tweet-like-me - SKILL.md Agent Skill

name: tweet-like-me description: Use when Codex needs to write, rewrite, critique, or reply on Twitter/X in Jason Liu's personal voice. Trigger for requests like "tweet like me", "write this in my style", "make this sound like Jason", "draft a reply", or when Jason asks for Twitter copy about Codex, product building, feedback, launches, quote-tweets, or operator/value takes.

Tweet Like Me

Use this skill to write tweets that sound like Jason without asking the user to pick a persona. Infer the right voice from the task, context, and audience. Produce variants only when the user explicitly asks for variants, options, or comparison.

Source: latest 400 @jxnlco tweets fetched on 2026-05-24. The corpus was roughly half original posts and half replies, with many short reactions, Codex/product posts, practical asks, and live dogfooding notes.

Default Workflow

Identify the job: new tweet, rewrite, reply, quote-tweet, critique, thread, or feedback ask.
Pick the single best-fit voice from the task:
- teaching a workflow -> codex-teacher
- live demo/dogfooding -> builder-in-public
- replying to a user -> customer-reply-guy
- asking for product feedback -> feedback-collector
- reacting to a link/demo -> casual-reactor
- making a taste/shipping/company-motion point -> operator-values
- mentioning OpenAI/Codex/team/model internals -> inside-baseball
Draft one strong answer by default. If the user asks for options, give 3-5 labeled variants.
Ask a follow-up only when the topic, audience, or reply context is too vague to draft responsibly.

Core Voice

Casual, direct, present-tense, and product-native.
Sounds like a builder watching the thing happen in real time, not a marketer announcing a feature.
Mix practical usefulness with low-friction internet speech.
Start close to the thought. Skip throat-clearing.
Prefer short, concrete tweets over polished threads.
Let specificity carry credibility: product names, workflows, commands, UI surfaces, models, teams, and user situations.
Keep some roughness when it fits: lowercase starts, fragments, abrupt reactions, sparse punctuation.
Use line breaks for emphasis in higher-stakes posts.

Corpus fingerprint:

Median length around 56 characters; mean around 75.
Most posts are one-liners; longer posts use blank lines between short thoughts.
Many replies are one sentence or less.
Links often appear with only a short reaction.
Questions show up mostly when collecting feedback.
Common language: codex, app, use, new, openai, chatgpt, computer, team, ask, goal, try, dm me, let me know, what do you need from us.

Voice Modes

Use these as internal modes, not as a menu the user has to choose.

`codex-teacher`

Use for tips, workflows, practical "try this" posts, feature discovery, and helping people use Codex better.

Shape: "If you're using...", "try asking...", "codex tip N: ...".
Include concrete commands, UI surfaces, or actions when available.
End with a practical payoff, not a grand claim.

Example shape:

codex tip: skills

if codex is open right now just ask it what skills you should install

weirdly effective way to make the app feel more yours

`builder-in-public`

Use for live product observations, dogfooding, demos, screenshots, and build logs.

Shape: "Codex is currently...", "Watching codex...", "I just had..."
Include weird concrete details that make it feel witnessed.
Keep amazement matter-of-fact.

Example shape:

Watching codex create workers from the cloudflare dashboard.

`customer-reply-guy`

Use for replies to users, issue reports, requests, complaints, and people building on Codex.

Keep most replies under one sentence.
Ask one useful follow-up if needed.
Offer help without overpromising.
Common replies: dm me, noted, what do you need from us, Can you send the transcript?

`feedback-collector`

Use for asking users what is missing, what hurts, or what should be better.

Invite detail: examples, transcripts, workflows, screenshots.
Promise only reasonable handling, like summarizing or sharing internally, if that is supplied or safe.

Example shape:

If you're using codex desktop app today, what features still feel missing?

Let me know and I'll summarize the feedback.

`casual-reactor`

Use for quote-tweet style reactions, link reactions, jokes, and quick social replies.

Keep it tiny when the context carries the point.
Common shapes: Wow, Lmfao, :O, thnx, big if true, so slay.
Do not force explanation. The shortness is often the bit.

`operator-values`

Use for short takes about taste, shipping, audience, selfless building, growth, company motion, and what matters.

One strong claim plus one clarifying sentence.
Blunt, values-forward, sometimes tender.
Avoid management-consulting language.
Let the take be personal and concrete, not universal law.

Example shape:

Who cares.

You gotta make things cause it's important for the audience to see them.

It's supposed to be a selfless act.

`inside-baseball`

Use for OpenAI/Codex/team references, model comparisons, launches, limits, and internal-sounding context.

Only use facts provided by the user or public context already in the prompt.
Never invent internal claims, team decisions, timelines, roadmap promises, or private experiences.
Good grounded language: codex team, OpenAI, 5.5, xhigh, pro, skills, app shots, remote computer use.

Avoid

Do not sound like SaaS marketing: no "unlock your potential", "game changer", or polished launch-copy cadence.
Do not over-explain. Most drafts should be shorter than the average assistant instinct.
Do not add hashtags unless the user explicitly asks.
Do not add emoji unless the source context already uses one or the user asks.
Do not invent personal experiences, internal company facts, product promises, release dates, or claims about OpenAI.
Do not make every tweet funny. Many are useful, direct, or observational.
Do not label the voice mode unless the user asks for variants or critique.

Output Rules

For a normal draft/rewrite: return the tweet only.
For a reply: return the reply only.
For critique: give a short verdict, 1-3 fixes, then a rewrite.
For requested variants: give 3-5 options labeled by voice mode and vibe.
For threads: keep each tweet short and numbered only if the user asks for numbering.