cankun.me
Oct 11, 2026 · 8 min read /

Taste: The Bar After Correct

Agents now produce work that passes every check and still reads generic. Taste is the next bar. Paul Graham's essay "Taste for Makers" argues that taste can be wrong, and that argument is what lets a skill teach it.

A sculptor stands back from a tall, smooth stone form, with a row of rougher earlier versions receding behind it

taste is a skill for work that is correct but unsatisfying. It gives the model a method and a set of tests for judging a design, a diagram, a document or a code structure, and for making it better. The skill paraphrases Paul Graham's 2002 essay “Taste for Makers”, and adds one technique of its own.

This essay explains why a skill about taste can exist, the case that shaped it, and how the skill was made. The best evidence comes from the skill's own first draft, which failed in the way the skill is meant to prevent.

Correct is not good

My agents now often produce work that passes every check that I write. The tests pass. The page builds. A script confirms that no label overlaps another. And the result still reads generic: a page that looks like every other page, a diagram that is a stack of boxes, a paragraph that says nothing wrong and nothing particular.

A check can measure only what I can write as a rule. Generic work satisfies rules easily, because rules describe the floor. Above the floor there is a second bar: the work is good for the person who uses it. I call that bar taste.

The word causes a problem. If taste is personal preference, then nobody can teach it, and a skill about it is only a list of my likes. So the skill needs an argument that taste is more than preference. Graham's essay supplies one.

Graham's argument: taste can be wrong

Graham starts from an observation about makers. If taste were only preference, then everyone's taste would already be perfect, and nobody could improve. But makers do improve. When they look back, they see that their earlier taste was worse, not merely different. So taste can be wrong, and the difference between good and bad work is something that a person can study.

He adds a second point. The same qualities mark good work in many fields: a short proof, a well-composed painting, a sound building, a clean program. So a lesson from one field can help to judge another. Most of the essay is a list of these qualities: good design is simple, timeless, suggestive, often slightly funny, hard, and so on.

The two points together make taste teachable to a model. A model cannot inherit my preferences. It can apply principles that recur across fields, and it can give evidence for each judgment.

The essay also warns against a sequence that I had in mind. I thought of correctness first and taste second. Graham's examples point the other way: a sense that something is ugly is often how a real flaw is found. He mentions Copernicus, who was bothered by the awkward machinery of the older astronomy. So the skill does not treat taste as polish after the real work. A diagnosis of ugliness can be a diagnosis of an error.

What the skill does

The skill has five steps and twelve tests.

The steps are the method:

  1. Diagnose. The model says what is ugly or unsatisfying, element by element, and backs each complaint with knowledge of the field, a comparable example, or how the thing is actually used.
  2. Name the problem. The model states who uses the thing, at what moment, and what they need then. It marks each part that it assumed.
  3. Find the organising idea. The model names the one idea that the work hangs on, and makes the hierarchy show it. This step is mine, not Graham's, and the skill marks it as an addition.
  4. Revise. The model makes a substantive revision aimed at the weaknesses from step 1 and compares it with the original. Graham notes that good work is usually redesign.
  5. Test. The model applies the tests that bear on the work, each with evidence.
Five steps arranged in a loop: diagnose, name the problem, find the organising idea, revise, test. An arrow from test returns to diagnose.
The five steps. A test that fails sends the work back to diagnosis.

The tests are Graham's principles written as questions. For example, the test for simplicity asks whether every element serves the need from step 2, or makes a stated contribution such as pleasure or identity. The test for symmetry asks whether repetition simplifies the work or stands in for a decision. Each step has a completion criterion, so the model knows when the step is done. Step 1 is done when each complaint names an element and its evidence.

Where taste bites: a diagram that passed every check

The third step comes from a real case from a few days before I wrote the skill. Agents drew diagrams of multi-step analysis plans: what runs, in what order, and where a person must decide. A script checked each drawing for what a script can check: every step was present, and no text overlapped or ran out of its frame.

The first round of drawings passed every check. Each one read as a document. Every step was a box with a paragraph in it, and every plan had the same shape: a stack of white cards. Nothing in the drawing said what this plan was about.

In the skill's first step, the diagnosis names an element and its evidence. The element was the boxes of text. The evidence was the use: the person looks at the drawing to see where work runs and where they must act, and the full description of each step is already in a side panel. So the drawing repeated the panel and hid the shape.

