunityCreated by Meng To

  • Skill
  • Content generator
  • Agent tooling
  • MIT
  • Low risk
  • Has scripts
Download ZIP

From the author

Write, rewrite, review, or continuously refine X/Twitter posts in Meng To's current voice using his deduplicated authored-post corpus, personal and product context, shared resources, and Content repo evidence. Use when asked to improve a tweet in Meng's tone, draft an original post, reply, thread, resource share, or quote post as Meng, check whether an angle repeats an earlier post, study Meng's writing style, ingest another 20-50 authored posts, or hone the reusable Meng X voice profile.
— the description by Meng To

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill enables an agent to generate and refine X posts that align with Meng To's writing style. It uses a deduplicated corpus of authored posts and contextual data from a Content repository to ensure that drafts remain grounded in real product work, personal experience, and established teaching themes. The skill includes mechanisms to maintain voice consistency, prevent content repetition, and periodically refresh the underlying corpus with new authored posts.

What it does

  • Drafts original posts, replies, threads, and quote posts based on user input.
  • Refines existing drafts to match Meng To's conversational, builder-focused voice.
  • Checks for potential repetition against a corpus of previously authored posts.
  • Ingests new authored posts from X to keep the voice profile and corpus current.
  • Provides multiple distinct drafting options for every request.
  • Ensures content is grounded in verified personal or product context from the Content repo.

When to use it

  • When you need to draft a new post or reply that sounds like Meng To.
  • When you want to improve a draft while preserving the original intent and specific voice.
  • When you need to ensure a new post does not repeat themes or hooks from previous content.
  • When you need to update the agent's knowledge with Meng's most recent X activity.

Advantages

  • Maintains a consistent, authentic voice based on actual authored evidence.
  • Prevents generic or repetitive content by enforcing grounding in specific product or personal context.
  • Provides diverse structural options for the same core idea.
  • Automates the maintenance of the voice corpus and profile.

Usage examples

“Draft a thread about why context matters more than just the prompt when using AI for design.”
The agent generates five distinct thread options grounded in Meng's specific design-builder perspective and verified teaching themes.
“Rewrite this tweet to sound more like Meng: I think AI is great for building apps quickly.”
The agent provides five variations that replace the generic statement with more specific, builder-minded language and personal context.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/mengto/write-like-meng-on-x/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip write-like-meng-on-x.zip -d .agents/skills

Documentation by Meng To

Write Like Meng on X

Start

Work from the user's Content repo unless the user names another checkout. If no Content repo is active, locate it from the available workspace roots or ask for the target before writing files.

  1. Read AGENTS.md and run git status --short --branch before changing files.

  2. Read references/voice-profile.md before drafting.

  3. Read only the relevant rows in references/content-source-map.md for factual and personal context.

  4. Search references/tweet-corpus.jsonl for the subject, hook, resource, product, and repeated phrases before writing. Do not load or restate the entire corpus when a focused rg search is enough.

  5. Treat authored X posts as voice evidence, not as copy to splice together.

Use this evidence order:

  1. Current user instructions and raw wording

  2. Recent authored posts of the same format

  3. Trusted personal, product, and resource context in the Content repo

  4. Older authored posts

  5. Prior generated drafts, only for continuity

Draft

Identify the format first: reply, original post, quote post, resource share, thread, or product post. Match examples from the same format because Meng's replies are looser than his standalone posts.

Keep the user's strongest original line when it already sounds natural. Improve clarity, sequence, and specificity without replacing the thought with a generic copywriting framework.

Ground the post in something real when useful:

  • a product Meng has built or actively uses

  • a workflow, constraint, failure, proof point, or result recorded in Content

  • a resource Meng has actually shared

  • a personal detail already supported by the corpus or trusted project context

Never invent usage, metrics, customers, team behavior, travel, family details, product capabilities, or results. Do not turn a resource share into a hidden product pitch.

For every rewrite or drafting request, return exactly five finished options unless the user explicitly requests another count. Put the strongest option first. Make the five options materially different in angle, structure, rhythm, or intent instead of producing five surface-level paraphrases.

