How to Turn Claude Into Your Personal Assistant

Claude can run the boring parts of your life, and setting it up is easier than you think. I use Claude Code as a personal assistant across three buckets — sales and productivity, research, and content — and it saves me 5 to 10 hours every single week. Here's my exact setup, the skills and automations I actually run, so you can copy the parts that apply to you.

Most "AI OS" videos focus on system design — skill architecture, how to structure your Obsidian vault. This is different. This is the real assistant: what it does every morning, what it drafts, what it researches, and where the human still has to stay in the loop.

What Should a Claude Personal Assistant Actually Do?

Start by finding the drudgery. The work that eats hours, doesn't require much decision-making, and is pure grunt work — that's what you offload to Claude Code.

For me, that splits into three buckets:

• Productivity and sales — tied to my AI agency business
• Research — staying on top of a fast-moving space
• Content — the social media side

Most of this translates even if you don't create content. If you work in sales or marketing, you have information coming in from clients, replies you owe, and daily research you need to stay current on. The specifics are mine. The pattern is yours to steal.

How Does the Email Triage Automation Work?

This is the part that gives most people the fastest payoff. Every morning at 8 a.m., a skill scans my Gmail inbox using the Claude connector, reads everything from the last 24 hours, and sorts it into buckets: leads, urgent, warm, sponsors, meetings, or noise.

Once everything is bucketed, Claude does different things depending on the bucket:

• Sponsors — it auto-drafts a reply with a boilerplate response and a link to my media kit, including pricing.
• Leads — it runs a deeper process (more on that below).
• Urgent — it drafts a potential reply.
• Everything else — it just marks and summarizes.

The lead process is where it earns its keep. On my website, every lead fills out a form with their budget, what they want, and their timeline. Claude Code picks all of that up in the triage. If it looks like a real inquiry — the email makes sense, they put in effort, and the budget aligns — Claude kicks off a web search to pull background on their company. Then it gives me a recommendation: here's what they want, here's what I found, should you move forward or not?

I'm still the arbiter. If I say yes, it creates a draft reply with a link to my calendar so they can book. If not, it skips them.

It also writes a full markdown report into my Obsidian vault — a summary I can read at a glance, including the noise I don't care about. The result: I don't read every email, I don't manually filter every lead, and the ones that matter arrive pre-researched with a draft already written.

To adapt this yourself, ask three questions: What emails do you normally get? How would you bucket them? And most importantly, what do you want Claude to do with each bucket once it's sorted?

How Do You Automate Proposals and Follow-Ups?

The other big time-saver on the sales side is proposals. This kicks in after a discovery call.

I use Calendly. Post-call, I get a recap in my inbox — an automatic summary with all the key points we discussed. That recap gets sent to Claude Code, which generates a branded PDF proposal: the scope of work, the pricing, the timeline, and a signature form.

I review it, obviously. But that beats building a proposal from scratch every time. Claude then generates a Drive link and a draft email to the client with the link included. The proposal breaks down the engagement, shows the timeline, lays out the investment, and leaves room for signatures at the bottom.

Follow-ups get more specific to the AI audits I run, so I'll leave those aside — they're technical and won't apply to most people. But the core lesson holds: the boring, repeatable parts of your sales motion are exactly what Claude should handle, so you spend your time on the actual conversations.

Where Do You Get Your Research, and How Do You Automate It?

Research starts with one question: where does your information actually originate? For the AI space, that's a handful of places — X (Twitter), GitHub, and sometimes YouTube. Once you know your sources, the setup is the same regardless of niche.

My main research automation is a daily brief that hits those sources and figures out what's going on:

• GitHub — uses the GitHub API to find trending repos. It pulls the top 10 trending for the week, the top 5 trending among repos created in the last 30 days, and the fastest-growing over the last 24 hours and 30 days. That combination tells you what's brand new versus what's been around and is suddenly climbing. This changes daily and is nearly impossible to track manually.
• YouTube — shows trending videos with the creator, views, and subscriber count. That last detail matters: I don't care if someone with a million subs gets 10,000 views, but if someone with 2,000 subs gets 10,000 views, they're clearly onto something people care about.
• X and general web search — pulls the big headlines and what's trending.

It does some basic content-opportunity analysis too, but honestly, AI is hit or miss at telling you what to make. What I actually care about is the raw data it gathers. The value is that I don't have to do it myself.

The email triage and the research brief are combined into a single morning automation, set up as a routine inside Claude Code. It's a local routine — it runs on my computer as long as it's open and fires off both automatically.

What About On-Demand Research Skills?

Beyond the daily brief, I run a few more research tools:

• X pulse — an app I built that lives on Railway, running 24/7. It sends me Telegram messages every hour with the top trending tweet in the AI space, tied to major creators including people on the Claude and OpenAI teams. If something big breaks, I know immediately without sitting on X refreshing.
• Deep research — this one's already built into Claude Code. Run /deep-research and you get a preloaded dynamic workflow that can spawn up to around 100 sub-agents doing adversarial information gathering. They browse the web, test findings against each other, and synthesize. It's a standard web search on steroids — it burns a lot of tokens, but most people never use it.
• YT pipeline — searches YouTube for videos relevant to my question, sends the URLs to NotebookLM, and lets NotebookLM handle the transcripts and synthesis. I use the notebooklm-py CLI, which gives Claude Code an unofficial API to NotebookLM. That means I get all of NotebookLM's power without spending tokens — Google's servers do the synthesis on the transcripts.

