Promptception: Nested Prompt Methods

• Promptception is a meta-prompting paradigm that treats prompts as dynamic artifacts capable of recursive generation and systematic refinement.
• It integrates methods like recursive prompt generation, structured adaptive pipelines, and sensitive evaluation to optimize large model performance.
• Applications include automated prompt engineering, reverse prompt design, and controlled prompt lifecycle management in multimodal systems.

Searching arXiv for papers on promptception and closely related prompt-generation/management research. arxiv_search(query="promptception prompt generation agent prompt management prompt sensitivity multimodal", max_results=10) Promptception denotes a family of prompt-centric methods in which prompts are treated not only as instructions for a model, but also as artifacts that generate, refine, manage, inspect, or systematically vary other prompts. In one line of work, it refers to recursive or meta-level prompting, such as a “parent” prompt that instantiates an agent which then produces task-specific “child” prompts; in another, it names a framework for measuring how strongly model performance depends on prompt phrasing and structure. Across these usages, the unifying idea is that prompts become explicit objects of computation, evaluation, and lifecycle management rather than opaque strings embedded in an application (Shen et al., 2023, Cetintemel et al., 7 Aug 2025, Ismithdeen et al., 4 Sep 2025, Dovdon, 15 Jun 2026).

1. Core meanings and conceptual scope

The literature uses the term in several closely related but non-identical senses. In Promptor, promptception is presented as “the idea of prompts that generate or refine other prompts,” realized through a “Bootstrapped Promptception chain—Prompt2Prompt” in which Promptor itself is produced by a “parent” prompt and then generates “child” prompts for intelligent text entry (Shen et al., 2023). In SPEAR, the same general idea is formalized as nested or self-referential prompt structures in which a refinement operator can invoke an LLM on existing prompt text to produce new prompt text (Cetintemel et al., 7 Aug 2025). In PromptMN, promptception is realized through true nesting of prompts, self-referential prompting, and reverse prompt engineering, with PromptMN blocks embedded inside later prompts and interpreted via semantic resolution (Dovdon, 15 Jun 2026). By contrast, the paper titled “Promptception: How Sensitive Are Large Multimodal Models to Prompts?” uses the label for a systematic framework that varies prompt wording and structure to diagnose prompt sensitivity in large multimodal models (Ismithdeen et al., 4 Sep 2025).

Strand	Representative mechanism	Primary concern
Recursive prompt authoring	“parent” prompt produces prompt-generation agent, which emits “child” prompts	Autonomous prompt generation
Structured adaptive pipelines	Prompt algebra over $(P, C, M)$ with REF and CHECK	Runtime prompt refinement
Prompt DSL and nesting	%-prefixed directives inside ∞ … ∞ blocks	Inspectable, reusable prompt artifacts
Prompt sensitivity framework	61 prompt types across 15 categories and 6 supercategories	Fair and reproducible evaluation

This distribution of meanings suggests that promptception is best understood as a meta-prompting umbrella rather than a single algorithm. A plausible implication is that the field is converging on two complementary questions: how prompts can rewrite prompts, and how prompt changes alter downstream model behavior.

2. Recursive prompt generation and the Prompt2Prompt paradigm

Promptor provides a concrete realization of recursive prompt authoring for intelligent text entry. Its architecture comprises a Designer Interface implemented as a web chat UI, a Promptor Agent based on GPT-4 loaded with a fixed parent system prompt, a Test Virtual Keyboard that submits generated child prompts to GPT-3.5, and an Evaluation Module that uses GPT-4 via OpenAI Evals to score outputs on format correctness, similarity, and coherence (Shen et al., 2023).

The workflow is explicitly staged. A designer first submits an initial request such as building a sentence-completion prompt for restaurant booking. Promptor then asks clarifying questions about user goal, profile, context, and data constraints; generates an intermediate prompt template; receives designer ratings on Relevance, Clarity, and Specificity; runs test rounds through the keyboard; and iteratively refines the template based on both designer feedback and GPT-3.5 outputs. When the process converges, Promptor emits a final child system prompt described as a structured, few-shot plus chain-of-thought template with a policy section (Shen et al., 2023).

Collaborative evaluation is represented by a weighted-sum score

$$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$

with $w_r=w_c=w_s=\tfrac13$, and Promptor proceeds when $\mathrm{Score}>4$ (Shen et al., 2023). The final child prompt embeds four examples, each with step-by-step reasoning using “Let’s think step by step,” thereby combining few-shot prompting with explicit chain-of-thought scaffolding.

