Category: Uncategorized

Do the math. If you spend 15 minutes a day briefing ChatGPT on who you are, who you sell to, and how you write — that’s 91 hours a year. Two full workweeks. Gone to a machine that forgets you the second you close the tab.

The #1 complaint about AI isn’t hallucination. It’s amnesia. And it’s fixable in an afternoon.

Here’s the exact system I run at Alington Media. Three layers. Build it once. It compounds forever.

Layer 1: The Master Context Doc

One markdown file. Not a Notion database. Not a folder of PDFs. One file you can paste into anything.

Mine lives in a doc called context.md and it has exactly these sections:

• Who I am — name, role, company, what I actually do day to day (not a LinkedIn bio)
• What we sell — the offer, the price range, the delivery model, what we do NOT do
• Who we sell to — ICP with specifics (industry, team size, current pain, what they’ve already tried)
• Voice rules — 5-10 concrete do’s and don’ts. Not “be conversational.” More like: “never start a sentence with ‘In today’s landscape.’ Use contractions. One-sentence paragraphs are fine.”
• Recent context — 3-5 bullets on what I’m working on this quarter, so the model knows the season I’m in

Keep it under 800 words. If it’s longer, you’re writing a brand book, not a context file.

How to actually write it

Open a fresh Claude or ChatGPT session. Paste 10 of your best pieces of writing (LinkedIn posts, emails, sales pages). Ask it: “Extract the voice rules a ghostwriter would need to sound like me. Be specific and prescriptive.” Edit the output. That’s your voice section done in 20 minutes.

Layer 2: A Prompt Library of 6, Not 60

Most people collect prompts like Pokémon. You don’t need 60. You need 6 that cover 90% of your actual work.

Here’s mine:

• Content draft — turn a raw idea or voice memo into a post in my voice
• Sales reply — draft a response to an inbound lead using my qualification frame
• Loom-to-SOP — turn a transcript into a documented process
• Offer critique — pressure-test a landing page or proposal against my ICP
• Weekly review — summarize what I shipped and what to prioritize next
• Research brief — pull structured notes on a prospect or topic

Every single one starts with the same first line: “Read the attached context.md before responding. Confirm you’ve internalized voice rules before drafting.”

Store them in a plain text file, a Raycast snippet library, or a Claude Project. Whatever you’ll actually open at 7am.

Layer 3: The n8n Auto-Injector

This is the part that turns a system into infrastructure.

Instead of manually pasting context.md into every session, I run an n8n workflow that does it for me based on the task. Here’s the shape of it:

• Trigger — a webhook, a Telegram message, or a new row in an Airtable “tasks” table
• Router — a Switch node that reads the task type (content, sales, SOP, etc.)
• Context loader — pulls context.md from Google Drive or a GitHub gist
• Prompt builder — a Set node that concatenates: system prompt + context.md + the specific task
• LLM call — Claude via the Anthropic node (I use Sonnet for drafting, Opus for critique)
• Delivery — result back to Telegram, Slack, or a Notion page

Total build time: about [2-3] hours if you’ve touched n8n before. If you haven’t, budget a weekend.

Why n8n and not a custom GPT?

Custom GPTs are fine for one-off use. But they live inside ChatGPT, they don’t compose with other tools, and you can’t swap the model. With n8n I can route the same context to Claude, GPT, or a local model, chain it to my CRM, and version the workflow in git. It’s infrastructure, not a toy.

What changes when you ship this

A few things I didn’t expect:

• My drafts stopped sounding like a LinkedIn guru because the voice rules are always loaded
• Onboarding a VA got easier — I hand them the same context.md I feed the model
• I stopped opening ChatGPT to “think” and started opening it to “produce” — the ritual friction was doing more damage than I realized

The 91 hours aren’t really about time. They’re about the cognitive tax of re-explaining yourself before you can even start working. Kill the tax once. Ship the system. Then spend those two workweeks on things that actually compound.

Start today, not next quarter

Don’t build the whole stack this weekend. Just do Layer 1. Write the master doc. Paste it into your next five ChatGPT sessions manually and see what changes. If it saves you even 5 minutes a session, you already know Layer 2 and 3 are worth building.

The best AI setup isn’t the fanciest model. It’s the one that remembers who you are.

