The fastest way to ruin your LinkedIn presence is to point an AI agent at a generic prompt and tell it to post five times a week. The fastest way to keep your LinkedIn presence is to point an AI agent at a real source of truth (your work, your notes, your shipped product) and let it draft, not publish.
This walkthrough covers a setup that has run for several founders without making them sound like a generic LinkedIn growth account. The agent drafts from concrete material, you approve in a 10-minute morning batch, and the queue does the rest. After 30 days you have the option to graduate one or two post types to auto-publish; most people never bother.
What this agent does
The agent reads three things every morning: your private notes for the past 48 hours, the latest entries from a designated Notion or Drive folder, and any newly shipped public artefacts (a commit, a blog post, a release note). From those, it picks one item per scheduled slot and drafts a post in your pillar style.
The output is a draft, queued in a "review" board you open once a day. The agent does not brainstorm topics out of thin air, scrape competitors, or "ideate". It works from your material because that is the only material that sounds like you.
For the broader pattern, see what an AI agent can actually do. For the related daily rollup pattern, see AI agent for Notion daily rollup.
Where it pulls from
Source quality determines post quality. Wire these in priority order:
- Working notes. A daily Notion or Obsidian file you write to as you work. The agent reads the past 48 hours.
- Shipped output. Commits, deploys, blog posts, customer emails. Anything that proves you did the work.
- Customer threads. Support, sales, onboarding feedback. Redacted before the agent reads it.
- Reading list. Articles you marked as "interesting" with a one-line reaction. The agent does not opine on articles it found on its own.
Things the agent must not use as sources: trending topics, generic prompt templates, previous LinkedIn posts that did well for someone else. Those produce average posts. The audience can tell.
Content pillars
Two or three pillars keep the feed coherent. For a bootstrapped founder building Gravity, the pillars look like this:
- Building in public. What I shipped, what I learned, what broke.
- Customer stories. Anonymised, real, with permission. The agent flags any that are too specific to ship without explicit consent.
- Decisions, not opinions. "We chose X over Y because Z" beats "X is the future". The agent has a hard rule against takes that are not backed by something we shipped.
Each pillar gets a rotation slot. If the agent drafts three posts per week, that is one of each pillar. The agent rotates strictly so the feed does not collapse to whichever pillar happens to have the most material that week.
Draft, approve, schedule
The first 30 days the agent runs in draft mode only. Each morning at 09:00 IST you open the review board, accept or edit each draft, and click schedule. The pipeline:
- 06:00 IST. Agent reads sources, drafts the day's post, places it in the review board.
- 09:00 IST. You spend 5 to 10 minutes editing or rejecting. Edited drafts go to the scheduling queue.
- 11:00 IST. Scheduler publishes via the LinkedIn API or your scheduling platform.
Edits matter. The first week, expect to edit every line. By week three, you will be editing a phrase or two per post. That delta is the calibration data; capture it (see voice calibration below).
The framing for graduating actions like this from draft to auto is in how to limit agent actions.
Cadence and weekly cap
Three posts per week is a defensible default. Five is the maximum before the feed feels like work. Daily posting is rarely correct unless you genuinely have a daily moment worth sharing.
Configure these limits in the prompt, not in your head:
- Weekly cap. Hard cap at 5 posts per week. The agent skips days if quota is met.
- Minimum gap. 24 hours between posts. No same-day double-posting.
- Skip when stale. If no new source material has appeared since the last post, the agent does not draft. Empty days are fine.
- Pause on personal events. If your calendar shows a flight, a funeral, or a launch, the agent pauses for that day.
For more on agent action boundaries, see how to restrict agent to business hours.
Voice calibration
The single biggest reason these setups fail is voice drift. The first week, the agent sounds like you because it is mostly your raw notes with light formatting. Week three onwards, the agent starts inserting safe LinkedIn phrasing (rhetorical questions, "the lesson is", lists of three nouns, exclamation marks). The audience checks out.
The fix is monthly. Once a month, take ten of your edited posts and add them to the prompt as examples of the final voice. Add three rules the agent must follow (yours will differ; mine are: no rhetorical questions, no "lessons learned", no claims that are not backed by something I shipped). Refresh the example set every 30 days; old examples grow stale because your voice grows.
For the broader idea of feeding edits back as training data, see how to train agent on company docs.
Common mistakes
- Generic prompt. "Write a thoughtful LinkedIn post about leadership" produces a post nobody reads.
- Auto-publish from day one. The first weeks are calibration, not throughput.
- No source material. If you have not done anything worth posting about, do not post.
- Five pillars instead of two. Too many pillars dilute the feed and confuse the rotation.
- Skip the monthly voice calibration. Voice drift is silent until week six, when reach drops and you wonder why.
Frequently asked questions
Can an AI agent post on LinkedIn for me automatically?
Technically yes through the LinkedIn API or the platform UI, but most founders should keep the agent in draft mode for the first 30 days. The agent prepares each post from your source material, you approve in a 10-minute morning batch, and the queue handles publishing. Auto-publish becomes safe only after the agent's voice has been calibrated against your real writing.
How does the agent know what to write about?
It reads from sources you wire up: a Notion or Google Doc with your weekly notes, your latest blog posts, your shipped commits, your customer support log. It does not invent topics. The prompt names two or three content pillars (for a founder typically: building in public, customer stories, technical decisions), and the agent picks one pillar per post in rotation.
How often should the agent post?
Three to five times per week works for most founders. Daily posting often produces filler that hurts engagement and reach. The agent should be configured with a hard weekly cap, a minimum gap between posts (24 hours), and skip days when no source material has changed since the last post.
Will the audience notice posts are AI-drafted?
If the agent posts without your edits, yes, immediately. The fix is the approval layer. You spend 90 seconds per draft tightening the opening line, swapping a generic phrase for a specific one, and trimming. Readers do not detect AI-drafted posts that you have edited; they detect AI-drafted posts that nobody read before publishing.
What is the biggest failure mode of a LinkedIn content agent?
Voice drift. The first week is fine, the second week starts to sound generic, by week four the posts read like every other AI-drafted thread. The fix is to feed back your edited versions into the prompt as examples, refresh the example set monthly, and set a tone rule the agent must respect (no rhetorical questions as openers, no list of three nouns, no claims you cannot back up).
Three takeaways before you close this tab
- Real sources, not topic prompts. Your notes, your commits, your customers. Nothing else.
- Draft, approve, schedule. 10 minutes a day for the first 30 days. Then decide whether to graduate.
- Voice calibration is monthly. Feed edits back. Refresh the example set. Or watch reach decay.
Sources
- LinkedIn Engineering, "LinkedIn Posts API Reference", retrieved 2026-05-10, learn.microsoft.com/linkedin/posts-api
- LinkedIn Marketing Solutions, "How LinkedIn's feed ranks content", retrieved 2026-05-10, linkedin.com/business/marketing/blog
- Aryan Agarwal, "Gravity content guardrails", internal v1, May 2026, About