Run a Content Channel with AI: A One-Person Operation That Ships Every Week

The goal

When you are done, you should have: an account that publishes for three straight months without a gap. Concretely, one or two issues a week, topics you never have to invent on the spot, a fixed writing process, a consistent visual template, two finished issues always in the bank, and a one-hour weekly review that feeds results back into the topic list.

Note that the goal says nothing about follower count. Followers depend on the niche, the platform's recommendation system and luck — no process guarantees them. Publishing consistently is the only thing a process can guarantee, and it is the precondition for everything else. Most accounts die between week three and week six, and the cause is almost never bad content. It is running out of things to write.

This path covers text and talking-head personal accounts: newsletters, long-form platforms, lifestyle platforms, industry-vertical accounts. It does not cover three things — paid promotion and sponsorship deals, multi-account syndication at scale (which most platforms classify as abuse), and daily news operations that require a team.

The tool stack

The weight of this stack belongs on topic input and scheduling, not on writing faster. Writing faster does not prevent a missed week.

Figure 1: The chain is a loop, not a line. The review must feed back into the topic bank, or you start every week from zero — which is the most direct cause of missed weeks.

Why this stack

Why the money and time go into the topic layer. Running an account alone, the time to write one piece is predictable. Finding a piece worth writing is not. Search-based AI earns its place here by telling you what people are actually searching for and how bad the existing answers are, which beats staring at a blank document. With 30 topics banked, your weekly job is choosing; with an empty bank, your weekly job starts from nothing, and two weeks of that ends the account.

Why structure and voice use different models. This isn't superstition. Models differ visibly in long-form structure, argument order and information density, and they differ even more in prose rhythm and how naturally they use idiom. Getting structure from one and the rewrite from another is usually faster than endlessly tuning prompts against a single model. Which one suits you is worth testing yourself against Which AI Is Actually Good in 2026: Six Models Tested.

Why a human must do the final pass. Readers detect AI phrasing far better than they did two years ago: too much parallel construction, a summary at the end of every paragraph, the inevitable uplifting close. This is not a style problem, it is a trust problem — once readers conclude they are reading machine output at scale, unfollowing is a quick decision. The human pass is not polishing. It is inserting the things only you can supply: the mistakes you made, the numbers you saw, your judgement, and the parts you are unsure about.

Why the bank is mandatory. Illness, travel, overtime and bad weeks will happen. With two issues banked, they are interruptions. Without a bank, they are day one of a gap — and after the first gap, the second one comes much more easily.

The full steps

Step 1: Positioning and columns

Write one sentence: who reads this, what they get, and what's different about it. If you cannot write the "different" part, don't launch yet — go read twenty comparable accounts first.

Then fix two or three recurring columns, something like "this week's mistake," "tool test," "reader questions." Recurring columns do three things: they give topics a direction, readers an expectation, and you a template.

Step 2: Build the topic bank

Collect 30 in one sitting. Sources in order of reliability:

1. Real search demand: use search-based AI to find what people look for and what the existing answers are missing.
2. Comments and messages: questions readers asked in their own words convert best.
3. Mistakes you made yourself: irreplaceable and impossible to copy.
4. Gaps in other people's content: not copying their topics — finding the point where they stopped halfway.

Record four fields per topic: title, who it's for, the reader's actual question, and what exclusive material you hold. Topics with an empty fourth field go to the back of the queue.

Step 3: Run the per-issue pipeline

Seven fixed actions: topic → outline → draft → human rewrite → visuals → headline and cover → publish. Every action ends in a deliverable. Do not run the whole thing inside one long chat — you lose track of progress and cannot resume after an interruption.

The draft stage can lean heavily on AI. The rewrite stage has to be you, with AI assisting.

Step 4: Write a "human voice" checklist

This is your own writing rulebook, with at least three parts:

• Banned words and constructions: list the AI clichés you personally hate and search for them before publishing.
• Mandatory elements: every piece contains at least one first-hand experience, one specific number or case, and one honest "I'm not sure."
• Tone baseline: three sentences that sound like you, three that don't, as a reference during rewriting.

