Writing for Literal Models

In 2024, the usual fix for a model that ignored a rule was to say it louder. In late 2026 the models are listening. Anthropic describes Claude Sonnet 5 as interpreting prompts “literally and explicitly,” and OpenAI says GPT-6 Astra is “better able to follow longer instructions, but can also be more sensitive to information in context.” Habits built for inattentive models now cause overreaction. See whether you can spot them before reading why.

Look at what stayed. The role line, the customer context, the competitor rule and the refund threshold are all fine. Context is never clutter, and a real business rule with its reason attached is exactly what a literal model needs.

Literal means literal

Anthropic says Sonnet 5 “does not silently generalize an instruction from one item to another, and it does not infer requests you didn't make.” The upside is precision: pipelines with carefully tuned prompts get predictable behavior. The cost lands on vague requests. Ask “can you suggest some changes to this function?” and, Anthropic notes, Claude will sometimes suggest changes rather than make them. Ask for a table in the first section and you get one table.

Volume is not priority

Prompts written to fix undertriggering on older models now overtrigger. Anthropic's guide: “Where you might have said ‘CRITICAL: You MUST use this tool when...,’ you can use more normal prompting such as ‘Use this tool when...’ instead.” It flags “If in doubt, use [tool]” as a cause of overtriggering too. OpenAI says the same of GPT-6 Astra: “overly specific guidance can now hinder results where it previously helped.”

Absolute words are read as absolute. Anthropic's Claude Fable 5.2 guide says the model “treats an instruction that contains ‘never’ as absolute,” and suggests “ask for approval before deleting files” instead of “never delete files” when you want the model to keep that responsibility. Broad approval rules cost you too: in Anthropic's testing, listing “sending messages to people outside this conversation” among the steps that need approval made the model hold back about half of its routine requests to colleagues.

Give the reason

Anthropic's favorite example: “NEVER use ellipses” works less well than “Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them.” In Anthropic's words, “Claude can generalize from the explanation”, so it can apply the same care to other symbols a speech engine would trip over. Its Fable 5 guide makes the same point for whole tasks under the heading “Give the reason, not only the request,” with a template:

I'm working on [the larger task] for [who it's for].
They need [what the output enables].
With that in mind: [request].

Rules for the release-notes writer:

Delete before you add

Prompts collect rules over time: one per incident, plus every fix for every model they ever ran on. Anthropic says skills written for earlier models are “often too prescriptive” for Claude Fable 5 and “can degrade output quality,” and advises removing older instructions if the default does better. OpenAI's advice for GPT-6 Astra: “frequently revisit each instruction and ask yourself whether it's still needed.”

Some things always stay. Context stays: audience, product facts and the reasons behind rules. Real business constraints stay, ideally with their reasons. Confirmation steps for risky or irreversible actions stay; Anthropic repeats that in nearly every model guide. Format rules your parser depends on stay, and are better still when enforced in code.