TL;DR — The model assignment table needs to be updated, but saying that “difficulty is always fixed” goes too far. My approach is to describe the task, consequences of error, and acceptance requirements first, then place the current model and settings in a separate dated table. Separating the two layers can reduce rewriting, but both still need to be reconsidered when conditions change.

Some time ago, I asked Claude to help assign my work. It produced a table: difficult work went to Opus, medium work to Sonnet, and high-volume, regular work to Haiku. That was its recommendation at the time, not a ranking of today’s models offered by this essay.

I then asked what should happen to the table if models continued to change.

A task may require more resources today and have another workable approach tomorrow. But changes in its data, tools, or requirements can also make it harder. An assignment table needs more than names; it needs the reasoning behind its choices and a way to reexamine them.

Why Does an Assignment Table Quietly Expire?

Writing a model name directly into every rule can turn a choice that made sense at one moment into a habit that nobody later rechecks. A tool’s continued ability to execute does not mean that the arrangement remains appropriate.

Capability is not the only changing factor. Task scope, input quality, prices, available permissions, and delivery requirements can also change. Expiration does not always appear as overspending. It can also mean missing a new risk or performing insufficient checks.

So the question I want answered is not merely whether the table can still be consulted. It is: what recent evidence was used to update this row?

Separate the Task Description from the Model Binding

Anthropic’s article on workflow design describes routing and orchestration patterns. These are useful arrangements to consider, but they do not endorse my classification system or any current model choice.

My design has two layers. The first describes the work and its acceptance requirements. The second records who currently receives it and which settings are used. The first layer does not have to be rewritten whenever model names change. The second is deliberately easy to replace.

For example, correcting punctuation and rewriting a claim that affects a decision are both forms of text editing, but they require different checks. This does not mean that the latter necessarily requires harder reasoning; the consequences of error and the allocation of responsibility are different.

Moving model names out of the decision rules can reduce vendor lock-in, but it cannot guarantee that the rules will never need revision. When new types of work, evidence requirements, or permission conditions appear, the classification should change as well.

Consider Difficulty, Risk, and Room for Judgment Separately

I originally treated the cost of error as the most important axis of difficulty. I now separate them. Difficulty concerns the challenge involved in completing the work. Consequences concern the impact of an error. Risk assessment must also consider the likelihood of error and the available controls. They cannot all be compressed into one score.

An operation that looks simple can still require stricter confirmation because it is irreversible, affects other people, or has a broad impact. A difficult exploration that produces only a disposable draft may not require the same release gate.

The cost of error is not permanently fixed either. The context of use, scope of dependency, reversibility, and presence of human review can all change the consequences. A model’s relative difficulty in completing a task can change, and so can the conditions of the task itself.

I therefore record separately what reasoning and data are required, which tradeoffs exist, what an error would affect, and who is responsible for acceptance. A single high, medium, or low score cannot replace those questions.

Recalibration Is Not Complete When the Newest Model Is Installed

I once had an upstream dependency fail in a scheduled workflow. The process stopped silently, and I did not notice until some time later. The experience made me care about another question beyond “is it still alive?” Does what it now produces still serve its purpose?

I keep a set of representative tasks drawn from real work and state the acceptable result before comparing candidate arrangements. About ten tasks can be a small-scale starting point, but one successful run cannot establish that every task is stable.

The check should record the version, data, settings, and acceptance conditions. A cost comparison should also include waiting, rework, review, and integration. A new model generation does not guarantee that a cheaper choice will pass, nor does it mean that all work should immediately be moved.

The date on the table creates a traceable record; it does not prove that the table is correct. Incomplete tests and uses that still raise concerns should be preserved alongside it.

Several Roles Can Temporarily Point to the Same Model

I allow multiple roles in the binding table to use the same model so that work is not divided artificially merely to preserve tiers. Whether and when the number of roles will converge cannot be declared inevitable in advance.

Separate roles, different models, and effective verification are not the same thing. When another perspective is needed, different data, prompts, or review arrangements may help. Merely opening another window does not guarantee independent judgment.

This continues the question I have explored in Dividing Work Across AI Windows: Materials, Permissions, and Verification and Looking Back at Long-Task Workflows Through the Release of Claude Fable 5: the value of dividing work lies in whether it helps complete the work, not in how many models are used.

An Exit Clause Is Also Part of the Design

Anthropic’s account of its multi-agent research system discusses the costs and limits of collaboration. It describes the experience of a particular system. It is not a fixed token multiplier for every workflow, nor does it establish that all programming work is unsuitable for division.

For my system, whether dispatch should remain depends on the quality it adds, its cost, and its maintenance burden. Vendor-provided routing can be a reason to reevaluate the arrangement, but controllability, permissions, data, and acceptance requirements still matter. A new feature is not, by itself, a reason to retire an existing process.

I previously retired an automated optimization system whose value did not justify its maintenance cost. That experience reminds me that a tool does not need to remain forever. But when it is retired, its dependencies should be explained, necessary data should be preserved, and the work should have another path forward.

Appendix: A Checklist You Can Fill In

Begin with three to five kinds of work you perform regularly. For each one, write down four things: its input and scope, the data and reasoning it requires, the consequences of error, and who will use which standard to accept the completed work.

Then create a dated binding table that records the role, model version, available settings, tools and permissions, and the reason for the choice. The table can offer recommendations. Actual cross-model execution, paid use, or publication must still follow the existing authorization.

Next, prepare representative tasks and counterexamples. Record what passed, what failed, and which scope remains untested. During recalibration, compare the complete delivery cost. Finally, write down the conditions for reevaluation or retirement rather than hiding them somewhere nobody will revisit.

This checklist is a starting point for design. Which models, tools, and permissions are actually available must still be confirmed in your own environment. A window can propose an assignment plan; that does not mean it can also execute the work across models according to the table.

Four-panel comic: The model assignment table needs to be updated, but saying that “difficulty is always fixed” goes too far. This does not mean that the latter necessarily requires harder reasoning; the consequences of error and the allocation of responsibility are different. A single high, medium, or low score cannot replace those questions. About ten tasks can be a small-scale starting point, but one successful run cannot establish that every task is stable. Separate roles, different models, and effective verification are not the same thing. The date on the table creates a traceable record; it does not prove that the table is correct.

Preserve the Method While Allowing It to Change

After expanding work across multiple windows and tasks, I once spoke with a friend about a feeling we shared: output became faster, while keeping track of context became more demanding. Both of us had started using handwritten notebooks again to organize our direction.

This does not mean that handwriting is necessarily better than tools. It reminds me that dispatch also consumes human attention. The amount saved on model fees cannot, by itself, determine whether an arrangement is worth preserving.

I would now replace my original principle that “difficulty is a constant” with something more practical: give tasks and acceptance requirements a clear place, make model choices rebindable, and preserve an entry point for correcting both. More reflections on workflows appear under AI and Human Order.