Configuring the Feedback-Routing Agent

A feedback-routing agent is a proactive AI specialist—built using tools like Dots or Claude Duet—that manages the intake of client comments. Unlike a general assistant, its primary purpose is to gather and flag information, not to make creative decisions. In a high-pressure production, the biggest risk is an automated system or a junior team member treating a client's casual preference as a final directive, which often leads to the "death by revisions" where the original idea is eroded.

To prevent this, you must define a strict boundary: the agent may organize and notify, but it may never approve. For example, if a client emails a request to "make the logo bigger," a correctly configured agent in Claude Duet will not tell the designer to execute the change. Instead, it logs the request, flags it as a potential change to the approved layout, and sends a decision request to you. You use this agent as a filter, ensuring that every change is judged against the brief before a single pixel is moved, keeping custody of the creative vision.

To turn an AI into a reliable specialist, you must provide a remit. A remit is a set of strict instructions that define the agent's expertise, the specific sources it is allowed to monitor, and its absolute limits. Without a remit, an agent behaves like a general-purpose chatbot; with one, it behaves like a production coordinator who knows exactly what to watch and when to stop.

For example, when configuring Dots, you wouldn't simply tell it to "watch for feedback." Instead, you write a precise remit: "Monitor the project email thread labeled 'Project X - Feedback'. Identify and extract every comment specifically directed at the 'Hero Film' asset. Ignore comments about the social banners or budget. When a comment is found, log it in the feedback tracker with the sender's name and timestamp."

By defining these boundaries, you ensure the agent doesn't create noise by flagging irrelevant conversations. You are essentially building a digital filter that protects your team from distraction, allowing you to spend your energy on judging the work rather than hunting for comments across multiple threads.

One of the fastest ways to kill a creative idea is by applying outdated feedback to a new version of an asset. This happens when a client's comment from round two—which was based on an early, flawed draft—is accidentally treated as a current requirement for round four. When comments are disconnected from their specific version, the team spends hours solving problems that have already been fixed, leading to a disjointed final product.

To prevent this, you configure your routing agent to tie every piece of feedback to a specific asset version number. For example, if a client notes that the "lighting is too harsh" while reviewing v02_Hero_Film.mp4, the agent tags that comment specifically to v02. When the team moves to v04_Hero_Film.mp4, where the lighting has been completely corrected, the agent flags the v02 comment as "superseded."

As a creative director, you use this version-locking to audit the feedback stream. Instead of seeing a long, confusing list of requests, you see a chronological map. This allows you to ignore ghost comments from old versions and ensure the team only executes changes that are relevant to the current state of the work.

Incoming feedback is rarely uniform; it is usually a mix of objective errors and subjective opinions. If you treat every comment as an equal directive, you risk shifting the creative direction based on a single stakeholder's whim. To avoid this, you configure your agent in Dots or Claude Duet to classify notes into three primary categories: Corrections, Preferences, and Strategic Shifts.

Consider two different comments on a draft. If a client says, "The product is missing from the frame," the agent labels this as a Correction. It is a factual error that must be fixed to meet the basic requirements of the deliverable. However, if the client says, "I don't like this shade of blue," the agent labels this as a Preference. This is a subjective opinion that may or may not align with the approved brief.

By automating this classification, you can instantly filter your feedback list. Instead of wading through a wall of text, you prioritize all Corrections first to ensure the work is technically sound, then review Preferences to decide which ones are worth implementing and which ones would compromise the core idea. This keeps you in control of the creative judgment.

The most dangerous moment in a feedback cycle is when a stakeholder's personal preference is mistaken for a final creative decision. When this happens, the creative director often feels pressured to 'average' the requests, leading to a version of the work that is technically compliant but conceptually hollow. By using your routing agent to explicitly separate these two, you create a buffer that protects the approved brief.

For example, imagine your approved brief defines the campaign tone as "bright and optimistic" to attract a young audience. A senior stakeholder reviews the latest cut and asks for a "moodier, more cinematic tone." Instead of immediately directing the editor to darken the grade, you look at the agent's log. The agent has already flagged this request as a Preference and attached a warning: "Contradicts Brief Objective: Bright and Optimistic Tone."

As the creative director, you now have the evidence needed to push back. Instead of reworking the direction, you address the underlying strategic question: "Are we shifting the target audience, or is this a personal aesthetic preference?" This shift in framing prevents unnecessary direction shifts and keeps the core idea intact.

The final step in protecting your creative direction is moving from manual sorting to an automated intake workflow. By integrating a proactive agent like Dots or Claude Duet directly into your project management tools (such as Asana, Monday.com, or Jira) and communication apps, you create a continuous, live stream of classified feedback. This removes the manual labor of hunting for comments across emails, Slack threads, and PDFs, ensuring that no request is missed and no outdated note is accidentally applied.

For example, you can connect Claude Duet to your client's shared feedback portal. Every time a client submits a comment, the agent instantly processes it: it identifies the asset being discussed, checks the current version number, classifies the note as a Correction or Preference, and routes it into the specific asset folder in your project tool with a classification tag.

As a creative director, you no longer act as a data entry clerk. Instead, you start your day by reviewing a pre-sorted dashboard of categorized requests. You can quickly approve the technical Corrections and strategically evaluate the Preferences, making high-level decisions without getting bogged down in the logistics of feedback collection.