Promptor’s evaluation is reported along two axes. For the KWickChat task, a Wilcoxon Signed-Rank test showed that prompted GPT-3.5 outperforms fine-tuned GPT-2 with $p<0.001$, while prompted GPT-3.5 and fine-tuned GPT-3.5 showed no significant difference with $p>0.1$. Agreement between human and AI scores was high, with Spearman’s $\rho$ between $0.94$ and $0.96$, and Cohen’s $\kappa \approx 0.86$ to $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$0 (Shen et al., 2023). In a user study with 24 designers across three use cases, self-designed prompts incurred format-correctness error of at least 30%, whereas Promptor-designed prompts remained below 5%; similarity improved from $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$1 to $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$2, a 35% increase, and coherence from $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$3 to $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$4, a 22% increase. The reported System Usability Scale for Promptor was 83.4, labeled “excellent” (Shen et al., 2023).

These results position promptception, in this sense, as an alternative to both manual prompt engineering and fine-tuning-heavy workflows. The paper states that manual prompt engineering is ad hoc, unstable, and often fails format constraints, while fine-tuning requires data collection and may yield little or no improvement over in-context learning via prompts (Shen et al., 2023).

3. Algebraic prompt management in adaptive LLM pipelines

SPEAR generalizes promptception from an interactive design workflow into a language and runtime model in which prompts are first-class components of a pipeline. A SPEAR pipeline operates over a triple $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$5, where $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$6 is a structured, versioned key-value store of named prompt fragments, $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$7 is runtime context, and $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$8 is metadata containing diagnostic or control signals such as confidence and latency. The core operators are $$\mathrm{Score}(\pi)=w_r\cdot \mathrm{Relevance}(\pi)+w_c\cdot \mathrm{Clarity}(\pi)+w_s\cdot \mathrm{Specificity}(\pi),$$9, each modeled as a pure function from $w_r=w_c=w_s=\tfrac13$0 to $w_r=w_c=w_s=\tfrac13$1 (Cetintemel et al., 7 Aug 2025).

Within this formulation, prompt composition is chiefly expressed through REF and CHECK. REF[action, f] constructs or refines one or more entries in $w_r=w_c=w_s=\tfrac13$2, and CHECK[cond, f] conditionally invokes REF when a runtime predicate holds. Derived operators include EXPAND, SWITCH, and [MERGE](https://www.emergentmind.com/topics/merge). The explicit promptception case arises when the refinement function itself invokes an LLM on prompt text already stored in $w_r=w_c=w_s=\tfrac13$3. The paper gives the example

$w_r=w_c=w_s=\tfrac13$4

which is described as a prompt to the LLM about another prompt: an explicit, self-referential form of promptception (Cetintemel et al., 7 Aug 2025).

SPEAR distinguishes three refinement modes. In manual mode, the developer directly appends or inserts text. In assisted mode, the developer provides a high-level hint and the system calls an LLM to rewrite the prompt. In automatic mode, refinement is triggered by runtime metadata such as $w_r=w_c=w_s=\tfrac13$5, and an automated refiner updates the prompt accordingly (Cetintemel et al., 7 Aug 2025). This classification makes prompt refinement a controlled runtime operation rather than an ad hoc pre-processing step.

Prompt management is correspondingly structured. A prompt entry stores the current text, an optional view, a version, and a ref history describing refinement steps. Introspection APIs include P.history(key), P.diff(key, ver1, ver2), P.viewNames(), and P.getVersion(key) (Cetintemel et al., 7 Aug 2025). This makes prompt evolution inspectable, which is especially important once prompts begin rewriting other prompts.

SPEAR also frames promptception as an optimization problem. Operator fusion can merge adjacent GEN operations with shared dependencies into a single generation step. On Qwen2.5/Instruct, the paper reports speedup up to $w_r=w_c=w_s=\tfrac13$6 with 4–8% accuracy drop for Map→Filter fusion. Prefix caching exploits stable prompt prefixes $w_r=w_c=w_s=\tfrac13$7 and small deltas $w_r=w_c=w_s=\tfrac13$8; in a prompt-refinement setting, the paper reports Static Prompt at 3.10 s with $w_r=w_c=w_s=\tfrac13$9 and CacheHit $\mathrm{Score}>4$0, versus Auto Refinement at 2.12 s with $\mathrm{Score}>4$1 and CacheHit $\mathrm{Score}>4$2 (Cetintemel et al., 7 Aug 2025). These results suggest that once prompts are made structured and versioned, promptception becomes compatible with classical execution-plan optimization.

