Controlling LLM Prose

Explanation

The Reader’s Memory Budget

Comprehension as a working-memory problem

Every reader holds a fragile handful of items in working memory. Miller’s famous “seven, plus or minus two” (1956) set the popular figure, but later work—notably Cowan (2001)—puts the real capacity for unrehearsed chunks closer to four. Either way the budget is small: prose that ignores it forces re-reading; prose that respects it feels effortless. Most “clarity” advice is really memory management in disguise.

Three moves do the heavy lifting. The given–new contract opens each sentence with something already known and ends on the new point, so the reader always has a hook to hang new information on. BLUF spends the most valuable position—the opening—on the conclusion, so a reader who stops early still leaves with the point. Chunking groups items into a few labeled units that fit the budget instead of one undifferentiated list.

These are the highest-leverage techniques for writing that must be understood the first time, because they reduce the number of things the reader must hold at once.

Example

Chunking eleven settings into three memorable groups.
  • Ungrouped (11 items)

    Configure host, port, username, password, timeout, retries, TLS, cipher suite, log level, cache size, and region.

  • Chunked (3 groups)

    Connection: host, port, region. Auth: username, password, TLS, cipher suite. Behavior: timeout, retries, cache size, log level.