# Superseded Techniques Lookup for techniques the older prompt-engineering literature treats as core. Each entry answers one question: **should I reach for this?** Most answers are no, but the reason matters — a technique absorbed into model defaults needs different handling from one relocated to another surface, and both differ from one nobody has re-measured. Status describes models that reason natively, follow instructions literally, and call tools without being taught to. On a model lacking any of those, the older recommendation may still apply — check before assuming. Status vocabulary: - **ABSORBED** — the model does it natively. Adding the instruction now compounds with existing behavior. - **RELOCATED** — still correct, but belongs on a different surface: harness, interface, or code. - **NARROWED** — survives in a specific case; the general recommendation does not. - **UNEVIDENCED** — no measurement located either way. Absence of evidence, not evidence of death. - **REMOVED** — returns an API error. --- ## Reasoning elicitation - **Chain-of-thought ("think step by step")** — SOLVED: models that would not decompose a problem unless told to. STATUS: ABSORBED, and RELOCATED for depth control. IN PRACTICE: reasoning depth is the effort setting. Vendor guidance for reasoning models advises against generic step-by-step instructions. STILL VALID WHEN: the reasoning is itself part of the deliverable — you want the steps in the output because a human will read them. - **Self-consistency (sample N, majority vote)** — SOLVED: high-variance single samples on arithmetic and symbolic tasks. STATUS: UNEVIDENCED. IN PRACTICE: no ablation located. Do not retain as a default recipe. STILL VALID WHEN: you already run N samples for another reason and voting is free. - **Tree-of-Thoughts** — SOLVED: single-path reasoning that could not backtrack. STATUS: UNEVIDENCED as a prompt pattern; RELOCATED in practice. IN PRACTICE: parallel exploration happens as native multi-agent decomposition rather than one prompt simulating a search tree. STILL VALID WHEN: nothing located supports reaching for it first. - **Reflexion (self-critique and retry)** — SOLVED: no mechanism to learn from a failed attempt. STATUS: ABSORBED for the generic form; NARROWED for the structured form. IN PRACTICE: these models catch and fix their own mistakes without prompting, and "double-check your answer" compounds into over-verification. STILL VALID WHEN: long-horizon runs, where a **separate fresh-context verifier subagent** with its own rubric outperforms self-critique. That is an architecture, not a sentence in a prompt. ## Demonstration - **Few-shot / multishot examples** — SOLVED: format and label ambiguity. STATUS: NARROWED, and split by surface. IN PRACTICE: still recommended for output format, tone, and genuinely ambiguous semantics — three to five, wrapped in tags. But for **tool behavior** they constrain the model to the demonstrated exploration space; encode that in parameter names, enums, and the tool description instead. STILL VALID WHEN: the target is a style or shape the model cannot infer from a description. - **Generated knowledge prompting** — SOLVED: models that lacked recall on the task domain. STATUS: ABSORBED. IN PRACTICE: the model retrieves or reasons its way there, or a tool does. STILL VALID WHEN: nothing located. - **Active prompting / directional stimulus** — SOLVED: example selection and steering for weak in-context learners. STATUS: UNEVIDENCED. IN PRACTICE: no work located. STILL VALID WHEN: nothing identified. ## Framing - **Role / persona prompting ("you are a senior engineer")** — SOLVED: register and tone drift. STATUS: NARROWED sharply. IN PRACTICE: specifying audience, register, length, and structure gives the model a target; a persona gives it a vibe, and vibes drift. Expert-persona framing as a _capability_ lever has no supporting evidence for these generations. STILL VALID WHEN: the role is itself the product contract — a customer-facing assistant whose identity is part of the spec. - **Emphasis scaffolding (`CRITICAL`, `IMPORTANT`, all-caps, emoji markers)** — SOLVED: instruction dilution in long prompts on weaker models. STATUS: ABSORBED and counterproductive. IN PRACTICE: when a rule is ignored, the file is too long and the rule is lost. Delete other rules rather than shouting this one. STILL VALID WHEN: nothing. - **XML tag structuring** — SOLVED: ambiguous boundaries between instructions, context, and data. STATUS: LIVE for multi-part prompts, UNEVIDENCED as a performance trick. IN PRACTICE: vendor guidance still recommends tags when a prompt mixes instructions, examples, and variable input. It is not a universal accuracy lever. STILL VALID WHEN: the prompt genuinely has parts that a reader could confuse. ## Orchestration - **ReAct (reason + act loop)** — SOLVED: models that would not interleave reasoning with tool calls. STATUS: ABSORBED as runtime behavior. IN PRACTICE: this is what an agent loop _is_. Naming the pattern in a prompt adds nothing. STILL VALID WHEN: you are building the loop rather than prompting inside one. - **PAL / program-aided reasoning** — SOLVED: arithmetic and precise computation in token space. STATUS: RELOCATED. IN PRACTICE: the model writes and runs code natively; deterministic filtering and aggregation belong outside the context window entirely. STILL VALID WHEN: the concept is permanent — the prompt recipe is not. - **Prompt chaining** — SOLVED: error propagation in one long prompt. STATUS: NARROWED. IN PRACTICE: adaptive reasoning and native delegation handle most multistep work internally. STILL VALID WHEN: you need to inspect, log, evaluate, or branch on an intermediate output. The common surviving case is generate → review against criteria → refine, as separate calls. - **Meta-prompting (ask the model to write the prompt)** — SOLVED: authoring effort. STATUS: LIVE but unremarkable. IN PRACTICE: it works and needs no technique name. STILL VALID WHEN: always available; not a discipline. - **APE / automated prompt search** — SOLVED: manual prompt iteration without a metric. STATUS: NARROWED. IN PRACTICE: requires an eval set to mean anything. Without one it optimizes noise. STILL VALID WHEN: you have a labeled eval and a production pipeline worth tuning. - **DSPy** — SOLVED: hand-tuned prompts in multi-stage pipelines. STATUS: UNEVIDENCED. IN PRACTICE: no ablation located for these generations. Do not call it obsolete and do not call it best practice. STILL VALID WHEN: you have labeled eval data and a compound pipeline — the original premise still holds even though the numbers are stale. ## API-level - **Prefilling the assistant turn** — SOLVED: forcing output format, skipping preamble. STATUS: REMOVED on APIs that reject a trailing assistant message. IN PRACTICE: structured outputs for schemas, a system instruction for preamble, and a quoted continuation in the user turn to resume. STILL VALID WHEN: the API still accepts it. - **Output-format constraints in prose** — SOLVED: unparseable output. STATUS: RELOCATED. IN PRACTICE: a schema, enum, or typed tool is the contract; prose describing a format is a weaker version of the same thing. STILL VALID WHEN: the consumer is human and no schema applies. --- ## How to answer "should I use X?" 1. If the model already does it — deleting the instruction is the improvement. 2. If it belongs to another surface — say which one and stop writing prompt text. 3. If nobody has measured it — say that, rather than inheriting the old recommendation or inventing a new verdict. The third case is the one most often gotten wrong in both directions. "Nothing measured" is a real answer and it is not the same as "dead."