---
name: content-refresh-workflow
description: Build an agentic content refresh workflow for the person, end to end. The workflow is kicked off by a task, finds where their brand is missing from AI search, researches keywords, writes the brief and a refreshed article, creates a video pack, sends everything for human approval, publishes to their site, and repeats on a weekly cycle. Use when someone wants to build this workflow for themselves, says their brand is not showing up in ChatGPT or AI search, wants to publish fresh content faster, or points you at the Show Me Your AI content workflow episode.
---

# Build a content refresh workflow

Companion to the *Show Me Your AI* episode on content workflows. Your job is to
build the workflow the episode demonstrated, for this person, using the tools they
actually have. You are building it, not just describing it.

## What the episode showed

A Notion task says "do my weekly content refresh." An agent then:

1. Confirms the setup and picks the article that should be refreshed.
2. Pulls the keyword cluster, volume and rankings (the demo linked to Ahrefs).
3. Builds the brief, topics and headings for the refresh.
4. Writes the updated content. In the demo that took 30 minutes instead of 30 days.
5. Writes a video pack, so a video is created in the same run.
6. Sends the result for approval, in Claude and in Slack.
7. Publishes the approved article and video to the site (the demo used Vercel).

The old way of doing the same job took about six weeks: the CMO spots the gap, emails
the team, the team emails the agency, the agency briefs the SEO team, drafts go back
and forth, and so on. LLMs reward freshness and volume of content, so one piece a
month does not rank. The goal of this build is to replace that chain with one
pipeline, with a human approving only where judgement matters.

Every stage is a swap point. Humans can do some stages, agents can do others, and
any tool can be replaced without changing the shape of the workflow.

## Before you start

Ask these first. Do not guess any answer, and do not start building until you have them.

1. Which brand and site is the refresh for?
2. Where should the task live that kicks it off? A Notion task worked in the demo,
but Slack, a ticket, a form or a scheduled trigger also work.
3. Which tool do they use for each stage? Fill in the table in Step 1.
4. Who approves content, and where? Slack, a web app, email or the task itself.
5. Where does approved content get published? A site deployed on Vercel, a CMS, or
something else.
6. Which stages should run on their own, and which need a human to approve?

Then check connectors. For every tool they named, confirm you can reach it from this
session. If one is missing, stop and tell them which connector to enable. Do not
carry on without it, do not fake what a tool would have returned, and do not quietly
swap in a different tool.

## Step 1: map the stages to their tools

Fill this table in with the person before building. The demo tools are examples only.

| Stage | Demo tool | Their tool |
|---|---|---|
| Trigger | A Notion task | |
| Find AI search gaps | Profound | |
| Keyword research | Ahrefs | |
| Brief and headings | An LLM instruction | |
| Draft | An LLM or writing agent | |
| Video pack | An LLM instruction | |
| Approval | Claude plus Slack | |
| Publish | Vercel | |

If a stage has no tool, agree a manual step for it and say what they lose by doing it
by hand.

## Step 2: set up the trigger

Create the task or message that starts a run. Keep it short: the brand, the word
"refresh," and the date. The agent should be able to start from that alone, and the
run should work whether it is started by a person or by a schedule.

## Step 3: write the instructions for each stage

Each stage gets a short written instruction the agent reads at run time. Use this
shape for every one:

```markdown
## When to use this
The situation that triggers this stage.

## Inputs
What the stage needs, and in what format.

## Steps
1. ...
2. ...

## Output
Exactly what gets produced, and where it goes next.

## Known failure modes
Things that went wrong before, and what to do instead.
```

The brief instruction needs the most care, because everything downstream depends on
it. It should require: the target reader, the primary and supporting keywords, the
questions the page must answer, and the full heading structure (one H1, then H2s and
H3s). Add a check that the headings are valid, since wrong H1 tags are a classic
QA failure.

Add rules that apply to every stage:

- Show your work. Say which tool you used and what it returned.
- Never invent numbers. Volumes, rankings and gap data come from a tool or from the
person. If the data is missing, say so.
- Keep what is already good on the live page, and change the rest.
- Do not invent statistics, quotes or product claims, and flag anything you could not
verify.
- If a stage fails twice, stop and report instead of improvising.

## Step 4: connect the pieces

Wire the stages together in this order, and test each connection as you go:

1. **Research connection.** Confirm you can pull AI search gaps and keyword data.
2. **Brief and draft.** Confirm the brief produces a valid heading structure and the
draft follows it.
3. **Video pack.** Write a short video script and a list of supporting assets from the
article. Ask the person how far they want this to go, since the format is theirs to decide.
4. **Approval.** Send the draft and a link to the full content to wherever the
approver works. Do not publish until you receive an explicit yes. If they ask for
changes, go back to the draft.
5. **Publish.** Push the approved article, and the video if there is one, to the
site. Confirm the live URL.

## Step 5: do one supervised test run

Run the whole workflow once on a real article, with a human watching every stage.
After each stage, write a one or two line status note. At the end, report:

- What ran, and what each tool returned.
- Anything you assumed, and anything that failed.
- The live URL, if it was published.

Fix the failures before turning on any automation. Add each failure to the relevant
stage's Known failure modes section.

## Step 6: put it on a cycle

Once the test run works, make it repeat. Log what was refreshed and when, so the next
run picks a different article. Ask the person how often they want it to run and
who gets told when it finishes.

Only remove a human from a stage if the person says so. The demo ran start to finish
on its own, but they chose where to keep a human approving.

## What to tell them at the end

Three sentences, no more:

- The workflow now runs from task to published page, and a human approves before it goes live.
- Any tool can be swapped without changing the steps.
- The Known failure modes sections are where it gets better, so log problems as they happen.

## Scope

What you have built is one workflow that refreshes one article per run. It does not
cover several brands, a content calendar or a full analytics loop, so treat those as
separate builds if they ask. Behind a demo like this sits real engineering:
connectors to each tool, specific instructions for each stage, an approval path and a
publishing connection. Expect the setup to take longer than a single run.