How to Build an AI Agent That Writes Tweets in Your Voice

Prerequisites

Before you start, you need three things: a Claude account (free tier works, but Pro gives you longer conversations and Projects), ten to fifteen of your best-performing tweets that represent your actual voice, and a clear picture of who you are writing for on Twitter/X. That last one matters more than people think. X rewards sharp, opinionated, compressed writing. An agent that writes for a generic audience produces generic tweets that get ignored.

Step 1: Define Your Voice Profile

Most people skip this step and go straight to prompting. That is why most AI-generated tweets read like corporate press releases squeezed into 280 characters. Your voice profile is the foundation. Without it, you are building on sand.

Open a new Claude conversation and paste ten to fifteen of your best tweets. Include a mix: standalone tweets, thread openers, quote tweets, replies that got traction. Then use the following prompt:

I'm going to share several tweets I've written. Analyze them
and create a detailed Voice Profile document that captures:

1. TONE: Am I sharp, casual, deadpan, provocative, informative?
   Give me a spectrum, not a single word.

2. STRUCTURE: How do I open? Do I use fragments or full sentences?
   Do I use line breaks within tweets? How do I close — punchline,
   question, trailing thought?

3. VOCABULARY: What words and phrases do I gravitate toward?
   What do I avoid? Do I use internet slang, abbreviations,
   or write formally?

4. SIGNATURE MOVES: What makes my tweets recognizably mine?
   (e.g., opening with a contrarian claim, dry one-liners,
   numbered observations, ratio-bait takes)

5. TOPICS & ANGLES: What themes recur? Do I approach topics
   through personal experience, hot takes, humor, data, or
   observation?

6. THREAD STYLE: When I write threads, how do I structure them?
   How long are individual tweets within threads? How do I
   hook people into reading the next tweet?

7. ANTI-PATTERNS: What does generic AI-written Twitter content
   look like, and how does MY writing differ from that?

Format this as a structured reference document I can reuse.

Here are my tweets:

[PASTE YOUR TWEETS HERE]

Claude will return a structured document. Read it carefully. Does it actually sound like you? If something feels off, push back. Say: “That’s not quite right. I don’t write in complete sentences most of the time. More fragments. More punch. Think less professional copywriter, more someone firing off thoughts at midnight.” This refinement is part of the process.

Save the final voice profile. You will use it in every prompt going forward.

Step 2: Build the System Prompt

The system prompt is the instruction set that turns Claude from a general-purpose AI into your personal Twitter writing agent. Think of it as the agent’s job description. Here is a production-ready template:

You are a Twitter/X content agent for [YOUR NAME], a [YOUR ROLE]
who tweets about [YOUR TOPICS].

## Voice & Style
[PASTE YOUR VOICE PROFILE FROM STEP 1 HERE]

## Audience
Primary: [e.g., Tech builders and operators, 25-45]
Secondary: [e.g., Founders interested in AI and automation]
What they engage with: [e.g., sharp observations, real results, contrarian takes]
What they scroll past: [e.g., motivational platitudes, engagement bait, thread-bro energy]

## Platform Rules
- Hard limit: 280 characters per tweet (this is non-negotiable)
- Threads: each tweet must stand alone AND hook into the next
- No hashtags unless I specifically ask for them (they look desperate on X)
- No emojis at the start of lines (looks like a LinkedIn post)
- Never use "🧵" or "thread:" — just start with a strong opening
- Brevity is a feature, not a limitation. Say less to mean more.

## Content Rules
- Never use these words/phrases: [e.g., "game-changer", "just dropped",
  "let that sink in", "here's the thing", "hot take:"]
- Never write in a way that sounds like a growth hacker or hustle bro
- Contrarian is fine. Needlessly provocative for engagement is not.
- If the tweet doesn't have a clear point, don't write it
- Match the energy of the format: standalone tweets are punchy,
  threads are structured, quote tweets add a sharp angle

## Output Format
For each tweet, provide:
1. The tweet text with exact character count
2. Format label: standalone / thread opener / quote tweet / reply
3. One alternative version with a different angle
4. If thread: all tweets numbered with character counts

Where to put this prompt: If you are using Claude Pro, create a Project and add this as the Project instructions. This means every conversation inside that Project automatically uses your agent configuration. If you are on the free tier, paste it at the start of each new conversation, or keep it in a text file for quick access.

