pith. sign in

arxiv: 2606.17164 · v1 · pith:AGU3R27Tnew · submitted 2026-06-15 · 💻 cs.CL · cs.AI· cs.HC· cs.PL· cs.SE

PromptMN: Pseudo Prompting Language

Enkhzol Dovdon This is my paper

Pith reviewed 2026-06-27 03:35 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.HCcs.PLcs.SE
keywords PromptMNpseudo-promptingdomain-specific languageprompt engineeringreverse prompt engineeringgenerative AIsoftware development lifecyclehuman-AI interaction
0
0 comments X

The pith

PromptMN annotates natural language with percent-prefixed directives so models resolve roles, goals, constraints, and complex logic by function type.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces PromptMN as a pseudo-prompting domain-specific language that adds compact, %-prefixed typed directives to natural language prompts for elements including roles, goals, requirements, priorities, constraints, plans, inputs, and outputs. Authors write in any order while models interpret directives semantically according to function rather than sequence or prose position. The language sits between informal prompting and pseudocode, enabling inspectable and reusable prompts for software development workflows. It also supports reverse prompt engineering, where models restate desired outcomes in PromptMN format to reveal inferred assumptions. Evaluation demonstrates that multiple frontier models correctly handle structures such as repetition, conditionals, methods, and a prime-checking task without fine-tuning or task-specific adaptation.

Core claim

PromptMN is a domain-specific language for annotating natural language with %-prefixed typed directives covering roles, goals, requirements, priorities, constraints, plans, inputs, and outputs. Semantic resolution by function type lets authors write in any order, and the same vocabulary applies across SDLC scenarios. Models resolve complex PromptMN instructions including repetition, conditionals, methods, and prime-checking without fine-tuning. Reverse prompt engineering produces reusable artifacts that surface inferred roles, goals, constraints, and missing assumptions.

What carries the argument

PromptMN, a pseudo-prompting domain-specific language that uses %-prefixed typed directives whose function type determines how the model resolves their meaning.

If this is right

  • Prompts become reusable artifacts that align people and AI tools across the software development lifecycle.
  • Reverse prompt engineering reduces repair cycles by exposing inferred roles, goals, constraints, and missing assumptions before execution.
  • The same PromptMN vocabulary applies to new codebases, maintenance, and redesign tasks without modification.
  • Complex structures such as repetition, conditionals, and methods become reliably executable by models.
  • Prompts gain inspectability and reviewability while remaining lightweight for analysts, managers, and developers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • PromptMN could address context ambiguities that cause agent failures even when model capability is not the limiting factor.
  • The approach might combine with existing prompt libraries to create hybrid interfaces that are both human-readable and machine-executable.
  • Larger-scale testing across additional task domains would clarify whether the same directive set generalizes beyond the presented SDLC examples.
  • Adoption could shift prompting from fragile prose to versionable, diffable artifacts in collaborative workflows.

Load-bearing premise

Semantic resolution of the percent-prefixed directives by function type works reliably across models and tasks without fine-tuning or task-specific adaptation.

What would settle it

A model given a PromptMN-formatted prime-checking task or conditional directive produces an incorrect result or misinterprets the directive's intended function.

Figures

Figures reproduced from arXiv: 2606.17164 by Enkhzol Dovdon.

Figure 1
Figure 1. Figure 1: Conceptual architecture of PromptMN. A pseudo [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-tick decision flow for the Snake game (left) and the result of the generated code running in a browser [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
read the original abstract