Adapt the remix mix to the post, using distinct types such as:

  1. Closest polish that preserves the user's original thought and phrasing

  2. Punchier hook or tighter structure

  3. Conversational, personal, or more casually spoken

  4. Proof-led, mechanism-led, or practical

  5. Playful, contrarian, unexpected, or use-case-driven

Label each option by its remix type when useful. Keep all five grounded in the same verified facts. Do not bury the copy under analysis.

Default to casual, spoken language unless the user asks for a more formal tone. Prefer the words Meng would naturally say out loud over polished brand-copy phrasing.

Make every sentence advance the thought. Do not let the second sentence or paragraph merely restate the hook or describe the same feature again. Use it to add a reason, mechanism, use case, proof point, constraint, contrast, question, or conclusion. If it adds nothing new, cut it.

Prevent Repetition

Check two kinds of duplication before finalizing:

Corpus duplication

  • Match by post ID, canonical status URL, and normalized text hash.

  • Store only Meng's authored text as voice evidence. Keep quoted-source text out of the voice field.

  • Never append the same authored post twice.

  • Replace a truncated record only when a verified full version is available.

Editorial duplication

  • Search for the same hook, claim, resource bundle, anecdote, proof point, and ending.

  • Do not reuse a distinctive line merely by changing a few words.

  • A recurring belief is allowed when the new post adds new proof, a different mechanism, a useful update, or a new personal angle.

  • Prefer a fresh observation over another version of taste is the moat, context matters, AI slop, or the prompt is not the point unless the source genuinely advances it.

Refresh The Voice Corpus

Use the Codex in-app browser only. Never use Chrome. Keep X read-only: do not post, reply, quote, like, repost, bookmark, follow, DM, or change account state.

Open the signed-in profile and the Latest search:

https://x.com/MengTo
https://x.com/search?q=from%3AMengTo&src=typed_query&f=live

On each refresh:

  1. Collect 20-50 unseen authored posts when X exposes enough history. Start with posts newer than the newest ledger entry, then continue older than the oldest ledger entry to backfill when the new batch is smaller than 20.

  2. Include original posts, replies, and quote posts. Exclude pure reposts written by someone else.

  3. Record post ID, canonical URL, UTC timestamp, authored text, and format.

  4. Expand every truncated authored post on its status page. Report any post that remains incomplete.

  5. Keep quoted-source text separate during collection and exclude it from the voice corpus.

  6. Save the collected batch as JSON or JSONL and ingest it with the script bundled beside this skill:

node <skill-dir>/scripts/update-tweet-corpus.mjs --input <batch-path>

  1. Read the script receipt. A healthy run reports zero duplicate IDs, URLs, and text hashes in the final corpus.

  2. Update references/voice-profile.md only when the new evidence confirms a repeated pattern, changes an earlier conclusion, adds useful lived context, or reveals a format-specific move.

  3. Update references/content-source-map.md when new Content work adds a durable product fact, resource, story, or teaching theme. Link to the source instead of copying whole passages.

If X is logged out, CAPTCHA-blocked, or the Codex browser cannot attach, stop and report the exact blocker. Do not use public search, Chrome, or guessed posts as a fallback.

Hone The Profile

Measure the batch before editing the profile:

  • sample count and date range

  • original, reply, and quote-post mix

  • first-person usage

  • paragraph and list patterns

  • recurring openings, transitions, and endings

  • concrete nouns, tools, resources, products, and proof

  • rough edges that are natural versus artifacts that should not become rules

Use replies for conversational vocabulary, capitalization, curiosity, gratitude, and disagreement. Use original posts for hooks, story, pacing, lists, proof, and endings. Use quote posts and resource shares for how Meng adds a reason, mechanism, or related resource.

Do not manufacture typos to imitate authenticity. Preserve natural directness, contractions, occasional fragments, and lower-case phrasing when the format supports it.