Step 3: Structure Your Input

The power of a well-built agent is that your input can be minimal. You should not need to write a brief every time you want a tweet. Here is the input template that makes this work:

Write a tweet based on this:

TOPIC: [One sentence describing what the tweet is about]

ANGLE: [Hot take, observation, personal experience, dry humor,
contrarian view, or signal boost]

TRIGGER: [What made you think about this? Something you saw,
built, read, or experienced?]

FORMAT: [Standalone tweet, thread, quote tweet draft, or poll]

Here is a real example of this template in action:

TOPIC: Everyone is building AI agents but nobody is building
the feedback loops that make them actually useful

ANGLE: Contrarian observation from running agents in production

TRIGGER: Saw three "I built an AI agent" threads today where
none of them mentioned evaluation or iteration

FORMAT: Standalone tweet, keep it under 200 characters

Notice how little effort this requires from you. Four lines. The agent handles the rest: the phrasing, the compression, the voice, the punch. Your job is to have the observation. The agent’s job is to articulate it in your voice within 280 characters.

Step 4: Create a Feedback Loop

This is where most AI content workflows fail. People generate a tweet, post it, and never close the loop. The agent never learns what landed. Here is how to fix that.

After Every Tweet That Performs

You do not need to debrief every single tweet. But when one hits — or completely misses — come back to your agent with a quick debrief:

Tweet performance update:

TWEET: [Paste the tweet you published, including any edits]

EDITS I MADE: [What did you change before posting and why?]

PERFORMANCE: [Impressions, likes, retweets, replies, bookmarks,
profile visits — whatever you track]

WHAT WORKED: [Did a specific phrase get quoted? Did people reply
with strong reactions? Did it start a conversation?]

WHAT FELL FLAT: [Low impressions? Wrong audience? Felt forced?
Got ratio'd for the wrong reasons?]

Update your understanding of my voice and what resonates on
this platform. Tell me what you'll adjust going forward.

Claude will internalize this feedback within the conversation (or Project) and adjust. Over time, the agent develops a feel for what works with your specific audience — not just Twitter in general, but the corner of Twitter you inhabit.

Monthly Voice Recalibration

Once a month, run a recalibration prompt:

Review all the feedback I've given you over the past month.
Update my Voice Profile with any new patterns you've noticed:

- What types of tweets consistently get the most engagement?
- What opening patterns hook people best?
- What topics does my audience respond to most?
- What formats work best (standalone vs. threads vs. quote tweets)?
- What should we stop doing?
- What should we experiment with next?

Give me an updated Voice Profile document I can use going forward.

This updated voice profile replaces the one in your system prompt. Each month, your agent gets sharper.

Step 5: Advanced Patterns

Once the basic loop is running, here are three ways to push it further.

Thread Builder

Threads are Twitter’s long-form format. They need different architecture than standalone tweets — each tweet must hook the reader into the next while standing on its own if someone lands mid-thread:

I want to create a Twitter thread on [TOPIC].
Target length: [5-10] tweets.

Structure it like this:
Tweet 1: Bold claim or observation that stops the scroll
Tweet 2: Context — why this matters right now
Tweets 3-N: The substance — one idea per tweet, each ending
with a reason to read the next one
Final tweet: The sharpest takeaway, something quotable

Rules:
- Every tweet must be under 280 characters
- No tweet should start with a number (e.g., "1/")
- No "🧵" or "thread" labels
- Each tweet should be strong enough to screenshot on its own
- End the thread with something people want to retweet

Draft all tweets with character counts.

Repurposing Engine

Turn longer content into Twitter-native formats:

I just published this blog post / gave this talk / wrote this
newsletter. Extract 3-5 tweet ideas from it.

For each one, tell me:
- The core insight worth sharing
- The best format (standalone, thread opener, or hot take)
- A draft in my voice, under 280 characters
- One alternative version with a different angle

[PASTE SOURCE CONTENT]

Reply and Quote Tweet Drafts

Use your agent to draft replies and quote tweets that add value instead of just noise:

Someone posted this tweet: "[TWEET]"

Draft a reply/quote tweet that:
- Adds a new angle or insight (not just agreement)
- Stays in my voice
- Is under 280 characters
- Doesn't come across as trying to hijack their engagement

Give me two options: one sharp, one thoughtful.