4. PromptMN and pseudo prompting as a nested prompt language

PromptMN approaches promptception through a domain-specific language that annotates natural language with compact, typed directives. A PromptMN program is delimited by the special token pair ∞ … ∞, and directives are %-prefixed constructs with text, block, or list arguments. The paper gives a minimal EBNF grammar covering DirectiveList, Directive, TextArg, BlockArg, ListArg, and a fixed keyword inventory (Dovdon, 15 Jun 2026).

The language defines forty-eight fixed keywords organized into seven clusters: Lexical Syntax; Context & Intent; Requirements Governance; Planning Orchestration; Control-Flow & Computation; Data Interface; and Meta & Lifecycle. Representative directives include %role, %goal, %domain, %reqfunc, %mustnot, %plan, numbered steps such as %1, %if, %repeat, %method, %in, %[out](https://www.emergentmind.com/topics/outer-automorphism-out), %diagram, %update, %newconcept, %format, and %meta (Dovdon, 15 Jun 2026). Each keyword is mapped internally to a semantic role tuple $\mathrm{Score}>4$3, and interpretation is driven by semantic role rather than textual position.

The semantic resolution algorithm is presented in five stages: tokenization around the ∞…∞ boundary; parsing into AST nodes; cluster assignment and type-checking; semantic reordering into a canonical sequence of clusters; prompt assembly; and execution with optional tracing. Within Planning Orchestration, numbered steps are sorted in ascending order regardless of source position. If %showplan is present, the system emits the assembled plan before taking action; if %trace is present, it parses the model response for reasoning logs and returns them under %out (Dovdon, 15 Jun 2026).

PromptMN’s promptception use case is reverse prompt engineering. A user may supply a natural-language specification and ask the model to restate it in PromptMN format, explicitly including roles, domain, goals, requirements, plans, inputs, and outputs. The resulting PromptMN block can then be inspected for missing %aware or %mustnot directives, revised with follow-up directives such as %newreq or %mustnot, and reused as the input to a later prompt. The paper states that the interpreter recognizes the nested ∞…∞ block, extracts it, runs semantic resolution, and feeds the assembled prompt to the LLM (Dovdon, 15 Jun 2026).

Feasibility is demonstrated across Claude Fable 5, Claude Opus 4.8, Gemini 3.1 Pro, and GPT 5.5 without fine-tuning. In the Basic Loop Test, all four models printed “Hello AI World!” exactly three times under a %repeat 3 times specification. In the Prime-Check Method task, they returned “true” for $\mathrm{Score}>4$4. In SDLC scenarios, models generated coherent code or documentation fragments from interleaved PromptMN descriptors. The paper also reports that the PromptMN-based reverse prompt yielded the richest, most actionable specification among three reverse prompting styles, and qualitatively states that no model drift or hallucination was observed when directives were clear, while hard constraints specified via %mustnot were strictly followed (Dovdon, 15 Jun 2026).

5. Prompt sensitivity and the evaluation meaning of Promptception

A distinct use of the term appears in the multimodal evaluation framework “Promptception: How Sensitive Are Large Multimodal Models to Prompts?” Here promptception is not a prompt-rewriting mechanism but a systematic methodology for measuring how strongly LMM performance depends on prompt phrasing and structure in multiple-choice question answering (Ismithdeen et al., 4 Sep 2025).

The framework comprises 61 prompt types organized into 15 categories and 6 supercategories. These probe choice formatting and presentation, linguistic and stylistic challenges, thought process and reasoning, context-aware and ethical guidance, task-specific instructions, and performance, feedback, and penalty framing. Examples include formatting variations such as “A. … B. …” versus “(A) …”, structured formatting via JSON, YAML, or Markdown, prompt position changes, poor linguistic formatting, prompt length, chain-of-thought prompts, ambiguity and probabilistic prompts, target-audience prompts, roleplay scenarios, anti-bias instructions, focus-driven prompts, answer handling, penalty-based prompts, incentive-based prompts, and competitive framing (Ismithdeen et al., 4 Sep 2025).