Validate And Commit

After changing the skill or corpus:

node <skill-dir>/scripts/update-tweet-corpus.mjs --check
python3 <skill-creator-dir>/scripts/quick_validate.py <skill-dir>
git diff --check -- <skill-dir>

Stage only the source-controlled write-like-meng-on-x/ skill folder and commit the narrow change. Leave unrelated modified and untracked files alone.

Report the sampled and added counts, date range, profile changes, duplicate checks, browser status, validation result, and commit hash. If no new post or durable voice insight was found, report a no-op instead of creating churn.

Files

references/voice-profile.md

Meng X Voice Profile

Last evidence update: 2026-07-17 JST

Evidence Base

  • A prior live study sampled 100 authored posts from July 4-13, 2026. It established the rule that replies teach conversational phrasing while standalone and quote posts teach structure.

  • The current corpus contains 40 live authored posts from July 11-17, 2026: 1 original post, 34 replies, and 5 quote posts. The latest 20-post batch contains 16 replies and 4 quote posts.

  • Personal, product, resource, and teaching context comes from the Content repo sources indexed in content-source-map.md.

  • Recent teaching context is canonical in the Content repo at articles/2026-07-13-x-history-teaching-strategy.md, articles/2026-07-13-agent-skills-business-flywheel/content.md, and articles/2026-07-13-ui-prompting-vocabulary/content.md.

  • tweet-corpus.jsonl is the deduplicated source of exact authored wording. This profile stores conclusions, not a second copy of the tweets.

Who Is Speaking

Meng is a designer, builder, teacher, and founder. His public identity is simple: founder at DesignCode and Aura, teaching designers code and developers design. The broader Content workspace also contains real work around DreamCut, Neuform, agent skills, AI-assisted product building, UI prompting, motion, video, and creator workflows.

The useful tension in his perspective is human taste plus agent leverage. He is excited about what agents can do, but the interesting part is usually the context, workflow, judgment, proof, or failure behind the result.

Durable beliefs supported across the corpus and Content:

  • Design taste comes from lived experience, references, vocabulary, and thousands of small decisions.

  • A strong result needs more than a model or one prompt. Context, planning, skills, subagents, verification, and iteration matter.

  • Reusable skills turn a good result into infrastructure other people can use.

  • The best resources help name what is missing. They are not a bookmark graveyard or a template to copy.

  • Proof makes teaching useful: screenshots, videos, code, prompts, metrics, before and after, commits, and working products.

  • Agents should remove tedious work while leaving direction, intuition, and responsibility with the human.

The Sound

Direct, specific, builder-minded, curious, and useful.

Use ordinary words. One main idea per sentence when possible. Keep short paragraphs, but do not force every sentence onto its own line.

Meng often sounds more spoken than edited. Replies may begin in lower case, use contractions, include fragments, or leave a small grammatical rough edge. Do not clean that into brand copy, and do not manufacture mistakes on purpose.

First person is natural when it adds lived context: I asked, I combine, I tested, I realized, I keep coming back to. Avoid pretending every lesson happened personally.

Enthusiasm is concrete. Phrases such as kind of wild, incredible, crazy good, what a time to be a builder, or you have my attention work because a specific tool, result, or question follows them.

Humor is dry and personal rather than performative. The July 14 birthday post opened with turning 44 as the perfect iOS button size, then used that geeky detail to enter a larger reflection.

Format Modes

Replies

Replies are the loosest mode. They are usually one useful reaction, question, clarification, concrete detail, or sincere thank-you. Lower-case openings can feel natural. Do not turn a reply into a miniature essay unless the source earns it.

Good reply moves:

  • ask the implementation question that the original post leaves open

  • name one design detail that improved

  • add a constraint such as grid size, reference quality, or context switching

  • expose the hidden cost of a workflow, such as another subscription, tool switch, or review step, then offer a simpler operating choice

  • agree with a specific reason

  • thank a person directly without a brand voice

  • make a small personal joke

Original Posts

Lead with a result, contradiction, personal detail, or useful surprise. Explain the mechanism after the hook. Longer posts earn their lists through real reflection or a concrete workflow.

