What we check for
9 dated patterns · last revised 26 August 2026
Prompt technique goes stale. Advice that genuinely helped in 2023 is now inert on models that reason on their own, and some of it actively costs you quality. Prompto scores every prompt against the list below and names the specific pattern it found.
Two kinds of date appear here, and they are not the same kind of claim. When a pattern stopped being good advice is given as an era, because nobody can honestly cite the day a technique stopped working. When we added or revised an entry is exact, because that is our own record and we can stand behind it.
Asking for JSON in prose
added 26 August 2026Sound through 2023
- What changed
- Providers shipped structured outputs and tool schemas that constrain generation directly, rather than requesting a format and hoping.
- What it costs now
- Asking in prose still produces the occasional malformed response that a schema would have made impossible.
json-by-request
Assistant-turn prefill to force a format
added 26 August 2026Sound through 2024
- What changed
- Putting words in the assistant's mouth to lock in a format has been restricted or removed on newer frontier models; support now varies by model rather than being assumed.
- What it costs now
- Where it has been withdrawn this is an error, not a technique. A prompt relying on it fails outright rather than degrading.
assistant-prefill
Chain-of-thought incantation
added 26 August 2026Sound through 2023
- What changed
- Models began reasoning through multi-step problems without being asked, and reasoning-focused models made the amount of thinking a setting you control rather than something you coax.
- What it costs now
- Inert at best. On a model that manages its own reasoning it can work against that control, and it spends context restating what the model already does.
chain-of-thought-incantation
Defensive boilerplate
added 26 August 2026Sound through 2023
- What changed
- The failure modes much of this guarded against — inventing sources, ignoring the requested format, breaking character mid-answer — became markedly rarer.
- What it costs now
- Paragraphs defending against behaviour the model no longer exhibits, displacing the part of the prompt that describes the actual task.
defensive-boilerplate
Emotional pressure, threats and bribes
added 26 August 2026Sound through 2023
- What changed
- The findings that emotional framing improved output came from a narrower generation of models and did not hold as instruction-following improved.
- What it costs now
- Noise. It occupies context that could have carried an actual requirement, and offering a tip the model cannot receive communicates nothing.
emotional-pressure
Fixed thinking-token budget
added 26 August 2026Sound through 2024
- What changed
- Effort controls replaced fixed reasoning-token budgets on the models that expose the setting at all.
- What it costs now
- A hardcoded budget either wastes the allowance on an easy task or truncates reasoning on a hard one. An effort level lets the model size it to the work.
fixed-thinking-budget
Instructing the model not to reason
added 26 August 2026Sound through 2023
- What changed
- Reasoning moved into a separate channel that the answer is drawn from, rather than being something the model wrote inline in its reply.
- What it costs now
- Instructing against it tends to push the working into the visible answer rather than removing it — the opposite of what was asked for.
suppress-reasoning
Repeating an instruction for emphasis
added 26 August 2026Sound through 2023
- What changed
- Long-context handling improved, and models stopped dropping instructions that were stated once in a short prompt.
- What it costs now
- Duplication invites contradiction. The second statement is rarely worded identically to the first, leaving the model to decide which one governs.
repeated-instruction
Role inflation
added 26 August 2026Sound through 2023
- What changed
- Instruction-following improved to the point that capability no longer had to be unlocked by flattering the model into character.
- What it costs now
- A role still usefully sets voice and audience — that use is not a defect. As a quality lever it does nothing, and the superlatives crowd out the specifics that would have helped.
role-inflation
Think something here is wrong?
This list is a claim about how current models behave, and claims like that age. If one of these is out of date, or a technique you rely on has been called dead unfairly, tell us — a rubric nobody corrects is exactly the problem it exists to solve.