Lighthouse Chart — Riley Chen

1. Opening / Bearings

You are running a 14k-subscriber product newsletter by yourself, writing every issue, holding the brand voice, and now trying to figure out whether AI can actually help — or whether it will quietly flatten the thing that makes the newsletter worth reading. That tension is the right one to be sitting with. The goal here is not to write faster by writing worse. It is to build a prompt library that protects your voice across drafts, cut the draft-to-send cycle down without losing what makes the work yours, and get a paid-tier essay series far enough along to show an editor.

Four years of professional writing means you already know what your voice sounds like. That is a significant advantage. Most people learning AI tools for writing have to build the voice and the prompts at the same time. You have the voice. The work here is translation: turning what you already know about your writing into instructions precise enough that a model can follow them without improvising.

Four weeks, three sessions a week, 45 minutes each. That is 12 sessions. Enough to build something real, not enough to waste on theory. Every session in this chart produces something you can use the same week.

2. Track Context

You are on the Creator/Comms track — the path for writers, editors, and communicators who want AI to amplify their voice and accelerate their process without replacing the craft that makes the work distinctive.

The central risk on this track is voice drift: the slow erosion of what makes your writing recognizable, replaced by a kind of competent-but-generic prose that AI produces when it has no strong constraints. The antidote is not using AI less — it is giving the model more precise instructions, not fewer. A vague prompt gets a vague draft. A prompt that encodes your actual voice rules gets something closer to a first draft you'd actually send.

The secondary risk is workflow complexity. Writers on this track sometimes build elaborate systems — multi-step automations, custom GPTs, API integrations — and then stop writing while they maintain the system. Given that becoming a developer is explicitly not a goal here, the plan stays entirely within Claude.ai and Claude Desktop (no code, no APIs, no new paid tools). The prompt library lives in Notion, which you already use. The AI layer lives in Claude.ai Pro, which you already pay for. Nothing new to learn to operate; everything new is about how you prompt.

The opportunity on this track is significant. A well-built prompt library functions like a writing partner who has read everything you've ever written and knows your rules cold. The draft-to-send cycle shortens not because the AI is writing your newsletter, but because the first draft you get back requires fewer structural repairs. You are editing toward your voice rather than rebuilding from scratch.

The paid-tier essay series is a natural extension of this. Longer form, more opinionated, more time-intensive to draft — exactly the kind of work where a voice-locked prompt library pays the most.

3. Goals and Success Metrics

Three outcomes, in priority order:

1. Build a personal prompt library that protects voice across drafts. This is the foundation. Everything else — faster drafts, the essay series — depends on having prompts that reliably produce output in your voice. The success signal is concrete: two newsletters shipped using the library, with draft-to-send time under 90 minutes. The library is not done when it exists in Notion. It is done when you have used it under real deadline pressure and it held.

2. Cut newsletter draft cycle time in half without losing voice. The "without losing voice" clause is load-bearing. A draft that takes 45 minutes but requires two hours of voice repair has not saved time. The metric is draft-to-send time, not draft time alone. Log your current baseline in Session 4 (the first live send) and again in Session 7 (the second send). Two data points is enough to see a trend.

3. Ship a paid-tier essay series outline with three drafted issues. The success metric here is external: the outline reviewed and approved by an editor. That means the deliverable is not just a document in Notion — it is a package you are comfortable putting in front of another person. The plan builds toward this in the second half of the four weeks, after the prompt library is established and tested.

What this plan does not do: It does not turn you into a developer. No code, no APIs, no automations. It does not replace your brand voice with a model voice — the entire architecture of the prompt library is designed to prevent exactly that. If a session produces output that sounds like a press release, that is a signal to fix the prompt, not to accept the output.

4. Baseline and Pacing Note

You have been using Claude.ai for less than a year, and your comfort with writing-with-AI is already at 4 out of 5. That is the number that matters most for this plan. You know how to work with a model on prose. What you are building now is structure around that instinct — prompts that encode what you already do intuitively, so you do not have to re-explain your voice every session.

Your comfort with building and automation is at 1. That is fine. Nothing in this plan requires either. The prompt library is a Notion document with a few pages. OpenBrain (covered in Section 7) is a one-time setup that takes one session. After that, it runs in the background.

You know what a system prompt is and you understand context windows. That is enough vocabulary to work with. The plan does not introduce new jargon — it uses those two concepts directly and practically.

Pacing note: 45 minutes is a short session. The sessions in this plan are designed to produce one concrete thing each — a prompt card, a draft, a logged time. If a session runs over, stop anyway and note where you were. Incomplete sessions still produce partial deliverables, and partial deliverables are still useful.

