Optimizing Models
Artificial intelligence can do everything from sorting a list of names to helping design a complicated computer system. However, that does not mean the same AI model should be used for every job. Some models are fast and inexpensive but work best when the instructions are simple. Others are better at writing, coding, or solving difficult problems. The most advanced models can even manage long projects in which the next step is not always known ahead of time. Using these models effectively requires more than choosing the smartest one available. A better method is to break a project into tasks based on the kind of thinking each task requires. To do this well, a person should define the goal, separate the different kinds of work, choose the proper model for each task, create clear handoffs between tasks, verify important results, and save human attention for decisions that actually require it.
The first step is to define what the finished project is supposed to accomplish. This sounds simple, but it is easy to confuse a goal with a list of instructions. For example, imagine that a school wants to hold a charity fun run. The goal might be, “Hold a successful fun run that raises at least $10,000 for the local food bank.” That statement explains the desired result without pretending that every step is already known. The budget, advertising, registration, volunteers, safety plan, and event schedule are all parts of the project, but none of them is the project’s final purpose. Starting with a clear goal gives both people and AI models something against which later decisions can be measured.
The second step is to divide the project whenever the kind of thinking changes. This is more useful than simply dividing a large project into equal pieces. For example, investigating why last year’s fun run lost money requires a different kind of thinking from writing this year’s sponsor letters. Writing sponsor letters is different from sorting 2,000 registrations into age groups. All three tasks may belong to the same project, but they should not necessarily be given to the same AI model. A good project plan therefore asks, “What kind of thinking does this task require?” rather than only asking, “How big is this task?”
This distinction is important because a large amount of work is not always difficult work. Imagine that the school receives 5,000 registrations and needs to separate them into four age groups. That is a large task because it contains thousands of records, but the rule for sorting them may be very simple. In comparison, the school might have only three conflicting estimates for the cost of police traffic control. Determining which estimate should be used could require research and judgment. The smaller task may actually require the more capable AI. In other words, the amount of work determines how much must be done, while the kind of thinking determines which model should do it.
Once the work has been separated, the third step is to match each task with an appropriate model. One simple system uses four levels: Haiku, Sonnet, Opus, and Fable. Haiku is useful when the rules are already known and simply need to be applied repeatedly. Sonnet is useful when the task requires meaningful work, such as writing or coding, but the goal and important rules have already been established. Opus is useful when the correct answer is not yet known and the AI must investigate, compare evidence, or solve a difficult problem. Fable is useful when the AI must help manage the project itself by deciding what to investigate next, keeping track of progress, and adjusting the plan as new information appears.
A short way to remember these roles is that Haiku processes, Sonnet builds, Opus solves, and Fable manages. These are not rankings of how important the work is. A very important project may contain thousands of simple Haiku tasks and only a few Opus tasks. Fable may not be needed at all. The purpose is to use more powerful reasoning only where more powerful reasoning adds value.
The fourth step is to make each task large enough for the AI to work independently. People often make AI tasks too small because they are accustomed to giving computers one command at a time. A person might tell an AI to open a document, then tell it to find a section, then ask it to summarize the section, and finally ask it to compare that section with another document. If the AI could have performed all four actions safely, the person has spent unnecessary time managing the machine.
Instead, the person could say, “Read these two documents, find the sections concerning event insurance, compare their requirements, and report any conflicts.” The AI now has enough responsibility to make useful progress without stopping after every small step. This leads to an important rule: give the AI the largest task it can reliably complete without unnecessary human help.
However, the fifth step is to make sure the task is still small enough to check. Giving an AI complete freedom over an entire project creates a different problem. Suppose the school tells one AI to investigate last year’s event, decide what should change, create this year’s plan, prepare the budget, contact sponsors, and approve all expenses. If the AI makes a bad assumption during the original investigation, that mistake may affect every later decision before anyone notices it.
A better approach is to create checkpoints. The investigation should end with a set of findings. Those findings can be checked before they are used to create the new plan. The plan can then be approved before anyone prepares the final materials. The completed materials can be checked before they are used. A good AI task is therefore large enough to avoid constant supervision but small enough that its result can be verified before an important new stage begins.
The sixth step is to write a prompt that tells the model exactly what responsibility it has. A useful prompt does not need to be extremely long. It should answer five questions: What is the objective? What information does the AI need? What limits must it follow? What does a successful result look like? When should it stop and ask for help?
For example, Opus might receive this assignment: investigate why last year’s fun run failed to meet its fundraising goal, use the available budgets and event records as evidence, do not assume that high expenses were the only cause, identify the most likely causes, and stop before recommending changes to school policy. That prompt gives the model a problem to solve without telling it what conclusion it is supposed to reach.
