Not knowing how to say it is the main barrier

What it is

Wanting an apology letter that can go straight to a customer is not the same as knowing to type “formal tone, no admission of legal liability, under two hundred words, keep the original order number.” The prompt formulation barrier is getting stuck compiling intent into the string a model will actually use — not failing to notice that the product exists. Even when the entry has already hinted at a range, people can still freeze at a blinking caret.

Graphical interfaces ask you to find a control. Open input asks you to invent a sentence that is not yet there. There is no menu to scan, only a search through language.

Why it happens

Norman’s gulf of execution here is two vocabularies: intent language and prompt language. Intent language is task words — apologize, revise the contract, align the headers. Prompt language is the dimensions the model actually tracks: role, constraints, format, counterexamples, output shape. Users do not already own a compiler between the two. When you speak to a person, missing slots get filled by follow-up questions. When you speak to a model, missing slots are filled by the mode of the training distribution, so “write an apology” becomes a correct and unusable sample letter.

There is a social cost on top. Freezing at an empty box in front of other people looks more like incompetence than hunting through a menu. Some people leave to search “how to ask AI,” moving the barrier out of the product and into a search engine.

Studying it

Give a concrete task (“turn these meeting notes into three follow-up emails”). Allow questions about the product, withhold examples. Record time from briefing to first submit, mid-flight rewrites, and think-aloud meta-questions (“do I mention tone,” “do I cap length”). Independent variables: whether the task already sits close to prompt language, whether constraint chips are offered, prior prompt-engineering experience. Dependent variables: start latency, first-draft usability, kinds of meta-question.

Split “cannot formulate” from “formulated, but the model is noisy.” If the same intent succeeds when dropped into a deterministic template, the barrier is phrasing, not sampling.

Where it stops holding

People who already speak a professional jargon — a lawyer writing “revise the defense to our position, introduce no new facts” — have prompt language overlapping intent language, and the barrier drops. Chat, translation, and summarization sit close to everyday speech, so starts are faster. Image generation, code repair, and data transforms hide constraints in jargon and raise the barrier highest. This entry does not explain why paraphrases diverge, nor whom people blame after a miss.

Applying it

• Put the missing slots beside the caret, not a sample essay: object, tone, length, forbidden moves. Checkboxes or short fields cut compilation cost while keeping the original intent sentence.
• Allow a very short intent first; let the system fill slots with visible follow-ups rather than silently using the mode.
• Do not make “write better prompts” the user’s homework. A product that depends on users teaching themselves prompt engineering has not finished the compilation at the entry.
• Check: people who cannot write prompts but can do the task (a support agent who can hand-write that apology). If they stare at the empty box for more than half a minute without a first sentence, or the first sentence omits constraints they would always put in a handwritten version, the barrier is at formulation.

Related

• Same group: L2.01.1 Open input does not cue the range of capability · L2.01.3 Differences in wording produce differences in results · L2.01.4 An empty box conveys no boundary; the first sentence is a guess · L2.01.5 Open input steers failure attribution toward “I said it badly” · L2.01.6 Synonymous phrasings yield different results, so users invent phrases to memorize · L2.01.7 Openness makes functions non-enumerable; the product can no longer show a complete capability list · L2.01.8 Error messages for open input stay vague because the system does not know what the user meant to do
• Nearby: L2.03 Examples and Template Guidance · L2.15 Ambiguity and Clarifying Follow-ups · L2.02 Discoverability of What Can Be Said
• Search terms: prompt formulation barrier · gulf of execution · intent-to-prompt compilation