Step 5: Visual consistency

Use a fixed cover template and change only headline and colour. Prefer your own screenshots and photos inside the piece. The value of a template is not beauty, it's recognisability — readers should identify you in a crowded feed at a glance.

Write three headline variants and publish one. Bank the other two for the next similar topic.

Step 6: Schedule and bank

Keep three states in the calendar: to write, written and queued, published. One rule: if the bank drops below two issues, stop starting new topics and refill it. The rule is almost insultingly simple, and it is the single most effective thing in this whole process.

Step 7: Publish and the first hour

Fix the publishing time so readers develop a habit. Reserve one hour after publishing for replies — early engagement affects distribution on most platforms, and it is your only direct contact with readers. Those replies are also next week's topics.

Step 8: The weekly review

Export the analytics and look at three rates: open rate (a headline and cover problem), read-through rate (a structure and length problem), and engagement rate (a topic and angle problem). Do not look only at likes — likes are the outcome, those three are the causes.

The output of the review has to land back in the topic bank: give the column that performs well a bigger quota, and record the headline patterns that open well.

Figure 2: The per-issue pipeline with the bank gate. Step G looks redundant, and it is the only part of the anti-gap mechanism that actually works.

Prompts

1. Build the topic bank

You are a content strategist. Help me expand the topic bank for the account below.

Positioning: {one sentence: who reads it, what they get}
Existing columns: {list 2–3 recurring columns}
My exclusive material: {your work history, data you hold, mistakes you made — be specific}

Produce 20 topics. Requirements:
1. Each maps to a concrete question a reader would actually search for. No
   "five tips for X" shells.
2. Annotate each with: who it's for / the reader's real question / what exclusive
   material I need before it's writable.
3. Explicitly mark the topics I should NOT write without my own material — the
   internet already has plenty of those.
4. Sort from "where I have the strongest advantage" to "anyone could write this."

No generic industry trend pieces.

2. Write the outline

Write an outline for the topic below. Order it by the reader's questions, not by the
logical order of the knowledge.

Topic: {title}
Reader: {who}
Where they start: {what they already know, what they wrongly believe}
What they should be able to do afterwards: {one concrete action}

Output:
1. The opening two sentences: answer the reader's most urgent question immediately.
   No background throat-clearing.
2. Three to five sections, each annotated with "which reader question this answers"
   and "what material answers it."
3. Mark where I need to insert personal experience or a specific case.
4. Ending: one concrete action. No uplift.

Target length {word count}. Outline only, no body text.

3. Rewrite for a human voice

The text below is an AI first draft. Give me edits that remove the model's voice.
Give suggestions and replacement sentences only — do not rewrite whole sections.

What to look for:
1. Every parallel construction, balanced pair, and cliché opener. Give a replacement
   for each.
2. Redundant summary sentences at the end of paragraphs. Mark the deletable ones.
3. Adjectives and adverbs carrying no information. Flag them.
4. Places where personal experience, a specific number, or an honest "I'm not sure"
   should go — tell me what to add, but do not invent the content for me.
5. Paragraphs whose information density is too low to keep.

Output a table: original / problem / suggested replacement / or just delete.

Text:
{paste the draft}

4. The weekly review

Below are this week's analytics and headlines. Do a review.

Rules:
1. Speak only from the data. No generic content-marketing advice.
2. Give possible causes separately for open rate, read-through rate and engagement
   rate. Every cause must point at a specific piece and a specific fixable action.
3. Identify this week's outlier (clearly high or clearly low) and hypothesise why.
4. Output three concrete actions for next week, each doable in under two hours.
5. Where the sample is too small to support a conclusion, say "sample too small"
   rather than inventing a cause.

Data:
{paste headlines, publish times, impressions, open rate, read-through, engagement}

Worked example: feeding a textbook AI draft into the voice rewrite

Test setup: macOS with the Claude Code CLI 2.1.226, --model sonnet, August 2026. The draft below was written deliberately as a typical AI first pass; the output was actually run.

Before you start: install the CLI

The demos use the Claude Code command line. Three commands to install and verify:

# 1. Node.js 22 or newer is required
node --version

# 2. Install globally
npm install -g @anthropic-ai/claude-code

# 3. Verify
claude --version        # should print something like 2.1.226 (Claude Code)
claude doctor           # checks that the installation is healthy

The first run needs a login: type claude in a terminal to open the interactive interface and follow the prompts to authorise your account (claude auth manages the login state afterwards). Once that is done, the one-shot -p calls below run directly.