The point of the YT pipeline is simple: I don't want to watch 10 YouTube videos on a topic. Give me the summary. Do the synthesis. Tell me where all 10 agree, where they don't, and why I should care.

Put it together and the research setup alone saves me 5 to 10 hours a week, minimum. If I tried to do all of it manually, I'd be drowning. And research pairs naturally with productivity — in most jobs, the emails you send and the things you create are based on research you had to do anyway. Combine them.

How Do You Use Claude for Content Without Sounding Like AI?

Content is the bucket most people won't need, but here's how I approach it — and the honest limits.

• Hooks and outlines — a single on-demand skill. I trained Claude on a bunch of Kallaway's videos (a strong creator on YouTube) so it understands how to build a hook and think about transitions. But I don't script my videos, so this is a brainstorming tool, not a writer. I don't rely on AI for the words I say — I use it for back-and-forth that helps my own ideas coalesce.
• Packaging — same deal. I do thumbnails entirely on my own. For titles, Claude uses those Kallaway-style psychological triggers and looks at which of my past titles performed and why. But it's a back-and-forth, not a hand-off. AI is bad at things that require taste, so a mixed approach wins.
• Content repurposing — this is where AI genuinely shines for me.

The repurposing setup runs twice a day, at noon and 8 p.m. It checks whether I've posted a new YouTube video, then runs a content cascade skill: it fetches the transcript and rewrites it into a blog post, a LinkedIn post, and a tweet.

How Do You Train Claude to Write in Your Voice?

This is the part people get wrong. There is no one-shot, self-improving skill that captures your voice. Anyone selling you that is wrong. The only way is a relentless human-in-the-loop cycle:

1. Give Claude examples of your own writing — writing you did 100% yourself.
2. Tell it to turn that into a skill.
3. It runs the skill and gives you an example output.
4. Destroy that example. Eviscerate it. This is wrong, this is wrong, this is right, here's how I'd change it.
5. It updates the skill and gives you a new example. Destroy that too.
6. Repeat about 10 times — and keep doing it live afterward.

Over time, you get a skill that actually sounds like you. There's a huge human component to any voice or humanizing work. I did this separately for the blog, because my blog voice differs from my LinkedIn voice, which differs from my tweet voice.

The payoff: my LinkedIn posts and blog posts get written and posted automatically from a single YouTube video, in my voice, on work I'd otherwise never get around to.

Why Does Obsidian Matter for a Claude Assistant?

Once you're generating briefs, drafting emails, and creating content, you need structure — a brain the assistant can reference. That's where Obsidian comes in.

Does Obsidian change how Claude Code works? Not really. The knowledge graph is a nice visual, but the real value is organization: a map you hand to Claude Code so it can answer your questions quickly and correctly.

The most common structure is the Karpathy method. Think of everything as folders:

• Raw — unstructured data. Claude dumps research here before any synthesis.
• Wiki — synthesized reports built from the raw material.
• Output — finished artifacts like slide decks.

Under each folder, you nest more subfolders by use case, and every folder gets an index file — a table of contents that tells Claude Code what's there and where to find it. That's what keeps it from getting lost.

Here's the real reason it matters: it's not that Claude literally gets lost. It's that a disorganized vault increases token cost and decreases accuracy as your file structure grows. Think of the vault as a filing cabinet for Claude Code. The knowledge graph is cool, but the filing cabinet is the value play.

Do You Need a Dashboard or Command Center?

All this generated information needs a home you can actually see. The terminal isn't it — that's not what it's for, and neither is the Claude Code desktop app.

That's the case for a command center or web dashboard. It's a one-stop shop that has to be 100% customized to you, which is exactly why you can't get it off the shelf. In my Obsidian-based setup, I have my metrics, my token burn, and every automation I've mentioned — all runnable at the click of a button. The audience section is a different view of the morning report. My web app version mirrors the same thing: metrics, skills, and the documents that pull up the exact reports I'd find in my vault.

The division of labor is clean: the dashboards give you observability, Obsidian keeps it organized, and the skills and automations do the actual work. Put those together and Claude works for you nearly hands-off.

Frequently Asked Questions

How much time does a Claude Code personal assistant actually save?

In my setup, 5 to 10 hours every week — and the research automation alone accounts for a big chunk of that. The savings come from eliminating email filtering, manual research across X, GitHub, and YouTube, proposal drafting, and content repurposing. Your number depends on how much of your work is low-decision grunt work.

How does Claude decide which leads are worth pursuing?

It reads the lead form (budget, needs, timeline), checks whether the inquiry is genuine and the budget aligns, then runs a web search on the company. It returns a recommendation, but you make the final call — the assistant drafts and researches, the human decides.

Can Claude really write content in my own voice?

Yes, but only through a repeated human-in-the-loop training cycle, not a one-shot skill. You feed it your real writing, it drafts, you tear the draft apart, it updates the skill, and you repeat roughly 10 times. There's no shortcut — anyone promising a self-improving voice skill is overselling it.

Do I need Obsidian to run a Claude assistant?

Not strictly, but you need some organizational structure. Obsidian works well because a folder hierarchy with index files acts as a filing cabinet — it lowers token cost and preserves accuracy as your file structure grows. Without organization, a big vault gets more expensive and less accurate over time.

What's the difference between the automations and the on-demand skills?

Automations like the morning brief (email triage plus research) and the X pulse run on their own with no input from you. On-demand skills like deep research and the YT pipeline are ones you trigger when a specific topic comes up and you want to go deeper.

---

If you want to go deeper into building your own Claude Code personal assistant, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.