.svg){kind=small}
Naga is the founder of Thought Leader, a content agency that helps B2B SaaS and AI SaaS founders grow on X. I can vouch for the output personally: his system generated 8.4 million impressions on my X account in less than six months.
The surprising part is what powers it. The entire agency runs on content orchestration: a single AI agent inside Claude Code, connected to Notion and a scraping API, that ideates, scores, writes, fact-checks, and ships every piece of content from one terminal.
Naga walked me through the full system on a recent call, layer by layer. This article breaks down exactly how the orchestration works, from the memory architecture to the algorithm scoring, and how you can build a version of it yourself.
1. The Architecture: One Agent, Three Layers
The system is built around a single content agent: an orchestrator running inside Claude Code that triggers skills, reads memory, and communicates with the outside world.
Naga describes it like a human brain. It needs its own memory, its own organs to act, and external input to process information. The content agent mirrors that with three layers: a memory layer stored in Notion, an action layer made of skill files, and an external communication layer built on scraping APIs and Slack notifications.
The tool spread is deliberately minimal. Connect Notion, connect Apify, optionally connect Slack, and the agent handles everything on X end to end: tweets, threads, and long-form articles. You never leave the terminal.
2. The Memory Layer That Makes It Smart
Naga's view is that any AI system is only as good as the data it has. The memory layer is where the intelligence lives, and it holds six components inside Notion.
Client profiles store every piece of publicly available, verifiable information about each client. The agent knows who it writes for before it writes a word.
The content library stores every post a client has published across platforms: X, LinkedIn, and YouTube transcripts. This becomes the raw material the agent can repurpose and learn from.
The drafts memory archives every draft the system produces. This solves the biggest weakness of AI writing systems: recycling the same ideas, phrases, and copywriting patterns. The agent rereads its own drafts before writing anything new, so it never repeats itself.
The idea bank logs every idea from every ideation session, organized per client. Nothing brainstormed gets lost.
Sentinel intel is the most original piece. A sentinel account is an inspiration account: a profile performing well on X that aligns with a client's niche. The system tracks these accounts continuously and extracts which post types, hooks, and posting times work for them.
Everything above feeds into a cache layer. Naga consolidates all six databases into a local cache because overloading an LLM with raw data degrades output quality. The agent only reads from the cache: the last 50 tweets, the last 30 LinkedIn posts, the latest YouTube transcripts, and the top-performing posts.
3. Skill Files and Voice Profiles
The action layer is a collection of Claude Code skills, and the depth of these files is what separates this system from a prompt template.
Each client gets dedicated writing skills per platform, because writing for X and writing for LinkedIn are different crafts. On top of that sits a voice profile: the phrases, energy, enthusiasm, and tone that make a person sound like themselves across platforms. A voice checker skill validates every draft against it.
Naga's process for building a skill starts with scraping every post the client has written, then classifying them by reach. He then infuses copywriting frameworks from direct response and VSL copywriters into a hybrid rule structure. One level is trained on the client's data, the other on writing fundamentals.
A single mega article skill file contains 14 traits: rules for the hook, the revelation, the opinion, voice application, pre-flight checks, verified numbers the agent is allowed to use, banned AI patterns like the "not X, but Y" construction, and sentence length variation.
4. The Ideation Loop
A typical session starts with one command: run content agent for a given client. The agent reads recent post performance, reports whether the account's average is trending up or down, and suggests content pieces and formats that have not been used yet.
It also reads YouTube. A weekly routine pulls transcripts from the client's latest videos, so the agent can flag topics covered in long-form video that never made it to X.
Two guardrails make the ideation trustworthy. First, cooldowns: the agent checks 45 days of history and flags any overlap with past content. During the demo, it ranked a topic S tier but recommended against a long-form article because it collided with a piece published weeks earlier.
Second, number caveats. When the agent takes creative liberty with a framing or a stat it cannot verify, it does not present it as fact. It flags the claim and asks the client to validate or substantiate it first.
This ideation loop, reading your ICP and your content history to surface fresh angles, is the same principle we productized at ColdIQ in a smaller package.
You can generate campaign ideas based on your ICP and content strategy in seconds, for free:
Campaign Ideation Tool
Enter your email to generate your campaign ideas.
5. Scoring Drafts Against the X Algorithm
Twitter published its recommendation algorithm on GitHub, first around 2017 and again recently with the updated version. Naga forked both repositories and turned them into an algo scorer skill that ranks every topic and draft.
The scorer reads what the source code rewards. Bookmarks and replies carry the most weight, so anything with high bookmark probability gets more distribution. Each draft receives scores for hook velocity (how punchy and simple the opening is), reply magnetism (how likely someone is to respond), and negative risk.
The skill also encodes patterns observed in practice. Time decay degrades a post's performance every few hours. And when one post goes viral, the next posts inside a 24-hour window rarely get boosted, because the algorithm prioritizes showing larger accounts first.
