AI for content ideation: the process that replaced our brainstorm meetings

AI content ideation has quietly become one of the most practical applications of AI in our day-to-day work at Choco Media. Not because it produces brilliant ideas on its own, but because it collapses the time between “we need content ideas” and “here are 20 filtered, prioritized topics worth writing about.” What used to take a 90-minute meeting now takes roughly 20 minutes — and the output is consistently more useful than what we were producing in the room together.

This post is for marketing teams and agency operators who are still running content brainstorms the old way: a calendar invite, a shared doc, a whiteboard session that goes sideways after the third tangent, and a list of ideas that half the team forgets about by Thursday. There is a better process. We use it every sprint, and it is not complicated.

What you will leave with: a step-by-step workflow for AI-assisted topic generation, the prompt structure that makes it actually useful, a filtering framework to cut the list fast, and the specific tools we use at each stage.

Why brainstorm meetings fail at scale

The traditional content brainstorm has a structural problem: it conflates two different cognitive tasks. Generating ideas requires loose, associative thinking. Evaluating ideas requires critical, comparative thinking. When you try to do both in the same meeting, people either hold back (afraid to suggest something bad) or the group latches onto the first idea that sounds good and stops generating.

The other issue is preparation asymmetry. One person in the room has read three competitor posts this week. Another has been deep in a client account and has no context. A third is thinking about a completely different quarter. The ideas you get reflect whoever prepared most, not what the audience actually needs.

• Idea generation and idea filtering require different mental modes — mixing them produces worse output of both
• Unequal preparation means the loudest voice in the room shapes the content calendar
• Meeting time is the most expensive time a team has — using it for idea generation is a poor investment
• Good ideas often surface in the two days after the meeting, when there is no mechanism to capture them

AI does not fix all of these. But it separates generation from filtering cleanly, removes preparation asymmetry by doing the research itself, and runs asynchronously — so you can run the process without pulling everyone into a room.

The four-stage AI ideation process we use

The process has four stages: research, generation, filtering, and prioritization. Each stage has a specific role. None of them require a meeting.

Stage 1: Research input

Before generating anything, we give the model context. This is the stage most teams skip, and it is why their AI brainstorms produce generic ideas. We pull three inputs: recent search intent data for our core topics (using a tool like Ahrefs or even a manual SERP scan), a list of topics we have already covered (to avoid redundancy), and any specific angles or campaigns the business is running in the next quarter.

This takes about five minutes to assemble. The research does not need to be exhaustive. It just needs to give the model enough signal to generate ideas that are specific rather than obvious.

Stage 2: Generation prompt

We run a single generation prompt that asks for a large volume of ideas — typically 30 to 40 — across a specific topic cluster. The prompt specifies audience, format constraints, and what we are trying to avoid.

The goal at generation stage is volume, not quality. You want enough ideas on the table that you can afford to throw out 70% of them and still have a strong calendar.

Stage 3: AI-assisted filtering

We paste the full idea list back into the model with a filtering prompt. This prompt asks it to score each idea against three criteria: search intent clarity (is there an obvious query this answers?), differentiation (does it say something different from the top 3 results currently ranking?), and fit for the stage of the funnel we are focused on this quarter.

Stage 4: Human final cut

The filtered shortlist — typically 8 to 12 ideas — goes to one person, not a team. That person makes the final call in under 10 minutes, knowing that everything on the list has already cleared a basic quality bar. They are not evaluating from scratch. They are choosing between already-filtered options.

The generation prompt structure that works

Most teams write weak generation prompts. They ask for “blog post ideas about content marketing” and get a list of titles that could appear on any agency site in the world. The fix is specificity at four levels: who you are, who you are writing for, what you already cover, and what you want the ideas to accomplish.

Here is the prompt structure we use, adapted for content marketing topics:

You are a content strategist for [agency/brand description].
Audience: [specific role, company size, industry, pain state].
We have already covered: [list 5-10 existing topics].
Goal this quarter: [specific business outcome — signups, retainer inquiries, organic traffic to X cluster].
Generate 35 content topic ideas that are: specific enough to rank for a clear search query, different from generic takes, and appropriate for a [awareness/consideration/decision] stage reader.
Format: Title | Target search query | One-sentence angle. No duplicates of the existing list.

• The “we have already covered” section is the single most important addition — it forces the model to generate new ground rather than recycling what you have
• Specifying output format (Title | Query | Angle) makes the filtering stage dramatically faster
• The stage constraint keeps ideas anchored to a specific funnel position rather than scattered across awareness and conversion topics randomly
• Adding “No duplicates of the existing list” sounds obvious but prevents the model from repackaging your existing titles with slightly different wording

We run this prompt in Claude or GPT-4o. We have found that longer, more structured prompts produce better output in both. If you are working from our AI prompt library, this prompt type belongs in the “strategic planning” category rather than the “content creation” category — it is a planning tool, not a writing tool.

The filtering prompt that cuts the list in half

After generation, we take the full list of 30 to 40 ideas and run a second prompt to filter it down. This is not asking the AI to pick the best ideas — it is asking it to surface which ideas meet specific criteria so a human can make the final call faster.

The filtering prompt:

Score each topic on a 1-3 scale across three dimensions:
1. Search clarity: Is there an obvious, specific search query this answers? (1 = vague, 3 = clear query)
2. Differentiation: Does this say something the top Google results do not? (1 = commoditised, 3 = genuinely different angle)
3. Funnel fit for [goal this quarter]: Does this reach the right reader at the right stage? (1 = off-target, 3 = strong fit)

Return the top 10 by total score. For each, state the score and one sentence explaining the differentiation score specifically.

