How to Design an Agent Content Topic Selection Workflow

When people talk about AI content creation, the first instinct is usually:

"let the model write."

But after working with real workflows for a while, a different bottleneck becomes obvious:

what to write, why it matters now, and whether the topic is worth writing at all.

That is topic selection.

If that layer is weak, the rest of the workflow becomes low-value repetition.
The model may write faster, but it is only producing weak content more efficiently.

That is why I increasingly split content workflows into three stages:

• front stage: source retrieval
• middle stage: topic judgment
• back stage: formatting and publishing

Among those three, topic selection is often the most important and the least explicitly designed.

Why topic selection deserves its own workflow

Writing usually answers questions like:

• how should this article be structured
• how should this point be expressed
• how should the argument flow

Topic selection answers earlier questions:

• which subject is worth covering now
• which angle fits the channel
• which sources are strong enough to support an article
• where the differentiation opportunity is

If those questions are not resolved first, drafting tends to collapse into:

• stitching together what the model has seen recently
• expanding a vague theme from memory
• repeating points that already exist everywhere online

That is why topic selection is not a side task. It is the judgment layer that determines whether the workflow should continue at all.

What an executable topic workflow should output

Many people reduce “topic selection” to a title.

That is too thin.

For a real workflow, the end product should be a compact topic brief with at least:

• the topic itself
• the recommended angle
• the target audience
• why the topic matters now
• which sources support it
• which repeated angles should be avoided
• what kind of article structure fits best

In other words, the output should not look like:

• "write something about AI"

It should look more like:

• "write an explainer for WeChat readers about a recent product update, focusing on what changed, who it matters to, and how it affects workflow decisions, while avoiding a generic announcement summary"

That is the kind of output a downstream writing agent can actually use.

Stage 1: retrieve sources first, do not let the model invent context

The first step of topic selection is not brainstorming.

It is source retrieval.

That is where jina-cli becomes useful.

It is a strong fit for:

• searching candidate sources
• reading articles, blogs, release pages, and X posts
• turning URLs into Markdown, Text, or JSON
• building a source pool that can be filtered later

Without this layer, topic judgment often depends on:

• model memory
• single-source interpretation
• vague notions of what is trending

With a readable source pool, the agent can start comparing actual material.

Stage 2: deduplicate, cluster, and group material by theme

Once the material is available, the next move should not be drafting.

A stronger sequence is:

1. deduplicate
2. cluster
3. summarize recurring themes

Because the same story often appears across many sources, but with different emphasis.

A useful middle layer should ask:

• are these sources describing the same event
• what is actually new across them
• which sources are original signals versus repackaged summaries
• which sources add genuine perspective

After this step, you no longer have a pile of links.

You have topic clusters.

That changes the quality of every later decision.

Stage 3: score candidate topics explicitly

Topic selection is really a ranking problem.

Not “is there a topic,” but “which topic deserves priority.”

I strongly prefer making the scoring dimensions explicit rather than asking the model for vague opinions.

A useful scoring set often includes:

1. timeliness

• did this just happen
• is there still a useful publishing window

2. audience relevance

• does this matter directly to your readers
• is it actionable or only mildly interesting

3. differentiation potential

• do you have a stronger angle than the obvious one
• has the topic already been exhausted by others

4. evidence quality

• do you have enough source material to support a solid article
• can multiple trustworthy sources confirm the key facts

5. cost-to-output ratio

• how much extra background is required
• is the topic likely to turn into a complete article efficiently

If the agent can score topics against those dimensions and explain the reasoning, the workflow becomes much more stable than intuition-only topic choice.

Stage 4: do not only choose the topic, choose the article form too

Many topic systems stop after saying:

• "this topic is recommended"

That is not enough.

The next question is:

how should this topic be written.

The same topic could become:

• a news brief
• an explainer
• an opinion piece
• a list article
• a tutorial

If the workflow does not decide the form early, the drafting stage becomes fuzzy again.

That is why a mature topic workflow should also include:

• recommended article type
• recommended structure
• possible title directions
• angles that should be avoided

That gives the writing stage a real starting point.

Stage 5: hand the result to the writing agent instead of restarting the workflow

At this point, a strong topic workflow should already produce enough structure to support writing.

The writing agent should receive:

• the topic brief
• source summaries
• key evidence
• recommended structure
• tone and style constraints

It should not have to rediscover the topic from scratch.

That is the real value of this workflow layer:

it resolves both “is this worth writing” and “how should it be written” before drafting begins.

That keeps the writing step focused on expression instead of redoing upstream decisions.

Stage 6: after drafting, move into formatting and publishing

This is also why I see jina-cli, topic workflows, and md2wechat as parts of the same chain.

A more complete content automation pipeline should look like this:

Front stage: source retrieval

• jina-cli
• search
• URL reading
• source pool creation

Middle stage: topic selection

• deduplication and clustering
• topic scoring
• angle selection
• topic brief generation

Back stage: writing, formatting, and publishing

• draft generation
• md2wechat formatting
• md2wechat Agent API draft creation
• asset upload and delivery into WeChat drafts

If you only solve retrieval, the workflow lacks judgment.
If you only solve writing, the workflow lacks reliable publishing.

That is why these three stages are best understood together.

A practical minimum loop

If you want to start small, the workflow does not need to be huge.

A good minimum loop could be:

1. use jina-cli search to find 10 candidate sources
2. use jina-cli read to extract their content
3. let the agent deduplicate and cluster the material
4. let the agent score the clusters and return the top 3 topic options
5. pick one and generate a writing brief
6. draft the article
7. format it with md2wechat
8. send it into WeChat drafts

That loop already proves an important point:

content automation is not only about writing speed. It is about whether each stage produces a stable handoff for the next one.

Common mistakes

1. treating “what is trending” as a topic

Trend signals are not the same thing as editorial topics.

A topic still needs:

• a target audience
• a clear angle
• enough supporting evidence

2. letting the model decide everything implicitly

Without scoring dimensions and decision criteria, output drifts quickly.

3. choosing the topic but not the form

That forces the writing stage to become fuzzy again.

4. focusing on drafting while ignoring publishing

If formatting, assets, and draft creation still happen manually, the workflow is still broken near the finish line.

Closing thought

An Agent content topic workflow should not ask the model to "invent a topic."

It should ask the agent to make a grounded editorial decision:

• retrieve sources
• organize sources
• compare themes
• score priorities
• produce a writing brief

When that layer is done well, drafting becomes steadier, formatting becomes easier, and publishing connects more naturally.

If jina-cli solves content retrieval and md2wechat solves formatting and publishing, then topic selection is the judgment layer in the middle that makes the whole pipeline worth running.

Continue Reading

• Retrieval layer: jina-cli
• Toolchain overview: From Content Retrieval to WeChat Drafts: The Content Creation Toolchain I Design for AI Agents
• Publishing workflow: What a WeChat Automation Workflow Should Include
• Related writing guidance: How to Write Prompts for WeChat Official Account Articles