Sonnet would receive a different kind of prompt. After the school approves the new event plan, Sonnet might be told to create the registration materials, volunteer instructions, sponsor letter, and event schedule according to the approved plan. It may decide how to organize those documents, but it should not quietly change the fundraising goal or invent a new safety policy. Haiku would receive even narrower instructions for processing registrations. Fable, if it were being used to coordinate the entire project, would receive the overall mission, the limits of its authority, and instructions about which decisions still require a person.
The seventh step is to create a clean handoff whenever work moves from one model to another. This is one of the easiest parts of a multi-model project to get wrong. A person may be tempted to give Sonnet the entire conversation that Opus used while investigating the problem. However, that conversation may contain incorrect guesses, rejected ideas, and conclusions that changed as new evidence was discovered. Sonnet does not need to repeat that journey.
Instead, the first model should prepare a short handoff containing five things: the objective, the facts that have been established, the decisions that have already been made, the limits that still apply, and the questions that remain unanswered. The evidence supporting those facts should still be available, but the next model should not have to read an entire conversation just to discover what was finally decided.
For example, Opus might finish its investigation by reporting that last year’s fun run had enough participants but spent too much on shirts and last-minute equipment rentals. It might also report that the evidence does not support the original theory that registration prices were too low. The school could then decide to limit shirt orders and require earlier equipment reservations. Sonnet would receive those established findings and decisions as its starting point. It would not need to repeat the investigation.
The eighth step is to move problems to a stronger model only when the work actually becomes more difficult. Suppose Haiku is processing registrations and discovers one that does not fit any approved age category. Haiku should flag the exception instead of inventing a new rule. Sonnet may be able to resolve the problem by applying the existing event instructions. If Sonnet discovers that two approved rules actually contradict each other, the issue should move to Opus for investigation and a recommendation.
The movement should also work in the opposite direction. After Opus resolves the conflict and the school approves a new rule, Opus should not spend its time applying that rule to 5,000 registrations. The work has become predictable again, so it can return to Haiku. This creates one of the simplest rules for managing several AI models: uncertainty moves up to stronger models, while predictable work moves back down to faster models.
The ninth step is to verify important work instead of simply trusting the model that produced it. Some results can be checked directly. A computer program can be tested. Totals can be recalculated. Records can be compared with their original sources. When direct evidence is available, that evidence is more useful than asking another AI whether the first AI appears to be correct.
Other questions are harder to test. An AI may have misunderstood a requirement, overlooked an unusual situation, or chosen a solution that creates a new problem somewhere else. When an error would have serious consequences, an independent AI reviewer can be useful. For example, an OpenAI model could be asked to examine the event plan specifically for weaknesses, while a Gemini model could be asked to study the original information and develop its own solution without first seeing the other model’s conclusion. If the models disagree, the goal should not be to hold a vote. The disagreement should identify a question that needs better evidence or human judgment.
The tenth and final step is to decide where human attention is actually worth using. Human time should be treated as a limited resource just like money or computing power. Choosing a cheaper AI model does not save much if a person must spend thirty minutes correcting it. Likewise, using the most powerful model for a simple sorting task wastes resources without providing much benefit.
Humans are most valuable when a decision requires responsibility, priorities, judgment, or authority. The principal may need to approve the event budget. A teacher may need to decide whether a safety risk is acceptable. The student committee may need to choose between spending more money on entertainment or giving more of the proceeds to the food bank. Those are meaningful human decisions. Finding the correct spreadsheet, sorting registrations, checking arithmetic, or comparing two versions of a schedule usually are not.
The fun run can now be seen as one project made from several different types of AI work. Opus investigates why the previous event failed to reach its goal. Humans use those findings to make important policy and budget decisions. Sonnet turns the approved plan into schedules, letters, forms, and other materials. Haiku processes the large number of registrations according to established rules. If the project becomes complicated enough that someone needs to track several workstreams, decide which unresolved problem should be addressed next, and keep the whole effort moving, Fable can manage that larger mission. At important checkpoints, tests, evidence, independent reviewers, and human approval are used to make sure the project is still moving in the right direction.
Breaking a project into AI-sized tasks is therefore not a matter of giving the easiest jobs to the weakest model and the hardest jobs to the strongest one. The process begins by defining the final goal and then separating the project whenever the kind of thinking changes. Each task should be large enough for the AI to work independently but small enough to verify before its conclusions affect the rest of the project. Prompts should clearly explain the model’s responsibility, and clean handoffs should carry established knowledge from one stage to the next. When uncertainty appears, the problem should move to a model capable of resolving it. When the work becomes predictable again, it should move back to a faster model.
Most importantly, the system should be designed around the value of human attention. Artificial intelligence is most useful when it removes routine work from people without removing people from the decisions that actually require them. The best multi-model workflow is not the one that always chooses the cheapest AI or the most powerful AI. It is the one that uses the right amount of intelligence at each stage, checks important results before moving forward, and allows people to spend their limited time on the choices where human judgment matters most.