The differentiation dimension is the one that produces the most useful signal. It forces the model to reason about what already exists in the SERP, which surfaces which ideas are crowded versus which have room to be the best piece on the topic.

How we integrate this into the content calendar

We run this process once per sprint, usually on the Monday or Tuesday before a new content cycle starts. The output is a ranked shortlist of 8 to 12 ideas. From that list, we assign 3 to 4 topics to the current sprint and park the rest in the queue for future sprints.

The person running the process is not a content writer — it is whoever owns the content strategy function. In our case that is the same person doing account work, which means the ideation is grounded in what clients are actually asking about that week. That context is hard to replicate in a team brainstorm where writers may not have direct client contact.

• Run ideation once per sprint, not once per post — batching the thinking reduces cognitive switching cost
• Keep a backlog queue so good ideas from one sprint do not get lost when you run out of room
• Rotate the person running the research input stage across team members — different people pull different signal
• Document which filtering criteria you are using so the process stays consistent across different operators

This connects directly to the broader question of how AI fits into AI content creation at scale. Ideation is the top of the funnel for a content operation. If it is slow or inconsistent, everything downstream — briefing, writing, publishing — inherits that inconsistency.

What AI cannot do at the ideation stage

There are two things the AI process does not replace, and it is worth being direct about them.

First: genuine insight from client conversations. When a client says something in a call that does not show up in any search data — a fear, a misconception, a question they were embarrassed to Google — that is ideation gold. No AI process surfaces that. It lives in call notes, in Slack messages, in feedback feedback forms. The best content ideas we have had in the past year came from verbatim quotes from client conversations, not from search data or AI generation.

Second: trend detection before the data exists. Search tools show what people are already searching. AI models are trained on data with a lag. The topics that will matter in six months are not well represented in either source. Staying close to what is actually happening in your industry — through newsletters, community forums, early-stage tools, and people doing experimental work — is still a human responsibility.

• Client conversation notes are the most underused ideation source in most agencies
• Search data tells you about existing demand, not emerging demand
• AI generation is good at exhausting known territory quickly — it is not good at inventing new territory

The tools we use and what they each contribute

The process is tool-agnostic but here is what we actually use. For research input: Ahrefs for search volume and competitor content gaps, and a manual SERP scan for the 3 to 5 queries closest to the cluster we are working in. For generation and filtering: Claude for structured prompt chains, because it handles the formatting constraints and multi-step reasoning more consistently. For the queue and calendar: a simple markdown file in our Notion content engine, which makes the backlog visible to anyone on the team without needing a separate project management tool.

Total cost at scale: the AI generation and filtering takes roughly 10 minutes of active time. The research input takes 5 to 10 minutes. The human final cut takes under 10 minutes. Total: 20 to 30 minutes, one person, no meeting required.

• Ahrefs (or any keyword tool) for research input — the data quality here affects the idea quality significantly
• Claude or GPT-4o for generation and filtering — both work, Claude handles the formatting constraints more consistently in our experience
• A simple queue document to capture ideas across multiple ideation runs, not just the current sprint
• No dedicated ideation tool required — the value is in the prompt structure, not the platform

Common mistakes teams make when switching to AI ideation

The most common failure mode is running the generation prompt without the research input stage and without the “already covered” list. This produces a usable list of ideas — but they are the same ideas everyone else is producing with similar prompts. Generic inputs produce generic outputs. The research context is what separates a useful AI ideation run from a mediocre one.

The second failure mode is skipping the filtering stage and taking the generated list straight to a human for evaluation. At 30 to 40 ideas, that evaluation takes longer than the original brainstorm meeting. The filtering prompt exists to do the cognitive sorting work before a human sees the list, not as an optional extra step.

The third failure mode is running this as a one-time experiment rather than a repeatable process. The value compounds over time as your “already covered” list grows, as you refine your filtering criteria based on what actually performs, and as the process becomes fast enough that you run it more frequently.

• Always include the existing content list — without it you will generate redundant ideas
• Always run the filtering stage before human review — otherwise the time savings disappear
• Treat the first three runs as calibration — refine your filtering criteria based on which ideas you end up using
• Keep a record of which AI-ideated topics actually performed, and feed that pattern back into future generation prompts

If you are thinking about how this connects to your broader content strategy infrastructure, our AI automation services page covers how to build the systems that make this kind of repeatable process sustainable at higher volume.

Putting it together: the 20-minute ideation run

For teams that want to run this immediately, here is the compressed version. Assemble three inputs: your existing content list, a short research note on the cluster you are planning for, and your business goal for the quarter. Run the generation prompt with those inputs. Take the output and run the filtering prompt. Hand the filtered shortlist to whoever owns the content calendar for a final 10-minute review. That is the full process.

The 90-minute brainstorm meeting is not producing better ideas than this. In most cases it is producing worse ones, because it rewards whoever speaks most confidently rather than whatever topics are most strategically sound. The AI ideation process does not have that problem. It generates exhaustively, filters against criteria you set explicitly, and leaves the judgment call to a single person with full context rather than a room of people with mixed preparation.

If you want to see how we build processes like this for clients — from ideation through production and distribution — get in touch and we can walk through what that looks like for your specific setup.