Case Study · Built for internal use
One dashboard.
Your entire operation.
Built for internal use. A custom AI-native operating system that runs a solo advisory and app development practice. It connects ClickUp, Gmail and Google Workspace, Slack, GitHub, Buffer, Kit, Plausible and Claude in one dashboard, and handles the repetitive parts of the work.
Situation
Northlight OS exists because no off-the-shelf tool connected ClickUp, Gmail, Google Workspace, Slack and Claude the way a one-person practice works, so I built one.
Running a solo advisory and app development practice means constant context-switching. Tasks live in ClickUp. Client emails arrive in Gmail. Finances are scattered across bank statements and email receipts. Content ideas sit in Slack. No off-the-shelf tool connected all of that in a way that matched how a one-person practice works.
The options were to pay for several separate tools and check them all by hand, or to build the right tool. I built it.
Objective
The goal was one place to see and run the practice, with repetitive work done overnight and a person approving everything that matters.
One place to see and run the whole practice, with the repetitive work done overnight and a person approving everything that matters. Four jobs in particular:
Overnight execution
Tasks assigned to Claude in ClickUp are drafted or researched by morning, ready to review.
Finance automation
Expense and revenue emails are parsed so entries can be logged in one click.
Content pipeline
An idea becomes a brief, a task and a tracked piece of content.
Idea to action
Ideas posted to Slack are assessed with a clear verdict: go, needs more thought, or park longer.
How it works
01
Overnight Execution Loop
Assign a task to Claude in ClickUp. By morning, it's done, research compiled, copy drafted, brief written. Review in the dashboard: Approve, Send Back with notes, or Take Over.
02
Finance Automation
Gmail labels trigger AI parsing. Vendor names and amounts are extracted from email receipts automatically. One click logs an approved expense or revenue entry to Google Sheets.
03
Content Pipeline
A chat-based intake interview with Claude captures the raw idea, audience, platform, and goal. Claude builds the brief, creates the ClickUp task, and tracks it through a 12-stage pipeline from draft to posted.
04
Idea to Action
Ideas posted to a Slack channel flow into the dashboard automatically. Claude writes a brief and runs a critic pass, returning a clear verdict: GO, Needs More Thought, or Park Longer.
Also in the dashboard
The dashboard has seven areas: Home, Focus, Marketing, Finance, Compass, Capture and System.
The four workflows above are the headline features. The dashboard itself is organised into seven areas, and each one does a job of its own.
Home
A morning command centre: what Claude finished overnight, what needs attention now, today's calendar, project health and the inbox. Goals, a wins board and an automatic weekly review every Friday also live here.
Focus
Client work, a learning log and projects, in one place for the work that needs sustained attention.
Marketing
The content pipeline from strategy to publication, a studio, operations, analytics and the marketing round table.
Finance
Revenue logged from Gmail, expenses queued for approval, and a monthly profit and loss report on the first of each month.
Compass
A daily news feed filtered for AI, marketing, advisory work and Canadian small business, the advisory board, and a parking lot for ideas that are not ready to become tasks.
Capture
A quick way to turn a thought into a properly formed ClickUp task.
System
Controls for the daily dispatcher, automations, settings and maintenance.
The advisory board and the marketing round table
The advisory board and the marketing round table are two groups of AI personas that debate a question, while a secretary who never debates writes up the decision and saves it to Google Drive.
Two groups of AI personas sit inside the dashboard. I give each group a question, they present their views and debate, and a secretary who never takes part in the debate writes up the decision. The record is saved to Google Drive. Each persona sees only the information relevant to its role, and each is built to challenge the idea, never the person holding it.
Business decisions
A CFO, a Chief Privacy Officer, a CMO, a Founder Coach, a Technical and Product lead and a Strategy lead. The Founder Coach sees overdue and stale tasks to spot avoidance patterns, and the Technical and Product lead sees the open product backlogs.
Content and marketing
A Strategist, a Drafter, a Critic, a CMO, an SEO agent and an Analytics and GEO agent. Together they work a piece of content from brief to draft to critique. The SEO agent is still a placeholder until it has a live data feed.
Decisions and the reasoning
Build the right tool
Connecting six platforms the way a one-person practice actually works was not something I could buy, so I built it around my own workflow.
A person approves everything
Every overnight result waits for review with three choices: approve, send back with notes, or take over. Nothing goes out on its own.
Route writes through one endpoint
When the ClickUp connector proved unreliable under load, I moved all task updates to a single batch endpoint in the dashboard and kept the connector for reading only.
Integrations
Northlight OS connects to sixteen tools, including ClickUp, Gmail, Google Workspace, Slack, GitHub, Buffer, Kit and Claude.
Northlight OS connects to sixteen tools, and all of them are part of how the system runs.
Product names and logos belong to their owners. They are shown only to identify the tools Northlight OS works with.
System architecture
AI skills framework
Northlight OS runs more than twenty reusable Claude skills across content, operations and finance, and they can be chained together or run on their own.
Northlight OS runs more than twenty reusable Claude skills, modular agents that can be chained together or run independently. Tasks in ClickUp are assigned to a skill, and the system handles the rest. All of them are listed here.
Content
Operations
Finance
Built with
Results
The system has run daily since May 2026, handling overnight drafts and research, expense and revenue parsing, a morning briefing, and the two AI boards.
The system has run daily since May 2026. It handles overnight first drafts and research, parses expense and revenue emails, produces a morning briefing, and runs the advisory board and the marketing round table. It has also failed in instructive ways, and most of what I learned came from those failures.
What I learned
The biggest lesson is to check the output, not the status: a pipeline once reported 12 posts as ready when 10 of them had no copy.
Check the output, not the status
The content pipeline once reported 12 posts as ready. On inspection, 10 of them had no copy, only refusal messages. A status that says done tells you the process ran, not that it worked.
Automation exposes the quality of the process
I automated first drafts, and the drafts did not sound like me. I paused drafting and moved to a stories-first approach, where the ideas come from things that really happened.
A check that catches the same problem every time is a warning
In one review, 5 of 6 debate rounds were flagged for invented context. The automatic regeneration kept producing a different wrong answer. The real cause was a prompt with no instruction for what to do when information is missing.
Silent failures are the expensive ones
A three-line helper that started tasks without waiting for them was hiding failures across the pipeline. Finding it meant reading the code for the symptom, not sampling at random.
Working and usable are different bars
A feature that worked end to end still failed the simple test of whether I could find what it produced. I now ask both questions.
Verify against the live system, not the documentation
Buffer's documentation and several searches disagreed about what its API could return. Calling it settled the question in a minute.
How this connects to my client work
When I help a business adopt automation or AI, I start with the process and not the tool: where work is actually getting stuck, who approves what, and how we will know the output is good. Building and running my own system is where I test those questions first.