Every command here passes --model sonnet so that you reproduce against the same model tier used for these runs; without it you get your account's default model and the output will differ.

Step one: create the material (copy and paste)

First, a paragraph that reads smoothly and says nothing — the kind of thing that ships when the human rewrite gets skipped:

mkdir -p ~/demo/channel && cd ~/demo/channel

cat > draft.txt <<'EOF'
In today's era of information explosion, how to efficiently manage personal knowledge has become an
important issue that every professional must face. With the continuous development and maturation of
artificial intelligence technology, more and more tools are entering our field of vision. These tools
not only help us quickly organize massive amounts of information, but also significantly improve our
work efficiency. First, we need to clarify our core needs. Second, we must choose the right combination
of tools. Finally, we also need to establish long-term sustainable habits. In conclusion, only by
organically combining tools and methods can we truly maximize the value of knowledge management and
stand out in fierce workplace competition.
EOF

cat > rewrite-prompt.txt <<'EOF'
The text below is an AI first draft. Give me edits that remove the model's voice.
Give suggestions and replacement sentences only — do not rewrite whole sections.

What to look for:
1. Every parallel construction, balanced pair, and cliché opener. Give a replacement
   for each.
2. Redundant summary sentences at the end of paragraphs. Mark the deletable ones.
3. Adjectives and adverbs carrying no information. Flag them.
4. Places where personal experience, a specific number, or an honest "I'm not sure"
   should go — tell me what to add, but do not invent the content for me.
5. Paragraphs whose information density is too low to keep.

Output a table: original / problem / suggested replacement / or just delete.
Keep it under 12 rows.
EOF

Step two: run it

{ cat rewrite-prompt.txt; echo; echo "Text:"; cat draft.txt; } > input.txt
claude --model sonnet -p "$(cat input.txt)"

View full-size image

×

Terminal window: entering the demo/channel directory, joining the rewrite prompt and draft into input.txt, calling claude, and printing a table that names each cliché and empty adjective with a replacement

Figure 3: The actual terminal output of the commands in this section. The closing observation — that the passage contains no concrete noun — works better as a checklist item than the ten individual edits above it.

Step three: read the result

Output (12 rows; five of them here):

| Original | Problem | Replacement |
| "In today's era of information explosion" | Cliché opener | Delete; open with the actual claim, e.g. "Professionals now manage more inbound information than they can read." |
| "not only help us quickly organize massive amounts of information, but also significantly improve our work efficiency" | Balanced "not only X but also Y" pair; "quickly," "massive," "significantly" are empty | Cut the pair structure; replace with one concrete claim + a number, e.g. "cut my weekly triage time from X to Y hours" — insert your own real figure |
| "First… Second… Finally…" | Three-part parallel list, textbook AI cadence | Break the parallelism, or replace with a single sentence naming the actual bottleneck you hit |
| "In conclusion, only by organically combining tools and methods…" | Cliché closer + redundant restatement | Delete entire sentence |
| Whole passage | Missing any first-person experience, number, or admitted uncertainty | Add one concrete anecdote (what tool you tried, what failed) and one honest hedge |

Note what it did not do: write the missing content for me. The last row says "insert your own real figure" and "add one concrete anecdote," rather than inventing three months of experience with a named product. That behaviour comes from the line "tell me what to add, but do not invent the content for me" — remove it and you get prose with a convincing personal voice built entirely on fabricated experience, which is considerably more dangerous than sounding like a model.

The other directly usable output is the closing observation: the passage contains no concrete noun. That works better as a checklist item than the ten individual edits — before publishing, scan your own draft and count the tool names, the numbers and the specific situations. If you can't find any, it isn't finished.

