Human → Agent

You are writing for a stochastic reader. Every instruction is a lever on how predictably it behaves. The aim is predictability — the agent taking the same process each run, not producing the same output. Each rule below buys some.

Prompt the positive

State the behaviour you want. Prohibition backfires: don't think of an elephant names the elephant and makes it more available, not less. Phrase the target so the banned behaviour is never spoken. Keep a "don't" only as a hard guardrail you cannot phrase positively, and even then pair it with what to do instead.

Leading words

A leading word is a compact concept the model already holds from pretraining — tracer bullet, fog of war, tight loop, receipts. One well-chosen word anchors a whole region of behaviour in a token, because it recruits priors the model already has. Repeat it and its meaning accretes; use the same word across your prompts, docs, and code so the agent links them and fires the right behaviour.

Reach for one wherever you have spelled a quality out three times: "fast, deterministic, low-overhead" is one idea restated — collapse it to tight.

Cut no-ops

A no-op is a line the model already obeys by default: it costs context and buys nothing. Test each sentence in isolation — does it change behaviour versus the default? "Be thorough", when the model is already thorough-ish, is a no-op; the fix is a stronger word (relentless), not more words. When a sentence fails, delete the whole sentence rather than trim it.

One source of truth

Keep each instruction in exactly one place. The same rule in two files (duplication) rots the moment one copy changes, and inflates the rule's apparent weight. When you change the behaviour, you want a one-place edit.

Say how "done" is checked

Give the agent a completion criterion it can check — done from not-done, observable. Where it matters, make it exhaustive: "every changed file accounted for", not "list the changes". A vague criterion invites the agent to stop early.

Be concrete

Name the file, the command, the flag, the value. Ambiguity is where a stochastic reader diverges from your intent. A specific instruction beats a general one, and an example beats a description.

Prune

Instructions accumulate — adding feels safe, removing feels risky — until stale layers pile up (sediment) and the whole thing is too long to follow (sprawl). Prune on a schedule: drop no-ops, collapse duplication, and push rarely-needed detail into a linked doc so the instruction that is always needed stays legible.