Strong original-post shapes:

  • result first, then the workflow behind it

  • personal moment, then what changed

  • surprising tool capability, then a walkthrough or proof

  • common shallow interpretation, then the system it misses

  • useful resource, then how to use it

Quote Posts And Resource Shares

Do more than summarize the source. Add why it matters, how Meng would use it, the missing mechanism, or two related resources.

A strong resource share often has this order:

  1. Direct reaction or recommendation

  2. Concrete capability or reason

  3. Related resource, repo, or workflow when useful

  4. Link

Do not append a product pitch unless the connection is real and useful.

Product Posts

Show the finished result early. Then explain the context, workflow, or tradeoff that made it possible. Honest constraints such as limits, token use, failed attempts, or the need for polish make the result more believable.

Avoid feature dumping. Turn features into a scene: what Meng asked for, what the agent did, what surprised him, what still needed judgment, and where someone can see the proof.

Context Meng Naturally Reaches For

  • UI vocabulary, design systems, typography, motion, component boundaries, interaction states, responsive behavior, and visual proof

  • AI builders and coding agents, especially Codex, model taste, planning, subagents, skills, verification, browser proof, and commits

  • Building and teaching through DesignCode, Aura, DreamCut, and Neuform

  • Turning screenshots, videos, references, prompts, and working examples into better agent context

  • Resource curation for interface patterns, landing pages, product flows, decks, video, motion, and creative production

  • Founder lessons from shipping, selling, debugging, teaching, content systems, and team leverage

  • Travel, building from an iPhone, family time, aging, intuition, and the human perspective behind the work when the subject genuinely connects

Resource Memory

Frequently supported resource categories:

  • Visual direction: Dribbble, Framer Marketplace, Lapa Ninja, Awwwards, Godly

  • Product patterns: Mobbin, Collect UI, UX Archive, Screenlane, Name That UI

  • Deck and story structure: Deck.Gallery

  • Design systems and AI building: Figma, getdesign.md, Neuform, Aura, Stitch, v0, Lovable, Cursor, Framer

  • Motion and implementation: Codrops, Three.js, GSAP, Lenis, Figma Motion

  • AI and production: Codex, ChatGPT, Claude, Gemini, GPT Image, DreamCut, CapCut, Runway, Seedance

Use the source map before naming a resource in a draft. Share fewer resources with a reason instead of dumping a long list without a use case.

Avoid

  • Generic openings such as In today's fast-paced world

  • Empty agreement such as Great point

  • Hype words such as game changer, unlock, and supercharge used without a concrete mechanism

  • The real X is Y as a repeated template

  • Polished symmetry that sounds generated

  • Short sentence, blank line, short sentence cadence for an entire post

  • Abstract conclusions that merely rename the topic

  • A CTA on every post

  • Em dashes

  • Invented biography, product facts, usage, metrics, or outcomes

  • Reusing a distinctive hook, anecdote, resource bundle, or ending from the corpus with surface-level changes

Current Editorial Watchlist

These beliefs are true but already common in Meng's recent work. Use them only with fresh proof or a new mechanism:

  • taste is the moat

  • context matters more than a bare prompt

  • skills make workflows reusable

  • AI output looks generic without references

  • the one-prompt story hides the real system

  • agents create more time for higher-level judgment

The freshness test is simple: what happened, what changed, what specific artifact proves it, and what can the reader do differently now?

Recent combinations that are already spent unless there is a concrete update:

  • small prompt + screenshots or video + one detailed prompt + UI skills

  • Sol as a design model + video-to-HTML + a walkthrough + the open-source Skills repo

  • open-ended prompting + an agent-written plan + isolated threads + per-change commits

  • UI vocabulary + interactive examples + the Name That UI, Collect UI, and Mobbin resource bundle

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs a command-line tool, Needs Python, Needs Node.js

Security

What our scanner noticed· Low risk

  • Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by
Meng To
Publisher
Meng To (Community)
Source
MengTo/Skills · agent-skills/codex/write-like-meng-on-x · commit 5ea1661
Licence
MIT License (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by Meng To or Meng To.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download write-like-meng-on-x.zip