Symbols a Newcomer Cannot Read Fail, and AI Can Fix Them── Whether a symbol works is decided by the first-time reader, so have the AI list every one and fix them until none are left

01The operator failed a design page that still carried symbols and coined terms

In September 2026, the operator had an AI rebuild a design page that explained how a review system worked. The request was to drop the description of the older version and rebuild the page around the newer one alone.

After reading the rebuilt page, the operator answered with a verdict, not a comment. The page contained symbols and coined terms that someone reading it for the first time could not follow, so it failed. The operator then had the AI rewrite it as a plain version.

The standard behind that verdict was neither the operator, who commissioned the page, nor the AI, which wrote it. It was the person opening the page for the first time. The makers understood every term. Even so, one term a newcomer could not read was enough to stop the page.

The question for this episode is whether that standard can become a pattern anyone can use.

02Whether a symbol works is decided by the first-time reader, not the writer

It is hard for the people who made a document to judge its symbols. They know what each one means, so the page reads clearly to them. A reader who already understands a term proves nothing by understanding it.

So the standard moves to the reader. Can someone opening this page for the first time get the meaning from the surrounding sentences alone? If not, the term needs fixing.

The word "fail" matters too. Treat unclear wording as a matter of degree, and a term that is "a little hard to follow" gets to stay. Treat it as pass or fail, and the page does not pass while a single unreadable term remains.

03One unreadable term makes the reader stop or guess

A newcomer who meets a term they cannot read has two options. They can stop reading, or they can guess the meaning and carry on.

If they stop, the page has failed at its job. If they guess, something worse can happen. A wrong guess gives no warning. The reader moves on to the next decision with a mistaken picture.

The makers do not see this loss. Readers rarely report what they did not understand. All the makers have is the fact that the page went out.

Experiments in economics show that people with more information misjudge how less-informed people will think. In a 1989 paper, Camerer and colleagues reported that people could not set aside what they knew, even when doing so would have paid. They named this the curse of knowledge. A maker who finds their own labels easy to read is showing one instance of it.

04The targets are abbreviations, labels, working names and unexplained coinages

"Symbols and coined terms" covers a lot, so this episode draws a boundary. A symbol here is an abbreviation of a longer name, or a number or version mark added to keep pieces of work apart. A coined term is a word the maker invented during the work, one that appears in no dictionary and no industry glossary.

Some terms fall outside the boundary. A word the reader uses every day can stay, even if it is technical. A page written for promotional material reviewers does not need to explain "material review". The test is not how hard a word is. It is whether this reader uses it in ordinary work.

There are three ways to fix a term. Replace it with an everyday word. If it must stay, explain it in one sentence where it first appears. If explaining it would not help the reader, delete it.

The Clear Communication Index published by the US Centers for Disease Control and Prevention (CDC) points the same way. It asks writers to choose the words their audience uses every day. When an unfamiliar term is needed, it should be explained where it is used, in the same sentence or right after. Abbreviations should be spelled out, explained and kept few. The guide adds that testing the material with the audience shows which abbreviations make sense to them.

05In pharma, it pays off in documents that leave the team and in AI drafts

Promotional material review and medical affairs produce many documents that leave the team. Review comments go back to the department that made the material, or to an outside agency. Internal procedures go to newly arrived staff. In each case, the reader does not share the makers' abbreviations.

Review records tend to collect abbreviations and numbers that only make sense inside one team. If the person receiving a comment misreads an abbreviation, the fix misses its target. Then the same comment comes back again.

The problem is more likely when an AI writes the draft. Over a long piece of work with an AI, the working names and numbers made along the way slip into the finished product. When you build a review process or an internal procedure with AI, decide first that someone outside the team will read the finished page. Then judge pass or fail by that person's standard.

The same check works for material aimed at physicians and pharmacists, with a different reader as the standard. Terms that healthcare professionals use daily can stay. Abbreviations made up inside the company, or names born during product planning, are new to healthcare professionals too.

06Symbols survive in AI documents because the AI writes from inside the work

There are two reasons unexplained symbols stay in documents an AI writes.

The first is that working names flow straight into the output. In long work, items get numbers and temporary names so they can be told apart. Those names were defined earlier in the conversation, so for the AI they count as explained. For a reader who never saw that conversation, they have no definition.

The second is that the AI has not been told who the reader is. With no reader named in the request, the only reader the AI has evidence about is the person who asked. That person knows the context of the work, so abbreviations read fine to them.

Anthropic's published prompting guidance suggests treating Claude as a capable new employee who does not know your norms or ways of working. It then gives a test: show your prompt to a colleague with little context on the task, and if they would be confused, Claude will be too. The same test works in both directions, for the requests people give to AI and for the documents AI gives back to people.

07From tomorrow: have the AI list every symbol, then test with a reader who lacks the context

There are four steps.

1. Name the reader in one sentence. Write something like "a colleague from another department opening this page for the first time", and give it to the AI.
2. Have the AI list every symbol and coined term. Abbreviations, numbers, working names and coinages, each with where it first appears and whether it is explained there.
3. Decide the fix term by term. Have the AI choose one of three: replace it, explain it in one sentence where it first appears, or delete it.
4. Test with a reader who lacks the context. Give only the page to an AI in a fresh conversation, or to a colleague who was not involved, and ask them to list every term they cannot follow. If even one comes back, return to step 3.

Step 4 uses a fresh conversation for a reason. An AI that has been doing the work knows the definitions of its working names. Ask it in the same conversation whether any terms are unclear, and it will tend to miss the names it made up itself.

The pass line is zero terms returned in step 4. "Mostly understandable" does not pass. Telling the AI this line at the start also cuts down the number of rewrites.