AI didn’t kill your LinkedIn reach. Pasting Claude’s first draft did.

I run most of my business on Claude and n8n. Research, outlines, transcript cleanup, CRM enrichment, lead scoring, meeting prep — the model touches almost everything. But it doesn’t touch the words you actually read from me. Not on LinkedIn, not in emails, not in reels.

Here’s why, and here’s the exact system I use so you can steal it.

The tell everyone is missing

Every AI-drafted LinkedIn post has the same fingerprints:

• Opens with a rhetorical question or a soft setup line
• Uses the phrase "here are X lessons" or "here’s what I learned"
• Bullet points with parallel structure that’s too clean
• A closing line like "What’s your take?" or "Which one resonates?"
• Zero specific numbers, names, or scars

Your buyer reads 30 of these a day. So does the algorithm. The moment your post pattern-matches to the other 29, distribution collapses and human attention slides right past.

This isn’t about "AI writing bad." Claude can write beautiful prose. The problem is that generic writing gets generic reach, and a first draft from any LLM is generic by definition — it’s the statistical average of the internet.

The line I draw

Simple rule: AI does the thinking work I can verify. Humans do the choosing work only I can do.

What that means in practice:

• AI does the research. Pull sources, summarize threads, find counterexamples, compare frameworks.
• AI does the outlines. Give me five angles on this topic, ranked by contrarian potential.
• AI runs the pipelines. Transcription, tagging, scheduling, repurposing snippets.
• The human writes the hook. Because the hook is a taste decision, not a probability distribution.
• The human tells the story. Because only you were in the room.
• The human picks the fight. Because opinions require skin, not tokens.

My actual pre-post workflow

Here’s what happens before I post anything on LinkedIn. It takes me around [X] minutes per post.

1. Voice memo the raw idea (2 min)

I record a voice note the second I have a thought worth posting. Usually triggered by a client call, a mistake I made, or something I disagree with. Raw, unstructured, my actual words.

2. Transcribe and structure with Claude (3 min)

I run the transcript through Claude with a prompt like: "Here’s a voice memo. Pull out the single sharpest claim. List every specific detail I mentioned. Do not rewrite anything in your own voice."

Notice what I’m asking for: extraction, not generation. Claude is a highlighter, not a pen.

3. Write the hook myself

I open a blank doc and write the first two lines from scratch. This is non-negotiable. The hook is 80% of whether the post works, and it’s the one part that has to sound like a person who has been through something.

Bad AI hook: "Most founders don’t realize this about AI content…"

Human hook: "AI didn’t kill your LinkedIn reach. Pasting Claude’s first draft did."

Same topic. One reads like a template. The other reads like a person picking a fight.

4. Draft the body in my own words

I write the body using the details Claude extracted. I keep specific numbers, real names, actual timelines. If I don’t know a number, I leave a bracket, not a fake stat.

5. Use Claude as an editor, not a writer

Final pass, I paste my draft in and ask: "What’s soft here? Where am I hedging? Which sentence sounds like every other LinkedIn post?"

Claude is excellent at critique. It’s mediocre at replacement. So I take the critique and rewrite those lines myself.

Why this actually works for reach

LinkedIn’s algorithm isn’t (only) hunting for AI text. It’s hunting for dwell time and meaningful reactions. Both drop when a post reads generic, because readers pattern-match and scroll.

Specificity is what stops the scroll. Specificity is what prompts a DM. And specificity is exactly what a first-draft LLM strips out, because averages don’t have specifics.

When I paste a Claude draft, I get [low number] impressions and zero DMs. When I write it myself and use Claude as an intern behind the scenes, I get [higher number] impressions and [X] DMs — and one of those DMs is worth more than the whole month of "5 lessons I learned" posts combined.

The one prompt swap that fixes 80% of it

Stop asking Claude: "Write me a LinkedIn post about X."

Start asking Claude: "I’m writing a LinkedIn post about X. Here are my notes. Ask me 10 questions that would make this post more specific before I draft it."

Now Claude is doing what a good editor does: pulling the story out of you. You still write the post. You just write a much better one.

The honest test

Read your last five posts out loud. If they sound like something you’d say to a friend at a bar, you’re writing. If they sound like a keynote nobody asked for, the model is writing them for you.

Use AI as your research intern. Not your ghostwriter. Your buyer can tell the difference — and so can the algorithm.

