TELeR
Turn · Expression · Level of detail · Role — Not a template but a measuring scale: a peer-reviewed answer to "how detailed does a prompt need to be", with seven rungs instead of an adjective.
Use it for
Benchmarking and quality control. When you are comparing models or prompt versions and need to show the comparison was fair, or standardising prompt quality across a team.
Avoid it for
Writing your first prompt. TELeR tells you how much to specify, not what to say, so on its own it produces nothing. The authors also scope it to complex tasks explicitly.
The slots
Each letter is a question the prompt has to answer.Turn Optional process step
One message, or a dialogue?
Choose single-turn when the run must be replayable and comparable, multi-turn when the task genuinely needs back-and-forth. Then record which, because it changes what a result means.
Example: Single-turn: the whole brief goes in one message so all three candidate models receive identical input.
Expression Optional process step
Questions or instructions?
Pick one style and hold it constant across every prompt in the comparison. Mixing the two confounds the result.
Example: Instruction style throughout: "List every function that mutates shared state", never "Which functions mutate shared state?"
Level of detail Optional process step
How specified is the directive, on a scale of nought to six?
Climb the ladder: one is a one-line goal; two adds what the output must cover; three splits it into named sub-tasks; four adds criteria for a good answer; five adds retrieved source material; six adds a demand for justification.
Example: Level four: goal stated, the review split into five named sub-tasks, plus "a good review flags severity and cites a line number for every finding, under 600 words".
Role Optional process step
Was a system role set, yes or no?
Binary. Record it, because an undefined role is a legitimate control condition rather than an oversight.
Example: Role defined: system message set to "You are a senior reviewer enforcing the team style guide."
Strengths
- The only citable, peer-reviewed answer to how much detail a prompt needs, which makes prompt quality auditable rather than a matter of taste.
- The detail ladder is cumulative and unambiguous, so two people grading the same prompt will usually land on the same number.
- It exposes two variables the popular templates ignore entirely: whether the prompt is one message or a dialogue, and whether a system role was set at all.
Weaknesses
- It produces no prompt text. Used alone it is inert — you still need a template underneath it.
- The authors scope it to complex tasks and say the levels do not apply to simple ones, which is most everyday work.
- It predates reasoning models. Its top rung is asking the model to justify its answer, which a thinking model now does by default, so the top of the scale has partly collapsed.
Sources and grounding
- 1TELeR: A General Taxonomy of LLM Prompts for Benchmarking Complex TasksSantu & Feng, Findings of the ACL: EMNLP 2023The peer-reviewed publication of record.
- 2TELeR (preprint)arXivFull text, including the definitions of levels nought to six.