What I Am Actually Running

Before I explain the workflow, here is the honest picture of what I am managing right now as a solo operator:

Bookkeeping and tax services 

— Get Online Bookkeeping (getonlinebk.ca), serving small businesses and newcomers in Calgary. CRA authorized representative. Month-end reconciliation, T1 filings, corporate returns, newcomer tax education.

Education and language consulting 

— TESOL-certified instructor and CELPIP preparation specialist. I develop bilingual curriculum for Mongolian L1 speakers, produce educational resources for newcomers to Canada, and publish language learning materials under the Little Nomad and Learning with Minjee brands on Amazon KDP.

KDP publishing 

— Self-publishing bilingual and heritage language books. Currently building out a 45+ book series for young heritage language learners. Each book is a small product launch with its own research, listing metadata, and promotion.

Calgary Mongolian Community Association (CMCA) 

— I am one of 75 founding members of this non-profit, incorporated in March 2026. The community website launches in June. There is real work here: content production, volunteer coordination, bilingual communications.

Four active areas. One person. A blog that is supposed to document all of it.  

The Problem I Am Trying to Solve

The problem is not ideas. 

I have a running list of things worth writing about — a CRA filing edge case I worked through with a client, a gap I found in CELPIP prep materials for Mongolian speakers, a KDP keyword research decision I made and want to remember, something I learned at a CMCA planning session.

The problem is execution time. 

Getting from "this is worth writing about" to "this is published" takes me two to three hours when I do it manually. Across four venture areas, at one post each per week, that is eight to twelve hours a week I do not have.

So I mapped the workflow. Every step from topic idea to published post. And then I asked: which of these steps actually require me, and which ones just require good information applied by something that can write quickly?

The Workflow I Designed

Here is what I am implementing. I will report back on what actually happens.

Step 1 — Monday topic planning (target: 15 minutes)

Each Monday I run a single prompt giving Claude my four venture areas, any timely angles for the week — a filing deadline, a new KDP launch, a CMCA update, a curriculum resource I just finished — and my audience context: newcomers to Canada, small business owners, Mongolian-Canadian community members, heritage language learners, and people preparing for Canadian language tests.

The output is a topic shortlist of six to eight options with a brief rationale for each. I pick four — one per venture area. That is my content plan for the week, made in one sitting.

What I keep for myself: the final selection. I need to choose topics I can speak to with genuine authority that week. AI cannot know which one I actually have useful things to say about right now.

Step 2 — Drafting (target: automated, 10 minutes of prompting)

Once I approve a topic, I run a drafting prompt. It produces a full post — working title, hook, body sections with subheadings, a practical takeaway, and a call to action relevant to that venture area.

I have fed Claude samples of my writing, my sentence patterns, and my tendency toward the direct and practical over the academic. The drafts should sound like me. Whether they actually do at scale is one of the things I am testing.

Step 3 — SEO metadata (target: 2 minutes, fully automated)

Meta title, meta description, URL slug, and tags — generated in the same session as the draft. These go directly into Odoo's SEO fields. I have been skipping this step manually because it felt like extra work after finishing the writing. Folding it into the AI run removes that friction entirely.

Step 4 — Cover image direction (target: 3 minutes)

Instead of hunting stock photos, I ask for a brief describing the ideal cover image — mood, visual angle, colour palette — and use that brief to search Unsplash or generate something. The brief makes the search fast and purposeful.

Step 5 — Review and edit (target: 20 to 30 minutes, always mine)

I read every post. I check facts. I fix anything that does not sound like me. I add the specific detail only I would know: the actual client scenario, the real number, the community context that makes the post genuinely useful rather than plausibly useful. This step does not get automated. Ever.

Step 6 — Publishing in Odoo (target: 10 minutes per post)

Paste the content, apply heading structure, set the blog category, upload the cover image, paste SEO metadata, schedule or publish. I plan to schedule all four posts at the start of the week so they go out spaced across the days without requiring me to revisit them.

Step 7 — Promotion (target: 4 minutes per post, fully automated)

Platform-specific captions in English and Mongolian for Facebook, Instagram, and my community channels. A newsletter excerpt. Each post gets a full promotion package generated in one prompt run.

The Actual Prompts I Am Starting With

These are the working prompts as of today. They will probably change.

Topic planning:

"Here are my four venture areas this week: bookkeeping and tax services, education and language consulting, KDP publishing, CMCA community updates. Timely angles this week: [insert]. My audience: newcomers to Canada, small business owners, Mongolian-Canadian community, heritage language learners, CELPIP candidates. Suggest six topic options across these areas with a one-line rationale each."

Drafting:

"Write a blog post for my 'Building in Public' section on [approved topic]. Voice: direct, practical, first-person. No jargon unless it serves the reader. Include a hook, three to four body sections with subheadings, a real-scenario example where possible, a practical takeaway, and a CTA pointing to [relevant service or resource]. Length: 600 to 900 words."

