Scheduled Chats and Agent Skills for Between-Meeting Client Check-Ins

A monthly review is too slow for some clients and too frequent for others. The ones who want to know what is going on in between tend to email on a Thursday asking if the numbers moved. The ones who do not care still need to see that someone is watching.

We solve both with scheduled, lightweight briefings. Here is the setup and the rules we follow.

What the tools are

Promptwatch's Agent Chat is a conversational analyst that works over live data. It runs multi-step workflows, delegates to read-only subagents, draws charts, builds and edits dashboards, generates PDF reports, and works the Action Board. A side panel opens on any page with the command-J shortcut, including inside the content editor.

Two features make it suited to check-ins.

Agent Skills. Saved routines such as a seven-day GEO briefing or a competitor gap analysis. Instead of rewriting the same question each week, we run the skill.

Scheduled Chats. A chat that runs on a schedule and delivers its output in the app or in Slack. A briefing arrives on its own.

Slack delivery is not available on white-label workspaces, which matters for agencies using that setup. We check this before promising a client a Slack briefing.

What we schedule

We keep the list short.

A weekly briefing for the internal team. The seven-day GEO briefing skill, delivered to the account channel on Monday. It gives us a starting point for the week, not a finished analysis. We read it, check anything that matters, and decide what to do.

A weekly or fortnightly note for the client contact who wants updates. Plain language, one screen. What moved, what we did, what is next. We review it before it goes anywhere near the client.

A competitor gap check. Monthly, ahead of the review meeting, so we have a fresh view of where rivals are gaining.

A crawler errors summary. Weekly for clients with active infrastructure changes. It points us at new errors before they cost citations.

What we do not automate

• The monthly review itself. That is a conversation.
• Any statement about cause. The agent can show data; we interpret it.
• Anything touching client strategy or sensitive claims.
• Unreviewed messages to the client. A person reads every message sent in our name.

Automation is good at recurring retrieval and summary. It is poor at judgement and tone.

How we check the output

Agents can misread data, and a confident summary is not a verified one. Our checks:

1. Open the underlying dashboard for any number we plan to quote.
2. Confirm the period and the prompt panel used.
3. Cross-check a surprising change in another view, such as the citation list or the crawler log.
4. Strip anything that sounds more certain than the data allows.

This takes a few minutes and has caught errors. We prefer the discipline over the risk of sending something wrong.

Designing a useful briefing

A good briefing has a fixed structure that readers learn quickly:

• Headline: one sentence.
• Movement: the key figures over the period, with the caveat on variance.
• Notable items: one or two things worth attention, such as a page that dropped out or a new competitor.
• Work: what shipped and what is planned.
• Ask: any decision we need.

If there is nothing to report, we say so in one line. A short note that says "no meaningful change, work continues on X" is better than a padded one.

Setting expectations with the client

We tell clients what the briefings are: a convenience between reviews. We explain that weekly movement is noisy and that the monthly review is where we interpret. We also tell them they can reply with questions, and a human will answer.

Using the editor panel

The side panel in the content editor is handy during drafting. We ask it questions about the gap data for a page: which prompts it targets, what competitors are cited, and what the current coverage looks like. It does not write the page for us and publish it. It supports the writer.

MCP for clients who like their own tools

Some clients want to ask questions themselves. Promptwatch has a hosted MCP server that works with tools such as Claude and Cursor, with scoped, read-only API keys for those who should not edit anything. We set up read-only access for clients who like to explore, and keep write access with the team.

Cost and plan notes

Features vary by plan, so we check before planning an automation around one. Agency plans include Agent Chat, MCP, and the API, but not the knowledge base or the automated Content Agent. If a client's plan lacks something, we adjust the scope rather than promise it.

Keeping it honest

If a scheduled chat produces something odd, we fix the skill or the prompt, not just the output. Over time the skills get better and the review time drops.

Our recommendation

Automate recurring retrieval, review everything before it leaves the building, and leave interpretation to people. We use Scheduled Chats and Agent Skills in Promptwatch for this, and keep the monthly meeting for judgement. You can find the platform at promptwatch.com.