5. Tasks and Projects

Weekly newsletter — the live proving ground

Your weekly newsletter is not a distraction from this plan. It is the test environment. Every prompt you build in the first two weeks gets used on a real issue in Weeks 2 and 3. This matters because prompts that work on hypothetical topics often drift when they meet a real deadline and real rough notes. The only way to know a prompt is reliable is to use it when something is actually due.

Week 1, Sessions 1–2: Build the voice rules and the first prompt card.

Start with your Notion brand voice guide. Read it as if you were a new contractor trying to follow it on day one. Mark anything vague — "warm," "conversational," "authentic" — and rewrite each vague rule as a concrete, testable sentence. "Be warm" becomes "use first person; avoid exclamation points; close with a take, not a question." Aim for 5–8 rules. These become the core of every prompt you write.

Session 2 turns those rules into your first prompt card: the voice-locked draft starter. The ready-to-paste prompt in your content assets is a starting point, but you will need to revise it after the first test. Read the output aloud. Mark every sentence that does not sound like you. Revise the prompt — not the output — to address the drift. The goal is a prompt that produces a draft requiring fewer than three voice repairs.

Week 1, Session 3 / Week 2, Session 1: Voice audit prompt and first live send.

The voice audit prompt (also in your content assets) is the second tool in the library. It does not write — it reads. Paste a draft in, get back a paragraph-by-paragraph verdict: on voice, slightly off, or off. Calibrate it on a draft you know is strong before you use it on a draft you are uncertain about.

Session 4 is the first live send. Start a timer when you open your notes. Use the draft starter, run the audit, make targeted edits, send from Substack, stop the timer. Log the number. This is your baseline. It may not be under 90 minutes yet. That is expected. You are measuring, not performing.

Week 2: Expand the library and hit the 90-minute target.

Session 6 adds an angle-finding prompt — the tool for the moment before drafting, when you have a topic but not yet an observation. This is often where the most time goes. A prompt that returns three possible angles, each with a one-sentence observation and one concrete example, compresses that exploratory phase significantly.

Session 7 is the second live send, this time using the full three-prompt sequence: angle-finding, draft starter, voice audit. Log the time and compare to Session 4. By this point, the draft-to-send target of under 90 minutes is within reach — not because the AI is writing the newsletter, but because the structural decisions (angle, lead, close) are happening faster.

By Week 2, your personal prompt library — living in Notion and queryable from any Claude surface via OpenBrain — will contain at least three prompt cards built from your own voice rules. This is the direct realization of your first stated outcome: a personal prompt library that protects voice across drafts, stored in OpenBrain memory by Week 2 so future Claude sessions inherit your voice rules without you pasting them in every time.

Weeks 3–4: The paid-tier essay series.

Session 8 is the concept session. Write one paragraph answering: what does a paid subscriber get that a free subscriber does not, and why would they pay for it? That paragraph is the brief for the outline prompt. Use the paid-tier outline builder from your content assets to generate a three-issue structure, then revise the prompt inputs until the output matches your vision.

Session 9 is calibration. Read the three-issue outline aloud. Do the issues feel distinct, or do they blur? Ask Claude to identify overlap and suggest how to sharpen the distinctions. Decide whether the paid-tier voice is the same as your newsletter voice or slightly different — longer sentences, more vulnerable, more technical. Write one paragraph defining it. Add it to OpenBrain.

Sessions 10 and 11 are drafting sessions: Issue 1 and Issue 2, using your prompt library adapted for essay length (1,000–1,200 words instead of 600). Run the voice audit on each. The goal is a draft you would be comfortable sharing with an editor — not a final draft, but a real one.

Session 12 assembles the package: pitch paragraph, three-issue outline, Issue 1 draft, Issue 2 draft. Run a final voice audit across both drafts. Send to your editor or schedule a self-review. The paid-tier series outcome is complete when the package is in one place and in someone else's hands.

6. Setup and Constraints

What you have: MacBook Air, Claude.ai Pro, Substack, Notion. That is the entire stack for this plan. Nothing else is needed.

No new paid subscriptions this quarter. This constraint is fully respected. Claude Desktop is free. OpenBrain (OB1) is open source and free to self-host. Notion is already in use. No paid tiers, no new accounts.

Claude Desktop is the one new piece of software to install — it is free and runs on macOS. It is what allows OpenBrain to work as a persistent memory layer. The installation takes about 15 minutes and is covered in Session 5.

Substack stays as your publishing surface. The plan does not touch your Substack workflow — drafts move from Claude to Notion to Substack the same way they do now, just faster.

Notion is your prompt library home. Each prompt card lives on its own page, using the fill-in template from your content assets. Keep the library in one Notion database so you can search across cards when you are mid-draft and need a specific tool.