What you have to supply

• A one-sentence positioning and two or three column names
• Your real history: what you have done, what you got wrong, what data you hold that others don't
• A list of twenty comparable accounts, used to find gaps rather than to copy
• Export access to your analytics (many reviews stall simply because nobody exported the data)
• A visual template: cover layout, colours, typeface
• A material library: your own screenshots, photos, work-in-progress records
• A fixed publishing time, and one immovable hour each week for the review

What you end up with

• A topic bank of at least 30 entries, refilled by every weekly review
• A seven-step per-issue pipeline where each step has a deliverable
• A voice checklist: banned words, mandatory elements, tone baseline
• A visual template and an archive of unused headline variants
• A calendar where the bank never drops below two issues
• A weekly review record that shows which topics and headlines genuinely work

Time

Estimated for a text account publishing twice a week:

Two issues a week plus the review comes to about five or six hours weekly once practised. Confirm you actually have those hours before you launch. If you don't, drop to one issue a week — a steady weekly cadence beats a high-frequency fortnight followed by silence.

Cost

• Required spend is close to zero: mainstream AI assistants have free tiers, editing tools have free versions, and a spreadsheet works as a calendar. Starting at zero is genuinely viable.
• Efficiency spend: one AI assistant subscription (mainstream personal tiers sit around 20 US dollars a month) plus one editing or design subscription. That is the actual monthly bill for most personal accounts.
• On-demand spend: voiceover, subtitle polish, outsourced covers, licensed stock. These are per-issue costs worth adding only after a format has proven itself.
• The most underestimated cost is your time. At five hours a week, three months is over sixty hours. Answer honestly before launching: where are those sixty hours coming from?
• Where not to spend: one-off "universal prompt packs," courses promising follower growth, and bulk distribution tools. Your own voice checklist beats the first two, and the third violates most platforms' rules.

How it fails

Figure 4: Seven ways an account dies. The first two are the most common, and they reinforce each other — the less you can write, the more you lean on AI, and the more you lean on AI, the fewer people read.

1. Readers spot the AI voice. The symptom is flat numbers you can't diagnose: acceptable open rate, poor read-through, silent comments. The cause is that the piece reads like every other piece on the subject. The only fix is putting in what only you have, and no prompt substitutes for it.

2. The topic bank empties. The most common cause of death, and it happens quietly: one week you have no topic and skip; the next week starting is harder and you skip again; by week three you no longer think of yourself as running the account. The bank gate exists to stop the first skip.

3. Ten pieces that read identically. Ten issues from one outline template tire readers, and platforms flag high content similarity. The fix is a different structure per column plus a deliberate template change every month.

4. Chasing news badly. Two ways to fail: chasing slowly and publishing after the moment passes, or chasing into a field you don't understand and getting it wrong. The test is simple — is this connected to your niche, and do you have independent information? If neither, skip it. Politics, disasters and other people's private lives are areas to leave alone entirely unless you genuinely work in them.

5. Watching likes only. Likes lag and are noisy. The three rates are what guide improvement: open rate points at headline and cover, read-through at structure and length, engagement at topic and angle.

6. Unclear sourcing. Images grabbed from search results and opinions lifted from someone else's article will eventually cause a problem. The rule is hard: only images you shot, generated, or explicitly licensed; attribute anyone else's judgement; never touch material marked as not for reproduction.

7. Launching on five platforms at once. Formats, lengths, cover dimensions and recommendation logic all differ, and cross-posting the same file usually performs badly everywhere. Get one platform working, then adapt rather than repost.

Alternatives

If you write slowly but talk quickly: switch to spoken drafts. Record ten minutes of thinking aloud, transcribe it with Tingwu, then have AI shape it into an article. For many people this is three times faster than facing a blank page, and it carries a human voice naturally.

If weekly is genuinely unsustainable: switch to series production. Spend two concentrated weeks making a six-part series, then release it weekly. To readers it looks like consistent publishing; to you it is one block of time you can actually schedule.

If you only want to make video: skip text entirely and go to Script and Edit Video with AI — the process and cost structure are quite different.

If you have a collaborator: keep topics and reviews for yourself and hand off visuals, editing and scheduling. Those two are the most separable parts and the least tied to personal voice. For team workflow, see Chinese AI Content Automation: An API and n8n Pipeline with a Human in the Loop.

Other tasks in this series: Script and Edit Video with AI, Run Product Research with AI, Build a Knowledge Base with AI.