Topics come back ranked in tiers, from S tier down to B, with the reasoning attached. Naga cross-references the open-source theory against real client post data, so the scorer reflects what works in practice rather than what the documentation claims.
LinkedIn has no public algorithm, so the LinkedIn scorer was built differently: pure scraping through Apify, cross-referencing post patterns and timing to understand what performs.
Packaging matters as much as the algorithm on LinkedIn, and seeing your post the way the feed renders it is half the battle.
You can preview how your LinkedIn content will look before publishing, for free:
LinkedIn Post Previewer Tool
Post Content
Paste your LinkedIn URL to auto-fill your profile
Preview
6. Production: Fact-Checks and AI-Tell Scans
Once a topic and format are locked, the agent runs a pre-flight checklist before writing: posting cadence, momentum, impression trends, and a full cache read.
Production is fast. A single content piece takes around 3 minutes, and the full journey from idea to polished draft lands around 30 minutes.
Naga appends one instruction to every production run: fact-check once done. Despite the entire memory architecture, Claude can still invent details, so the agent finishes every draft by separating verified claims from unverified ones.
It then runs a voice and AI-tell scan on its own output. The system counts banned constructions (a maximum of three "not X, but Y" patterns per 2,000 words), checks for machine-gun sentence rhythm, and validates the draft against the client's voice profile.
The format decision stays human. The agent can recommend whether a topic deserves an article, a thread, or a single tweet, but Naga reserves that judgment for himself and never generates every version up front, which he considers wasted tokens.
7. Distribution From the Terminal
Finished threads ship to Typefully through MCP. The agent pushes drafts, schedules them, and tags them without anyone opening the Typefully interface.
Long-form articles route to Notion instead, because Notion formatting transfers cleanly into X articles, while other editors break the formatting on paste. X rolled out article scheduling only recently, so articles still get published manually.
Notion doubles as the article registry. Every published piece is logged with its content pillar, the key stats used, the core thesis, the closing line, and the word count. Slack and Telegram are wired in for pings when routines finish.
8. The APIs Behind the System
The external layer runs on a short list of APIs, and the volume is bigger than you would expect: over 1,000 tweets and posts scraped on a normal day.
Apify is the backbone. Naga uses the Harvest API for LinkedIn data, plus the Tweet Scraper and the LinkedIn Profile Scraper. One Apify token inside Claude Code covers nearly every platform he needs to scrape.
The YouTube Data API handles video transcripts on its free tier, with Apify as the fallback when it hits limits.
The Claude API powers the scheduled routines: a sentinel scan, an hourly loop, a daily sync, and a daily digest. These run as cron jobs with the API routed in as a fail-safe, so the system keeps working even when nobody is at the terminal.
Tracking accounts for inspiration is one flavor of signal monitoring. We apply the same principle to buying signals at ColdIQ, where the companies showing intent are the ones worth contacting first.
If you want to see which companies are actively researching solutions in your space right now, you can do it for free here:
Intent Signals Tool
Fields marked with * are required
Quick examples:
9. How to Build Your Own Version
Naga spent a couple of weeks building this incrementally, with 30 to 40 hours of total compute. He started with elaborate prompts, turned them into skill files, evolved those into an orchestrated agent, then added the memory layer and the routines. The whole system lives in a GitHub repository he updates continuously.
His advice for replicating it is refreshingly simple. Start with one orchestrator that can write, read what is already online, and pull data.
Connect Notion to Claude Code through MCP as your storage. Grab an Apify token and ask the agent to scrape your last year of posts across LinkedIn, X, and YouTube. That becomes your content library.
For writing skills, use meta-prompting. Feed Claude the best copywriting templates you can find, and it builds its own skill files from them. If you admire a specific creator's style, scrape their posts and ask the agent to extract the patterns.
Then make one rule permanent: before writing anything, the agent checks the database for what already went out and what performed. The scorers, the sentinels, the cache, and the cron jobs are refinements you add once the core loop works.
The system is also modular by design. When Naga onboards a freelance writer, they receive only the memory layer and the skill files they need, never the full system.
10. Why Content Orchestration Works
Naga is explicit that speed is not the point. The reason Thought Leader orchestrates everything through Claude Code is intelligence: surfacing angles and insights no human could find, because nobody can read thousands of posts a week and remember what worked in each one.
The judgment stays human at every decision that matters: which idea to pursue, which format to use, and whether the final draft sounds right. The agent handles the reading, the scoring, the drafting, and the shipping.
Plenty of AI content systems look impressive in a demo and produce posts that get 20 likes. This one generated 8.4 million impressions on my account in under six months, books meetings from X, and averages 5 to 10 million monthly impressions across Naga's client base. The complexity connects to output, which is the only test that matters.
Benchmarking your own motion against systems like this one is the exercise we built our GTM reports around.
If you want to understand where your GTM motion stands today, see below how your current approach compares to these specialized models:
GTM Report Tool
Which part of this system would you build first: the memory layer, the algo scorer, or the voice profiles?
---