One practical note: Claude.ai Pro's context window is large enough to hold your full voice rules, a prompt card, and a draft in a single conversation. You do not need to manage context manually — paste everything into one conversation and work from there.

7. OpenBrain — Foundational Setup

OpenBrain is a personal knowledge layer that travels with you across Claude.ai, Claude Desktop, Claude mobile, and any other Claude surface you use. Think of it as a persistent memory that holds your voice rules, prompt cards, and working context — so that every Claude session starts knowing who you are and how you write, without you having to paste your brand voice guide in from scratch each time. It is designed to be set up once by anyone, not just developers.

Setup on your MacBook Air:

1. Install Claude Desktop from claude.ai/download. It is free and runs natively on macOS.
2. Clone or download the OB1 repository from https://github.com/NateBJones-Projects/OB1 and follow the README to add the OpenBrain MCP server. The README walks through each step; no terminal experience beyond copy-paste is required.
3. Once connected, open a new Claude Desktop conversation and type: "What do you know about my newsletter voice?" If OpenBrain is working, it will surface whatever you have stored. If the answer is "nothing yet," that is correct — you have not added anything yet.
4. Add your first entries: paste in your 5–8 voice rules and your first two prompt cards. These become the foundation of your memory layer.

Session 5 in this plan is dedicated to this setup. Do not try to rush it into another session — give it the full 45 minutes.

Three usage patterns for your track:

• Before drafting: Open Claude Desktop and ask "What are my voice rules for the newsletter?" Confirm they surface correctly before you start a draft. This takes 30 seconds and prevents the most common drift scenario: forgetting to include voice context in a prompt.
• After a prompt revision: When you improve a prompt card — tighten a rule, fix a drift pattern — update the OpenBrain entry immediately. The library is only as good as its most recent version.
• When starting the essay series: Add your paid-tier voice definition to OpenBrain as a separate entry from your newsletter voice. Ask Claude "What's the difference between my newsletter voice and my paid-tier essay voice?" — if it can answer accurately, the distinction is stored correctly.

The learn-session loop: At the end of any meaningful working session — a session where you revised a prompt, made a voice decision, or discovered something about how the model handles your writing — type "learn this session" in Claude Desktop. This captures the session's context so future conversations inherit it. It takes ten seconds and compounds over time: by Week 4, Claude Desktop will have four weeks of your voice work in memory.

8. Resources and Links

5 minutes

OB1 — OpenBrain MCP server (README) https://github.com/NateBJones-Projects/OB1 Read the README before Session 5. It is short. Knowing what you are installing before you install it makes the setup session go faster.

Anthropic prompt engineering overview https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview Skim the sections on role prompting and few-shot examples. Both are directly applicable to voice-locked prompts. You already know what a system prompt is — this fills in the surrounding context.

30 minutes

Anthropic prompt library — writing and editing examples https://docs.anthropic.com/en/docs/resources/prompt-library Browse the writing-focused examples. You are not looking for prompts to copy — you are looking for structural patterns (how role + rules + task are sequenced) that you can apply to your own voice rules.

Claude.ai Projects feature https://claude.ai (navigate to "Projects" in the left sidebar) Claude Projects lets you set a persistent system prompt for a project — your newsletter, your essay series — so the voice rules are always loaded without you pasting them. This is a free feature within your existing Pro account and complements OpenBrain. Worth 20 minutes to set up a "Newsletter" project with your voice rules as the system prompt.

Deep dive

Substack's guide to paid tier positioning https://on.substack.com/p/paid-posts Useful background before Session 8 (the paid-tier concept session). Not about AI — about what makes a paid tier worth paying for. Read it before you write your one-paragraph brief.

Anthropic's guide to long-context prompting https://docs.anthropic.com/en/docs/build-with-claude/long-context-tips Relevant for the essay series drafting sessions (Sessions 10–11), where you will be working with longer inputs and outputs than the weekly newsletter. The key principle: put the most important instructions at the beginning and end of a long prompt, not buried in the middle.

9. Closing / Signals from the Keeper

The work in this plan is not about learning AI. It is about encoding what you already know — four years of professional writing, a voice guide you maintain, a newsletter you ship every week — into a form that a model can follow reliably. That is a translation project, not a technology project.

The prompt library is the foundation. Build it carefully in the first two weeks, test it under real deadline pressure, and revise it when it drifts. By the time you reach the essay series in Weeks 3 and 4, you will have a set of tools that know your voice well enough to be genuinely useful on longer, harder work.

The signal that the plan is working is not speed. It is that you stop rewriting for voice and start editing for substance. That shift — from structural repair to genuine revision — is what the 90-minute target is really measuring.

The path is clear. The first session starts with your Notion brand voice guide and a willingness to make the vague concrete. That is the whole thing.