A marketing workspace queues agent work for approval and puts paid calls behind a gate

The task

A blank chat with a coding agent starts from nothing each time. The same audit requested on two different days comes back different, because nothing says what a good audit contains for this business, which sources count, what the agent must not invent, or when to stop and ask. Paid data calls add a second concern, since each one costs money.

What they did

The marketer built a workspace in a single Obsidian vault of plain Markdown files. It holds step-by-step skills with inputs, steps, output, and limits for each job, sourced knowledge packs, checklists tied to official documentation, and written roles. House rules tell the assistant to use only facts from the folder, to say what it could not verify, and to ask before spending or publishing. A request goes to an orchestrator role, which picks the specialist roles and the review the work needs. The specialist follows the matching skill and checklist. A separate reviewer role, played by the same assistant under different instructions, checks the result against its brief and sources. The assignment, result, review, and decision are saved as linked notes. Before a paid call runs, the agent drafts an approval note that lists the precise action, where it points, the most it could cost, and which effects cannot be reversed. The agent then pauses. The example in the post is a DataForSEO backlink pull for the author's own site.

What the source reports

According to the source, the example approval note covered up to 4 read-only API requests with a worst-case cost of 0.33 USD, checked against a 1.00 USD cap before each call. According to the source, in one test an agent following the blog workflow skipped the SEO and editorial reviews, and the destination site's validator rejected the result with 35 errors. According to the source, a rerun of the local SEO knowledge pack together with its skills satisfied all assertions in 16 of 33 test cases.

Where the approval sits

Finished work lands on an activity board that shows who did it, who reviewed it, and what it was checked against. The person answers accept, ask, or deny. An agent's own statement does not count as acceptance, and work counts as checked only after a reviewer role passes it. Paid API calls wait at the approval note until the person releases them. The source describes both controls as written rules the assistant follows; nothing in the vault enforces them, apart from some record checks. It suggests also setting spend caps with each provider.

Worth knowing

The author built this system and has a commercial interest in it, so this is a first-party account. Each role is a set of written instructions for one assistant. The assistant writes the records itself, and nothing flags a session that skips them. The activity board cards shown in the tour were demo items, and the source tells readers to review every output before using it.

More in this category

• Marketing operations

  A marketing lead has Claude Code draft six weekly meeting docs and reviews every line

  A marketing lead at a startup runs one Claude Code command each Sunday evening to prepare the next week's meeting documents. The lead reviews every line, and nothing reaches a live document until the lead signs off.

  Read the entry
• Marketing operations

  A small marketing team has agents read six systems and approves what reaches customers

  A marketing lead at a small company has AI agents read six marketing systems in one pass and return one page on what changed. The agents draft and analyze, and a person approves anything that goes to customers.

  Read the entry