Metadata:

"Based on this post, write: a meta title under 60 characters, a meta description under 155 characters, a URL slug, and five relevant tags."

Promotion:

"Write social captions for this post in English and Mongolian: Facebook (100 to 150 words, community tone), Instagram (60 to 80 words with hashtags), and a 60-word newsletter excerpt."

What I Am Watching For

I have not run this at scale yet. Based on what I know about my own tendencies — and the known failure modes of AI-assisted content workflows — here is what I am genuinely uncertain about:

Voice drift. 

AI drafts can start sounding smooth in a way that loses the rough-edged specificity of real experience. If the posts start reading like they could have been written by anyone, I will know something has gone wrong.

Topic repetition. 

Without explicit guardrails, topic suggestions tend to circle back to familiar ground. I will need to feed the system a running list of recent topics to prevent retreads.

The education and curriculum posts specifically. 

Bookkeeping and KDP posts can be reasonably general and still be useful. CELPIP prep and heritage language curriculum content requires genuine subject-matter precision. The drafts will need heavier editing in those areas, and I am curious whether AI can hold the specificity or whether I will essentially be rewriting from scratch.

Consistency across weeks. 

The real test is week four and week eight, not week one. Systems that feel manageable at launch often quietly collapse under the pressure of everything else happening at once. I am going to document that honestly if it happens.

Why I Am Documenting This Here

I could just implement this quietly and report back in three months with polished results. That is not what this section is for.

The "Building in Public" premise is that the interesting part is the process — the decisions made under uncertainty, the things that did not work, the adjustments made in real time. A case study written after the fact smooths all of that out. A case study written as it happens keeps it visible.

So here is the plan: I post the workflow today (this post). In two weeks I post an honest update — what ran as designed, what broke, what I changed. At the end of month one I do a fuller review with whatever data I have.

If this works, you will see it working in real time. If it does not, you will see that too.

What Comes Next

Next post in this series: the first week's content in review. Did the drafts sound like me? Did the Monday planning session actually take 15 minutes? Did anything get skipped?

I am also planning separate posts on what the education and language consulting content looks like specifically — the CELPIP prep angle and the heritage language curriculum angle are different enough from the bookkeeping and KDP posts that they probably deserve their own workflow notes.

If you are also running multiple ventures solo and trying to figure out how to maintain any kind of content presence, follow this series. I will tell you what I actually find.

How This Post Was Made

This post was produced collaboratively with Claude (Anthropic), which served as a structural thinking partner, workflow architect, and drafting collaborator throughout this project. The process worked as follows:

The workflow design itself — mapping the seven phases, identifying what requires human judgment versus what can be delegated — was developed through a working session with Claude in which we built and critiqued the framework together, including producing a visual diagram of the automation breakdown. Claude also produced the initial draft of this article, which I then reviewed, corrected for factual accuracy (venture descriptions, CMCA membership), reframed from retrospective to real-time case study, and edited for voice and precision.

The editorial judgment, factual content, venture context, and the decision to document this in real time rather than in hindsight are mine. The structural analysis and first-draft production are Claude's. Both contributions are real; neither is sufficient without the other.

I will maintain this disclosure standard across all posts in this series.

Collaboration role: Claude (Anthropic) — workflow design, structural critique, drafting partner, editorial iteration Human role: Minjee Ganbaatar — factual authority, voice, editorial judgment, corrections, framing decisions

References and Source Materials

Primary source — workflow developed in this project:

Ganbaatar, M., & Claude (Anthropic). (2026, May 22). AI-powered blog publishing workflow: Automation map and phase analysis [Working session transcript and workflow diagram]. Produced in claude.ai. This article documents the implementation of that workflow beginning on the same date.

Internal reference document:

Ganbaatar, M. (2026, May). Creating AI agent flow [Internal planning document]. Google Drive, personal workspace. This document, developed in May 2026, outlines the conceptual architecture for building AI agents across three operational roles — research, content, and operations — and informed the content agent framework underlying the blog workflow described in this article. The content agent concept (ideation through scheduling, quality gates, human final pass) is directly reflected in Steps 1 through 4 of the workflow above.

Platform referenced:

Odoo S.A. (2026). Odoo Website: Blog module [Software]. https://www.odoo.com. The blog publishing platform used for scheduling, SEO metadata fields, category management, and post formatting described in Step 6.

Tools referenced:

Anthropic. (2026). Claude (Sonnet 4.6) [Large language model]. https://claude.ai. Used as the AI drafting and workflow partner throughout this project.

Unsplash (2026). Free stock photography platform. https://unsplash.com. Referenced in Step 4 as one option for cover image sourcing.