The second round asked each drawing to commit to one visual idea. The best one was a subway map. Each line was a place where work runs. The main line stepped down like a staircase as the work moved from one place to the next. A part of the work that took a different route left the main line as a dashed detour. A decision became a gate that the person stands at. The step that the plan hinges on was drawn larger, and housekeeping steps were small.

Two schematic drawings of the same plan. Left: five equal text boxes stacked in a column. Right: the same steps as stations on two subway lines, with a dashed detour and a gate before the final step.
The same plan twice, as a schematic. Both versions pass the same layout checks. Only the right one shows where work runs and where the person decides.

Both rounds passed the same checks. Only the second could be read at a glance. When I looked at it, I also found that a text column beside the drawing was now unnecessary, because my attention stayed on the map. That is the simplicity test in practice: the column no longer served the need.

The examples in step 3 of the skill (a subway map, a gate where the user stands, one table that threads through every output) come from this case.

How the skill was made

I asked Claude to draft the skill from the essay. Claude used the levers from my writing-for-agents skill: steps with completion criteria, a description that starts with its leading word, and instructions in positive form.

Then a GPT model (GPT-6 Astra) reviewed the draft independently. It fetched the essay, compared each principle with the skill, and checked how an agent would behave with each line. It found no copied passages. Its verdict was still “fix,” for one main reason: the draft had turned Graham's qualified observations into rigid rules.

Four examples:

  • Size. The draft said that the part the work hinges on must be the largest. The rule echoes the diagram case above, where a larger step was the right choice for that plan. As a general rule, it is wrong: Graham says nothing of this kind, and the rule can damage a design. In a chart, size often encodes a quantity. On a form, the main button needs prominence, not the most space. The fixed step says that the central task must be recognisable at first look, by position, weight or size as the medium allows.
  • Ornament. The draft told the model to strip ornament. Graham's objection is narrower: ornament is fine until it disguises a weak form. The fixed test removes decoration that hides a weak structure, and keeps decoration that adds to a sound one.
  • Redesign. The draft required a second version with a different central idea. Graham says that experts expect to throw early work away, and that good tools make mistakes easy to fix. He does not say that each revision must replace the idea. The fixed step asks for a substantive revision of the weaknesses that step 1 found. The central idea may survive.
  • Nature. The draft listed a map, a ledger and a hull built like a ribcage as “proven forms” under the principle about nature. Graham's point is specific to natural forms, and he adds a warning: early aircraft that copied birds failed because the technology could not yet support the copy. The fix split the test in two. One test covers natural forms and asks why the borrowing fits the current constraints. The other covers human precedents and asks the model to name the precedent and why its answer applies.

The review also found a dropped qualification at the end of the essay. Graham says that great work often starts from dissatisfaction, but dissatisfaction is not enough: the maker needs enough knowledge of the field to see what to fix. The first draft kept the dissatisfaction and lost the knowledge. The fix moved it into step 1, which now requires evidence for each complaint.

I think this episode is the best argument for the skill. The draft was a correct-looking document. Every line was clear, positive and checkable. And it was wrong in the way that generic work is wrong: it replaced judgment with rules that are easy to check. The lever that helped most, a completion criterion for each step, also pulled the draft toward rules, because “the largest element” is easier to check than “recognisable at first look.” The fix kept Graham's qualifications and put the demand somewhere else: each judgment now needs evidence.

A second model found the problem because it read the source with different blind spots from the model that wrote the draft.

What the skill cannot do

The skill does not supply knowledge of a field. Graham notes that good work tends to happen where capable people work on related problems. When the model lacks the knowledge to judge, the skill tells it to find the field's strong work or a practitioner's criticism, and to judge against that.

The tests are questions, not scores. A model can still answer a question with a weak reason. The completion criteria ask for evidence, which makes a weak reason visible, but they do not make it impossible.

The skill is new. I have the design reasons, the review, and the diagram case. I do not have a measured result yet.

How to use the skill

taste is in my skill collection, and the skill instructions are public. Read Graham's essay too. The skill paraphrases it, and a judgment that turns on his argument should go back to the source.

For a review, I start with this request:

Use taste on this. Say specifically what is unsatisfying and why, name who uses it and when, find the idea it should hang on, and propose a concrete revision. Test the revision against the principles that apply, with evidence.