The framework defines several formal metrics. Accuracy swing is

$\mathrm{Score}>4$5

measuring worst-case spread across the 61 prompts. Robust central tendency is summarized by a 10% trimmed mean $\mathrm{Score}>4$6. Relative metrics are PRA and PRAD, defined against a baseline prompt accuracy $\mathrm{Score}>4$7 by

$\mathrm{Score}>4$8

The experimental setup covers 10 LMMs, including eight open-source models and two proprietary models, evaluated on MMStar, MMMU-Pro, and MVBench under zero-shot conditions (Ismithdeen et al., 4 Sep 2025).

The reported findings are substantial. Certain prompt types induce swings of up to approximately 15 percentage points, exemplified by Category 12 variations on GPT-4o in MMMU-Pro. Proprietary models exhibit larger standard deviation across prompts; GPT-4o and Gemini 1.5 Pro have the largest $\mathrm{Score}>4$9, whereas mid-sized open-source models from 8B to 38B are steadier, and 1B models are more variable due to limited context capacity. Category-level sensitivity differs by model class: proprietary models exceed $p<0.001$0 pp across all 15 categories, while open-source models exceed that threshold in exactly 6 categories: 2, 6, 7, 8, 9, and 12. Benchmark sensitivity also varies, with MVBench slightly more variable than MMStar and MMMU-Pro; within MMStar, ScienceQA is the most sensitive subset and Math the single most sensitive capability (Ismithdeen et al., 4 Sep 2025).

The paper’s interpretation is model-class-specific. Proprietary models are described as more sensitive because extensive instruction-tuning, RLHF, meta-prompting, and high-quality data cause them to “lock on” strongly to user instructions. Open-source models are steadier across prompt structures, but less responsive to subtle cues such as structured JSON or CoT reasoning. The resulting prompting principles diverge accordingly: open-source models are best served by short, direct, neutral prompts with simple formatting, whereas proprietary models tolerate or benefit from longer structure, explicit step-by-step reasoning, and even some penalty or reward framing (Ismithdeen et al., 4 Sep 2025). In this sense, promptception designates a robustness problem as much as a capability.

6. Limitations, misconceptions, and open research questions

The current literature consistently frames promptception as useful but constrained. Promptor notes that GPT-4 and GPT-3.5 calls introduce inference latency relative to local, rule-based auto-complete in tight typing loops, that automation is only partial because designers still manually copy prompts into the keyboard UI and configure parameters, and that domain transfer beyond intelligent text entry requires re-crafting the parent prompt’s facts and examples (Shen et al., 2023). SPEAR identifies termination and confluence, debuggability of nested prompt graphs, nonzero latency for each refinement step, model drift away from developer intent, and the absence of full denotational semantics for nested REF as open questions (Cetintemel et al., 7 Aug 2025). PromptMN reports token-overhead for full specifications and treats large-scale validation as future work (Dovdon, 15 Jun 2026). The multimodal Promptception framework shows that prompt sensitivity can make evaluation neither fair nor reproducible if only a single hand-tuned prompt is reported (Ismithdeen et al., 4 Sep 2025).

Several misconceptions are therefore ruled out by the available results. Promptception is not synonymous with arbitrary prompt tinkering: Promptor operationalizes it as a six-step Engage→Intermediate→Evaluate→Test→Refine→Finalize pipeline, SPEAR embeds it in a typed operator algebra, and PromptMN formalizes it with a directive language and semantic resolution (Shen et al., 2023, Cetintemel et al., 7 Aug 2025, Dovdon, 15 Jun 2026). Nor is it equivalent to fine-tuning: Promptor explicitly reports that prompted GPT-3.5 and fine-tuned GPT-3.5 show no significant difference on the KWickChat comparison, while the prompt-based approach avoids data collection and fine-tuning costs (Shen et al., 2023). Conversely, promptception does not guarantee universal improvement. The multimodal evaluation results show that some prompt variants degrade accuracy by 10–15% PRA, and that prompt effectiveness is contingent on model family, task format, and benchmark (Ismithdeen et al., 4 Sep 2025).

The field’s future directions are correspondingly concrete. Promptor proposes full end-to-end automation from conversational goal to deployed keyboard code, a voice interface, and a library of domain templates (Shen et al., 2023). SPEAR points toward richer runtime control over refinement modes, view reuse, and cost-based planning for deep nesting (Cetintemel et al., 7 Aug 2025). PromptMN suggests broader reuse across SDLC scenarios and more systematic interpreter support for concise and full specifications (Dovdon, 15 Jun 2026). Taken together, these directions suggest that promptception is moving from a narrow prompt-engineering tactic toward a broader systems view in which prompts have syntax, history, execution semantics, optimization opportunities, and evaluation protocols.