From "what should I write?" to two published posts, in one conversation

This is a lightly edited transcript of an agent session. The company is fictional (Wirefield, a press-release distribution service), but the tools called, the timing, and the shape of the output are exactly what you'll see. The agent was Claude Code with the Letterstory MCP server connected (KB-402).

"Give me some topics I can write about for Wirefield"

The agent's first move was to read before it wrote. It called five tools: the organization's positioning, its content inventory, the collection's topic map, and the scored headline suggestions, and reported back:

Wirefield's collection is brand new: 56 spokes across 7 clusters, nothing covered yet, so you have a wide-open field. Here's how I'd prioritize, grounded in the positioning (accessible PR, wholesale pricing, PR × SEO × AI search) and the headline scores.

Start here (highest signal): the top-scored suggestions already queued, all with zero competitor coverage:

1. How to search a journalist database for the right beat reporter, "media outreach", 140 searches/month, difficulty 36. The anchor piece for the outreach cluster.
2. Wire service costs for small businesses, compared side by side, sets up the wholesale-vs-retail argument without being a sales page.
3. What white-label press release distribution actually means for agencies, difficulty 7, easy to rank.
4. How earned media gets your brand cited in AI search results, the first-mover piece for the AI Search Visibility cluster.
5. How to write a podcast pitch email that hosts actually reply to.

Strong follow-ons by cluster: Press Release Distribution, AI Search Visibility (the differentiating cluster: own it early), Earned Media & SEO, PR for Founders … (four or five per cluster)

One thing to watch: several briefs lean on Wirefield's features as the working example. That's fine for high-relevance spokes, but the AI-search and SEO pieces will earn more authority if they stand on their own and mention the product lightly.

If you want, I can queue any of these, every spoke above has an open topic_id, so I can set it for the next planner slot or draft it now.

Notice three things. It grounded the list in your strategy and topic map rather than inventing. It ranked by search demand and competitor coverage because those numbers are in the tools. And it offered the next action without taking it.

"Okay, write about the Earned Media & SEO and PR for Founders"

The agent picked one spoke from each cluster ("the highest-value spoke in each"), then called run_collection twice with each spoke's topic_id:

Both runs are going. Research plus a kernel run takes 20 to 60 minutes each, so I'll check back when they finish.

Cluster	Spoke	Run
Earned Media & SEO	Google News inclusion criteria for press release content: eligibility signals, and which wire paths get a release indexed	2289bfc5…
PR for Founders	PR strategy for bootstrapped companies without agency budgets: the 90-day DIY plan: news calendar, first two releases, a 30-outlet media list, cost estimates at wholesale rates	844b90fd…

Both were bound to their topic_id, so the topic map will mark those spokes covered and the cadence won't re-queue them. Each one spends this week's cadence budget rather than a separate credit. One thing to be aware of: this pipeline publishes and rebuilds the site once the draft finishes, if you'd rather review before it goes live, say so and I'll unpublish it when it lands.

Then it set a ten-minute timer and polled get_collection_run until both read published. Each post went through Research → Write → Enhance passes (links, sources, visuals, metadata) → publish → site rebuild, and the agent came back with the two live URLs.

What to take from this

• Ask open questions first. "What should I write?" gets you a ranked, grounded list. You choose; the agent executes.
• Name the cluster or the headline, not the tool. The agent maps your words to list_topics, run_collection, and the rest. You never have to.
• Say how you want to review. The one-shot pipeline publishes. If you want a draft first, say "write it but don't publish" and the agent will use the draft-only tools instead (KB-502).
• Let it wait. Long jobs are normal. A good agent tells you the run ids and checks back.

More phrasings that work well in KB-505.