twitter-agent

name	twitter-agent
description	Twitter, tweet, timeline, mentions, X, social media, feed, engagement, replies, drafts, scoring (project)

Twitter Agent

Twitter/X social management agent. Extracts feed, scores posts for relevance, and creates draft responses for the user's approval.

Critical Rules

1. DRAFTS ONLY - Never auto-post anything. All content must be approved by the user.
2. TOP 0.01% ONLY - Only surface posts scoring > 0.9 (extremely selective)
3. Follow Operating Manual - See references/operating-manual.md for voice, audience, and approval gates

Quick Start

# Extract posts from timeline (requires Playwright MCP login)
python twitter.py extract --limit 20

# Score unscored posts (surfaces high-relevance only)
python twitter.py score-feed

# View high-relevance posts
python twitter.py high-relevance

# List pending drafts
python twitter.py drafts

# Create a draft reply (does NOT post)
python twitter.py draft-reply POST_ID "Response text"

Workflow

Timeline ─> Extract ─> Score ─> Present High-Relevance ─> User Reviews
                                        │
                            Create Drafts (if warranted)
                                        │
                              User Approves/Rejects
                                        │
                              User Posts Manually

Setup: Playwright MCP

Twitter extraction requires Playwright MCP with persistent login:

# Install Playwright MCP with persistent browser session
claude mcp add playwright -- npx @playwright/mcp@latest --user-data-dir ~/.playwright-data/twitter

# First time: log in via the browser
# Subsequent runs: session persists

Target Audience (Score HIGH)

Per the user's Operating Manual, prioritize these audiences:

• AI for science researchers with serious engineering taste
• HPC, scientific computing, data infra engineers
• Lab automation and autonomous experimentation operators
• Hard tech founders and early employees (materials, energy, defense)
• Defense and energy program people who care about capability
• Deep tech investors with real full stack instincts
• Materials science graduate students and postdocs
• Industry R&D engineers (batteries, catalysts, semiconductors, aerospace)

Score LOW (Avoid)

• Generic AI hype accounts
• Engagement farmers
• People who only want vibes and hot takes
• Content about generic AI trends, not scientific computing

Engagement Rules

The user should ONLY engage when he can:

• Add a missing system detail or infrastructure layer
• Ask a pointed, specific question
• Clarify a technical distinction
• Provide HPC/simulation workflow insight
• Build a useful relationship with someone in target audience

Functions

Feed Extraction

from twitter import extract_feed, store_extracted_post

# Get extraction instructions (uses Playwright MCP)
result = extract_feed(limit=20)
# Follow the instructions with browser tools

# Store each extracted tweet
store_extracted_post(
    post_id="1234567890",
    author_name="Display Name",
    author_handle="username",
    content="Tweet content here",
    engagement={"likes": 50, "retweets": 10}
)

Relevance Scoring

from twitter import score_relevance, score_feed, get_high_relevance_posts

# Score a single post
result = score_relevance(post)
# Returns: {score, audience_match, engagement_fit, reasoning, suggested_action}

# Score all unscored posts and get high-relevance ones (>0.9)
high_relevance = score_feed(min_score=0.9)

# Get already-scored high-relevance posts
posts = get_high_relevance_posts(min_score=0.9)

Draft Creation (NEVER Auto-Posts)

from twitter import draft_reply, draft_quote, draft_dm, get_pending_drafts

# Create draft reply to a post
result = draft_reply(post_id=42, suggested_text="Great point about HPC...")
# Returns: {success, draft_id, message: "the user must review and post manually"}

# Create draft quote tweet
result = draft_quote(post_id=42, suggested_text="This connects to...")

# Create draft DM
result = draft_dm(handle="username", suggested_text="Quick question...")

# List all pending drafts for the user to review
drafts = get_pending_drafts()

Draft Management

from twitter import approve_draft, reject_draft

# the user approves a draft (still must manually post)
approve_draft(draft_id=1)

# the user rejects a draft
reject_draft(draft_id=1)

CLI Commands

# Extract tweets from timeline
python twitter.py extract [--limit N]

# Store a tweet manually
python twitter.py store POST_ID --author "Name" --handle "username" --content "Text"

# Score all unscored posts
python twitter.py score-feed [--min-score 0.9] [--show-all]

# Score a single post
python twitter.py score-post DB_ID

# Show high-relevance posts
python twitter.py high-relevance [--min-score 0.9]

# List pending drafts
python twitter.py drafts

# Create drafts (DOES NOT POST)
python twitter.py draft-reply POST_ID "Response text"
python twitter.py draft-quote POST_ID "Quote text"
python twitter.py draft-dm HANDLE "DM text"

# Approve/reject drafts
python twitter.py approve DRAFT_ID
python twitter.py reject DRAFT_ID

When User Says...

User Says	Action
"Check my Twitter"	Extract feed, score, show high-relevance posts
"Any good engagement opportunities?"	python twitter.py high-relevance
"Draft a reply to that post"	draft_reply(post_id, text) + show draft
"Reply to @username about X"	Search posts, create draft_reply
"DM that person"	draft_dm(handle, text)
"Show pending drafts"	python twitter.py drafts
"Approve draft 5"	approve_draft(5) + remind the user to post manually

Approval Gates (From Operating Manual)

Can post without asking:

• Technical observations with no sensitive claims
• Neutral questions
• Replies that are not political and not personal

Must get user approval on:

• Any thread
• Anything that names a person as a claim or critique
• Anything that references a partner, lab, employer, or program
• Anything with a strong defense or policy stance
• Anything mentioning the user's product names, fundraising, hiring, or partnerships

Voice Guidelines (From Operating Manual)

• Direct, technical, calm
• Curious and skeptical
• High signal, low drama
• Precise about what is known vs assumed
• Short sentences
• No emoji unless it adds clarity (max one)
• No hashtags
• No em dashes

Ban list:

• programmable matter (unless milestone)
• revolutionize, disrupt, democratize, game changer, huge

Database Tables

social_feed

Stores all extracted posts:

Column	Description
id	Database ID
platform	Always "twitter"
post_id	Twitter's post ID
author_name	Display name
author_handle	@handle (no @)
content	Tweet text
engagement	JSON: {likes, retweets, replies}
relevance_score	0.0-1.0 from scoring
suggested_action	reply/quote/like/ignore
action_taken	What was done
captured_at	When extracted

twitter_drafts

Stores draft responses:

Column	Description
id	Draft ID
post_id	FK to social_feed (for replies/quotes)
draft_type	reply/quote/dm
target_handle	For DMs
draft_text	Draft content
status	pending/approved/rejected
created_at	When drafted
approved_at	When approved

Related Skills

• browser-automation: How to use Playwright MCP
• vault-operations: Captures for context
• relationship-intelligence: Track people engaged with
• pcp-operations: System overview

References

• references/operating-manual.md - the user's full X Operating Manual

---

Remember: This is a DRAFT-ONLY system. Never auto-post. the user must manually post all content.