The pitch is everywhere: a solo founder did $400M with zero employees, so you can run your agency with Claude, n8n, and a stack of agents that costs less than a junior hire’s lunch budget. I bought the pitch. I rebuilt my ops at Alington Media on a roughly $400/month stack. Thirty days in, here’s the unvarnished version.

Three things worked. Three things broke. One I had to hand back to a human.

The actual stack (and what each piece costs)

• Claude (Pro + API): ~$[100]/mo. SOP drafting, voice agents, summarization.
• n8n (self-hosted on Hetzner): ~$[15]/mo. Glue layer for every workflow.
• Vapi: ~$[80]/mo. Inbound voice for qualification.
• Twilio + OpenAI: ~$[120]/mo combined. SMS follow-up and embeddings.
• Cal.com + Notion + Airtable: ~$[60]/mo. Calendar, docs, ops database.
• Loom + Otter: ~$[25]/mo. Recordings into transcripts for SOPs.

Total: roughly $400/month. No employees, no contractors on retainer, just me and a folder full of agents.

What worked

1. Claude drafting SOPs from Loom transcripts

Every time I did a task manually, I recorded a Loom. Otter pushed the transcript into a Google Drive folder. An n8n workflow watched that folder, sent the transcript to Claude with a prompt template (“Convert this into a numbered SOP with prerequisites, steps, and a ‘done when’ check”), and dropped a draft into Notion. I edited about 20% of each draft. In 30 days I went from [3] documented SOPs to [27]. This is the single biggest unlock of the stack.

2. n8n pulling usage data into one sheet

I had no idea what my real cost per client was. I built an n8n flow that hits the Vapi, Twilio, and OpenAI APIs every morning, attributes usage by client tag, and writes it to one Google Sheet with a daily Slack summary. The first week I found a client whose voice agent was burning [4x] what they paid. I fixed it that day. Without the pipeline, I’d have caught it at month-end, if at all.

3. The invoice sub-agent

A small Claude agent reads the weekly usage sheet plus my Notion “work log” and drafts invoices with line items including the small stuff I always used to forget: extra calls, integration changes, urgent requests. I review and send. Revenue per client went up because I stopped underbilling myself.

What broke

1. The voice-match writing agent drifted

Week one it sounded like me. By week two it sounded like a LinkedIn ghostwriter on melatonin: more em-dashes, more “in a world where,” zero edges. The problem wasn’t the model. It was that I kept giving it feedback in chat instead of updating the system prompt and the few-shot examples. Each new conversation started from a slightly stale baseline. Fix: I now version the prompt in Git and require new examples (good + bad) every Friday. It’s better. It’s still not me.

2. The support agent invented a refund policy

A client asked about refunds. The agent confidently quoted a 14-day policy. We don’t have a 14-day policy. We don’t have any refund policy written down, which is the actual root cause. The agent filled the vacuum with plausible nonsense. Lesson: agents amplify whatever’s missing in your documentation. I pulled it offline until I’d written the real policy and grounded the agent in a retrieval doc with a hard “if not in source, escalate” rule.

3. The scheduler double-booked me

The auto-scheduler tried to be helpful by holding tentative slots across two calendars. It double-booked me [4] times in two weeks, including once on a paying client call. I killed it. Cal.com with one round-robin link and a manual buffer rule is now doing 95% of the job at zero risk.

What I had to take back

Client strategy calls. I tried to use an agent to prep a strategy doc, run a partial discovery via async video, then have me only join for the close. Clients hated it. They didn’t pay me for a workflow, they paid me for the judgment in the room. The async layer was extracted to a pre-call intake form. The call itself is mine again.

The real lesson

The stack is real. It’s just not a team. It’s an extremely fast intern with no taste, no memory, and no skin in the game. That framing changes how you design the system:

• Give it bounded tasks with a clear “done when” and a source of truth to ground in.
• Review everything that touches a client for the first 60 days, minimum.
• Document before you automate. Agents amplify the doc; if the doc doesn’t exist, they invent.
• Keep judgment work human. Strategy, pricing, hard conversations, anything where being wrong is expensive.

Your job didn’t disappear. It moved. From doing the work to catching where the work goes wrong. That’s the actual one-person unicorn skill, and nobody’s selling a course on it yet.