The Last-Twenty Audit

To automate your content without losing your voice, you first need to know exactly what is repetitive and what is unique. Jim Louderback, former CEO of VidCon, suggests the "last-twenty" test. This is an audit of your most recent history to separate manual labor from creative judgment.

For example, look at your last twenty YouTube videos and Instagram carousels. You might find that while the core story of each video was a unique creative decision, the process of gathering links for the description, sorting viewer comments into categories, and drafting basic captions felt like manual labor. These are repetitive tasks.

By reviewing these twenty pieces, you can identify a pattern of work that doesn't actually require your personal taste or lived experience to execute. You aren't looking for what an AI could do in theory, but what you actually did that felt like a chore. The result of this audit is two piles: a "no pile" for high-judgment work that stays with you, and a "yes pile" of repetitive preparation tasks you can delegate to a bounded AI agent.

The "no pile" contains high-judgment work that must remain with you. This work is built on your lived experience, personal taste, and unique perspective—things an AI cannot have because it has no physical presence in the world. This is where you provide the "reason to trust" that your audience follows you for.

Consider a video about a cliff walk. An AI can find the official distance of the trail or the hours the park is open. However, the specific detail that your nephew sat down in exhaustion on the 200th step is a "no pile" task. The AI wasn't there; it didn't see the tired child or feel the incline of the path. That observation is an original fact that only you can provide.

When auditing your last twenty posts, move any task to the no pile if it requires a first-hand account, a subjective opinion based on taste, or a final decision on whether a claim is honest. Your goal is to protect these moments, as they are the only parts of your content that cannot be replicated by a machine.

The "yes pile" consists of repetitive preparation tasks that can be delegated to a bounded AI agent. A bounded agent is a proactive AI tool—like Dots or Claude Duet—that works toward a specific goal under a strict set of limits, rather than just answering one-off prompts. These tasks are ideal for delegation because they involve sorting, organizing, or summarizing information rather than inventing new ideas or sharing lived experiences.

For example, imagine you have forty viewer questions spread across TikTok, Facebook, and Instagram. Manually reading every comment to find common themes is a chore. Sorting these questions into groups—such as "trail difficulty," "parking tips," and "shuttle times"—is a classic yes pile task. The agent doesn't need to know your personal taste to group these; it only needs to recognize keywords and patterns.

To build your yes pile, look at your audit and identify every step that feels like "legwork." This includes gathering sources, creating initial draft lists, and formatting content for different platforms. By moving these to the yes pile, you clear your schedule to focus entirely on the high-judgment creative work in your no pile.

Once you have a 'yes pile,' you can delegate the most tedious part of content prep: the initial research. Instead of asking an agent to simply 'find information,' you task it with creating a research packet. A research packet is a structured document where every single claim is paired with a direct source, a link, and the date the information was accessed. This removes the risk of the AI 'hallucinating' facts or providing outdated data without you knowing it.

For example, if you are prepping a video about a park visit, you don't just ask for the shuttle times. You instruct your agent to find the current timetable and present it as: "Last shuttle departs at 5:30 PM [Link to Park Timetable Page, Accessed Oct 12, 2025]." By requiring this format, the agent isn't making a decision; it is acting as a digital filing clerk.

You then take this packet and check it against your no pile. You don't trust the agent's summary—you click the link to verify the claim yourself. This allows the agent to do the heavy lifting of searching while you maintain total editorial control over the facts.

There is a critical boundary between generating a draft and making an editorial decision. To maintain your authenticity, you must treat every agent-produced text as a suggestion, not a finished product. In this workflow, the agent is responsible for the 'first draft'—the act of putting words on the page—while you are responsible for 'deciding' which words are true to your voice and accurate to the facts.

For example, imagine you are prepping a Reel about your cliff walk. Instead of asking the agent for 'the perfect caption,' you task it with drafting three distinct versions based on your research packet: one that is helpful and direct, one that is curious, and one that is slightly more storytelling-focused. The agent does the labor of phrasing and formatting. However, you then step in to review them. You might find that version two captures your natural tone, but version three has a better hook. You combine the two, correct a phrasing error that sounds too 'AI,' and verify that the shuttle time is still correct. By separating drafting from deciding, you use the AI to beat the blank page without letting it dictate your perspective.

To maintain audience trust, you must set a hard boundary around your Name, Image, and Likeness (NIL). Your NIL is the set of unique identifiers—your face, your specific voice, and your personal identity—that signal to a viewer that the content is actually from you. While a bounded agent is excellent at the legwork of research and drafting, it must never be given the authority to simulate your physical presence.

For example, you may allow your agent to write a script for a YouTube Short based on a research packet. This is a text-based task. However, you must explicitly forbid the agent from using AI voice-cloning tools to generate the audio or using image generators to create a synthetic version of your face. If an agent can 'speak' or 'appear' as you without your direct physical involvement, you have moved from using a tool to creating a digital twin, which risks eroding the authenticity your audience values.

When configuring your agent's boundaries, list your NIL assets—your voice, your face, and your signature sign-offs—and mark them as strictly off-limits. The agent may prepare the map, but you are the only one allowed to be the guide.

The final step in scoping your agent is to map your new workflow. A bounded workflow is a sequence of tasks where preparation is handled by the agent, but final approval is gated by the human creator. This ensures that nothing goes live without your eyes on it, effectively solving the fear that an agent might 'decide' things you didn't approve. Instead of the agent having a direct line to your publishing buttons, it has a direct line to your review desk.

For example, in your weekly content cycle, the process now flows linearly: first, the agent gathers viewer questions; second, it builds a sourced research packet; third, it produces first drafts of shot lists and captions. Only after these three steps are complete does the work reach you. You review the research for accuracy, refine the captions to match your voice, and approve the shot list. You are no longer the person doing the filing; you are the showrunner signing off on the production. To implement this, write down your production sequence and mark exactly where the agent's authority ends and your approval begins.