Prompting has become the primary interface between humans and generative AI, yet many natural language prompts remain fragile: roles, goals, constraints, and expected outputs are often buried in prose or left implicit. In agentic and software development workflows, a misread at the first handoff can propagate through every step, since a significant portion of agent failures stem from context ambiguities rather than model limitations. This paper introduces PromptMN, a pseudo-prompting domain-specific language that annotates natural language with compact, %-prefixed typed directives covering roles, goals, requirements, priorities, constraints, plans, inputs, and outputs. Semantic resolution lets authors write in any order while the model interprets directives by function. PromptMN sits between informal prompting and programming-style pseudocode: structured enough to be inspectable and reusable, yet lightweight enough for analysts, managers, developers, and stakeholders across the software development lifecycle (SDLC). PromptMN also pairs with reverse prompt engineering. Asking a model to restate a desired outcome as PromptMN lets users inspect the inferred roles, goals, constraints, and missing assumptions before acting, reducing repair cycles and yielding a reusable artifact for aligning people and AI tools. PromptMN's feasibility is evaluated across several frontier models, including Claude Fable 5, Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5. The models correctly resolved PromptMN instructions, including complex structures such as repetition, conditionals, methods, and a prime-checking task, without fine-tuning. The same vocabulary applies across new codebases, maintenance, and redesign in the SDLC scenarios presented. While large-scale validation remains future work, these early results suggest PromptMN is a practical step toward clearer, more reviewable human-to-AI interaction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces PromptMN, a lightweight pseudo-prompting DSL that annotates natural-language prompts with compact %-prefixed typed directives (roles, goals, requirements, priorities, constraints, plans, inputs, outputs). It claims semantic resolution allows models to interpret directives by function regardless of authoring order, and reports that four frontier models (Claude Fable 5, Claude Opus 4.8, Gemini 3.1 Pro, GPT-5.5) correctly handled complex PromptMN structures including repetition, conditionals, methods, and a prime-checking task without fine-tuning. The same vocabulary is said to apply across SDLC scenarios, with an additional suggestion for reverse prompt engineering to produce reusable PromptMN artifacts.

Significance. If substantiated, PromptMN would offer a practical middle ground between free-form prompting and full pseudocode, potentially reducing context ambiguities in agentic and software-development workflows. The absence of any concrete evaluation data, however, prevents assessment of whether the approach delivers measurable gains over ordinary instruction following.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'the models correctly resolved PromptMN instructions, including complex structures such as repetition, conditionals, methods, and a prime-checking task, without fine-tuning' is unsupported; no PromptMN source, model outputs, definition of 'correct,' error rates, or baseline comparisons are supplied, rendering the feasibility evaluation unverifiable.
  2. [Evaluation paragraph] Evaluation paragraph (referenced in abstract): the assertion that 'the same vocabulary applies across new codebases, maintenance, and redesign' and that resolution works 'reliably across models and tasks' lacks any test cases, consistency metrics, or cross-model analysis, which is load-bearing for the paper's empirical contribution.
minor comments (1)
  1. [Abstract] Abstract: model names 'Claude Fable 5' and 'GPT-5.5' are non-standard; clarify whether these refer to specific released versions or are illustrative.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the need for verifiable evaluation details. The manuscript presents PromptMN as an early-stage proposal with illustrative examples rather than a comprehensive empirical study, and we agree the current text overstates the strength of the feasibility claims without supporting artifacts.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'the models correctly resolved PromptMN instructions, including complex structures such as repetition, conditionals, methods, and a prime-checking task, without fine-tuning' is unsupported; no PromptMN source, model outputs, definition of 'correct,' error rates, or baseline comparisons are supplied, rendering the feasibility evaluation unverifiable.

    Authors: We agree the claim as worded is unsupported by explicit evidence in the manuscript. The abstract refers to observations from our testing of frontier models but omits the actual directives, outputs, and criteria used. In revision we will rephrase the abstract to describe these as preliminary illustrations, add a dedicated evaluation section containing sample PromptMN source, corresponding model responses, and a definition of successful resolution, and explicitly note the absence of quantitative metrics or baselines. revision: yes

  2. Referee: [Evaluation paragraph] Evaluation paragraph (referenced in abstract): the assertion that 'the same vocabulary applies across new codebases, maintenance, and redesign' and that resolution works 'reliably across models and tasks' lacks any test cases, consistency metrics, or cross-model analysis, which is load-bearing for the paper's empirical contribution.

