Prompting is becoming routing

Prompting tells an AI what to do. Routing decides where the work should go. As models and specialist tools multiply, that second decision is becoming more important.

For a while, using AI well mostly meant getting better at asking one model for what you wanted. You opened ChatGPT or Claude, explained the task, added context, refined the prompt and tried again.

That still matters. But it is no longer the whole job.

Once you have several capable models, web search, coding agents, research tools, spreadsheets, your own files and persistent knowledge, the first question changes. It becomes less how do I prompt this model? and more what is the right place to do this piece of work?

I think that is routing.

Routing is just choosing where work goes

The word can sound more technical than the idea.

Imagine you are preparing an important presentation. You might use a fast model to summarise 30 pages of background material, a stronger reasoning model to find the argument, a spreadsheet or coding tool to analyse the numbers, web search to check a current claim, and another model to challenge the recommendation before you present it.

You might also decide that one part should not go through AI at all because the judgement is yours to make.

The skill is not writing one enormous prompt that somehow performs every stage. It is deciding which job should go where, what context each stage needs, and where another perspective is worth the extra effort.

That is routing.

Why this matters now

When there was one obvious AI destination, routing barely mattered. You opened the chatbot you used and prompted it.

Now the landscape is different. Some tasks are routine transformations. Some need stronger reasoning. Some are better handled by a specialist tool. Some benefit from a second model because independence matters more than raw capability.

That creates a simple economic question as well as a quality question. The most powerful model does not need every task.

I have written before about when the best AI model stops being worth it. The principle is the same here: use sufficient intelligence, not maximum intelligence.

If a fast model can clean up notes reliably, use it. If the task is difficult synthesis, escalate. If the job is code, use the coding environment. If a factual claim matters, verify it rather than assuming a fluent answer is correct.

The interesting shift is that model choice becomes part of the work.

I realised I was already doing this

I noticed the pattern through my own Editorial Intelligence work, but the principle is not really about editorial work.

An idea usually starts with me. ChatGPT might help me develop it, connect it to previous thinking and turn it into a workable argument. Useful context is stored in GitHub so it does not disappear inside one conversation. Claude or a coding agent can then approach the work from a different angle, rewrite it, challenge it or implement something on the site. I decide what survives.

I had been doing that manually without thinking of it as a system.

The important point is not that one model is permanently “the writer” and another is permanently “the critic”. Those roles can change. The useful pattern is assigning the work by role rather than assuming everything belongs in the same chat.

That is why I still like the idea in the second model is not a proofreader. A second model can be useful because it creates separation of perspective, not because it automatically knows better.

Then I made routing too complicated

Naming the pattern created a predictable temptation: I tried to turn it into a process.

For one article, a quick check of the existing work was useful. It showed what I had already said, what evidence the new argument needed and where the claim might be weak.

Then I kept adding stages.

The piece was going to get an independent challenge. That created a challenge brief. The brief had to be stored. Different branches had to see the same material. Before long I was spending more attention arranging the handoff between systems than thinking about the article itself.

The router had become the work.

The frustrating part is that the extra scrutiny did eventually find something valuable. The proposed article overlapped heavily with work I had already published, so I improved the existing pieces instead of adding another one.

I did not conclude that the checking was pointless. I concluded that the checking had become too expensive.

That distinction matters. Good routing should reduce unnecessary work while keeping the scrutiny that actually changes the outcome.

Routing is a decision, not a pipeline

My first version of the idea looked too much like this:

idea → novelty check → evidence → development → persistence → independent challenge → human gate → drafting → verification → logging

There are situations where that level of scrutiny is justified. A major methodology change, an important public claim or a difficult decision may deserve independent challenge and explicit verification.

But making that the default turns a useful principle into bureaucracy.

A better default is closer to:

work arrives → choose the right route → do the work → escalate only when needed

For my own publishing, that often becomes:

idea → quick duplication and evidence check → develop and draft → optional second-model pass → human decision → publish

The intelligence is mostly in knowing when an extra stage is worth adding.

Three questions are usually enough

The routing layer I actually want is small enough to disappear into the work.

Before starting, it should be able to answer three questions:

1. Does this already exist?
2. What evidence or context does it need?
3. Would another model or specialist tool improve the result enough to justify the handoff?

Those questions apply well beyond publishing.

For a market analysis, the second question may trigger current web research. For a financial model, the third may send the calculation to a spreadsheet or code rather than a chatbot. For a routine rewrite, the answer to the third question may simply be no.

The point is not to maximise the number of AI systems involved. The point is to make a better decision about the work before generating more of it.

The durable context matters more than the route

There is one part of my original thinking I have not changed.

I do not want useful knowledge trapped inside whichever model happened to produce it. Models change, relative strengths change and the tool I prefer today may not be the one I use next year.

The durable layer should sit outside them: evidence, previous arguments, decisions, research, source material and things learned from earlier work.

That is the idea behind the workflow should outlive the model.

Routing works better when the context can travel. But that does not mean every handoff needs a formal package or a complicated system. Often it is enough that the relevant material exists somewhere the next tool can retrieve it.

The infrastructure should support the work, not become another job to maintain.

Prompting still matters, one level down

None of this makes prompting obsolete.

Once the work has been routed, the chosen model still needs a useful brief. Good context still matters. Bad instructions can still produce bad output.

The difference is that prompting is increasingly happening one level lower in the stack.

First decide whether the job needs a fast model, a stronger reasoning model, web research, a coding environment, another perspective, a specialist tool or no AI at all. Then tell that system what role it needs to perform.

That feels closer to how serious AI use is developing. The unit of work is no longer always one person talking to one chatbot. It is a person deciding how a set of different capabilities should contribute to the job.

That does not require an autonomous agent or an elaborate orchestration layer. Most of the time, routing can remain a human decision.

The test is simple: does the route make the work easier and the result better?

If not, remove it.

Routing should make complicated work simpler. It should not make simple work complicated.

Prompting tells an AI what to do. Routing decides where the work should go.