The Complete Workflow at a Glance

Step	Action	Frequency
1	Build Voice Profile	Once (update monthly)
2	Set Up System Prompt	Once (refine as needed)
3	Input Idea → Get Draft	Daily (3–10 tweets)
4	Post, Track, Debrief	After high/low performers
5	Monthly Recalibration	Once per month

Why This Works (And Why Most AI Tweets Don’t)

The difference between this approach and asking ChatGPT to “write me a tweet” comes down to three things.

Compression. Twitter demands you say something meaningful in 280 characters. A generic prompt produces bloated output that you have to cut down. Your voice profile and platform rules train the agent to write tight from the start — fragments, punch, no filler.

Voice fidelity. On Twitter, your followers know how you sound. One off-voice tweet and people notice. The voice profile combined with anti-pattern rules prevents the agent from defaulting to that bland, helpful AI tone that reads like a chatbot tried to be relatable.

Feedback. The debrief loop is the entire game. Without it, you are using a static tool. With it, you have a system that adapts to what your specific corner of Twitter responds to. After two months of consistent feedback, the agent will produce first drafts you post with minimal edits.

What to Build Next

This blueprint covers Twitter/X, but the architecture is portable. Once you have your voice profile and system prompt working, you can adapt it for:

1. LinkedIn posts — Same voice core, longer format, more professional framing (see Blueprint #001)
2. Email newsletters — Same voice, longer format, add a section for curated links
3. Internal comms — Same structure, different audience definition, more context on company culture
4. Short-form video scripts — Same voice profile, add pacing and timing constraints

Each of these is a separate blueprint. If you want to see them, let us know in the comments which one to build next.

---

Frequently Asked Questions

Does this work with the free version of Claude?

Yes. The free tier handles everything in this blueprint. The only difference is that Claude Pro lets you use Projects, which means your system prompt persists automatically across conversations. On the free tier, you paste the system prompt at the start of each new conversation. The output quality is the same.

Can I use ChatGPT or Gemini instead of Claude?

You can adapt this workflow to any capable language model. The prompts are model-agnostic. That said, I built and tested this with Claude because its instruction-following and voice matching are strongest for this type of work. If you use a different model, you may need to adjust how you phrase the voice profile constraints.

How many tweets should I provide for the voice profile?

Ten to fifteen is the sweet spot. Fewer than that and the model does not have enough signal to identify real patterns. More than twenty and you start diluting the analysis with mediocre examples. Pick your best-performing tweets – the ones that sound most like you and got the most engagement.

Will people know my tweets are AI-generated?

If you skip the voice profile step and use generic prompts, yes – people will notice immediately. If you follow this blueprint properly, the agent produces drafts that sound like you because it has been trained on your actual writing. You should still edit every tweet before posting. The agent writes the first draft. You own the final version.

How do I handle the 280-character limit?

The system prompt includes a hard character-count rule, and the output format requires the agent to report the exact count for every tweet. Claude is good at respecting this constraint, but occasionally runs a few characters over. Always verify the count before posting. For threads, each individual tweet has its own 280-character limit.

How long before the agent gets good at my voice?

The voice profile gives you a strong starting point from day one. After two weeks of consistent use with feedback, you will notice the drafts require fewer edits. After two months, the agent will produce tweets you can post with minimal changes. The feedback loop is what accelerates this – skip it and the agent stays static.

Should I post AI-generated tweets without editing them?

No. Always review and edit before posting. The agent gives you a strong first draft, not a finished product. Your editorial judgment – knowing what lands with your audience, what is worth saying right now, what tone fits the moment – is something the agent cannot fully replicate. Think of it as a writing partner, not an autopilot.

Can this workflow handle quote tweets and replies?

Yes. Step 5 includes specific prompt templates for both. Quote tweets and replies need a different energy than standalone tweets – they need to add a new angle to someone else’s point without just restating it. The prompts are designed for that.

---