    Authors: We accept that the evaluation paragraph is insufficiently substantiated. While the manuscript references SDLC scenarios, it does not supply the concrete test cases or cross-model observations needed to support the claims. We will revise this section to include specific PromptMN examples drawn from new-codebase, maintenance, and redesign contexts, along with qualitative notes on model behavior, and will remove or qualify language implying broad reliability pending larger-scale validation. revision: yes

Circularity Check

0 steps flagged

No derivation chain or self-referential steps; empirical language proposal only

full rationale

The paper proposes PromptMN, a %-prefixed DSL for prompting, and asserts that frontier models resolved its directives (including repetition, conditionals, and a prime-checking task) without fine-tuning. No equations, fitted parameters, uniqueness theorems, or derivation steps appear in the abstract or described content. The evaluation claim is presented as direct observation rather than derived from prior results or self-citations. Absence of any mathematical or definitional chain means no opportunity for circular reduction exists; the work is self-contained as a language specification plus qualitative feasibility report.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The proposal rests on the untested assumption that current LLMs can reliably parse and act on the new directive syntax across tasks, plus the domain assumption that prompt ambiguity is a primary failure mode in agentic systems.

axioms (1)
  • domain assumption Frontier LLMs can interpret %-prefixed typed directives by semantic function without fine-tuning.
    Stated in the feasibility evaluation section of the abstract.
invented entities (1)
  • PromptMN directive vocabulary and semantic resolver no independent evidence
    purpose: To provide compact, typed annotations for prompt elements that models resolve by function.
    New language construct introduced by the paper; no independent evidence supplied.

pith-pipeline@v0.9.1-grok · 5852 in / 1273 out tokens · 45343 ms · 2026-06-27T03:35:34.831243+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 2 linked inside Pith

  1. [1]

    Cambridge university press, 1995

    Frederic Charles Bartlett.Remembering: A study in experimental and social psychology. Cambridge university press, 1995

    1995

  2. [2]

    How much do language models memorize?arXiv preprint arXiv:2505.24832, 2025

    John X Morris, Chawin Sitawarin, Chuan Guo, Narine Kokhlikyan, G Edward Suh, Alexander M Rush, Kamalika Chaudhuri, and Saeed Mahloujifar. How much do language models memorize?arXiv preprint arXiv:2505.24832, 2025

    arXiv 2025

  3. [3]

    Prompt engineering for chatgpt: a quick guide to techniques, tips, and best practices.Journal of Engineering Technology, 43(1), 2026

    Sabit Ekin. Prompt engineering for chatgpt: a quick guide to techniques, tips, and best practices.Journal of Engineering Technology, 43(1), 2026

    2026

  4. [4]

    A comprehensive taxonomy of prompt engineering techniques for large language models.Frontiers of Computer Science, 20(3):2003601, 2026

    Yao-Yang Liu, Zhen Zheng, Feng Zhang, Jin-Cheng Feng, Yi-Yang Fu, Ji-Dong Zhai, Bing-Sheng He, Xiao Zhang, and Xiao-Yong Du. A comprehensive taxonomy of prompt engineering techniques for large language models.Frontiers of Computer Science, 20(3):2003601, 2026

    2026

  5. [5]

    Context engineering 2.0: The context of context engineering.arXiv preprint arXiv:2510.26493, 2025

    Qishuo Hua, Lyumanshan Ye, Dayuan Fu, Yang Xiao, Xiaojie Cai, Yunze Wu, Jifan Lin, Junfei Wang, and Pengfei Liu. Context engineering 2.0: The context of context engineering.arXiv preprint arXiv:2510.26493, 2025

    arXiv 2025

  6. [6]

    Context engineering: A principal-agent theory of ai-augmented knowledge work.Available at SSRN, 2026

    Nicos Savva. Context engineering: A principal-agent theory of ai-augmented knowledge work.Available at SSRN, 2026

    2026

  7. [7]

    Moscow rules: A quantitative exposé

    Eduardo Miranda. Moscow rules: A quantitative exposé. InInternational Conference on Agile Software Development, pages 19–34. Springer, 2022

    2022

  8. [8]

    A survey of techniques, key components, strategies, challenges, and student perspectives on prompt engineering for large language models (llms) in education

    Wan Chong Choi and Chi In Chang. A survey of techniques, key components, strategies, challenges, and student perspectives on prompt engineering for large language models (llms) in education. 2025

    2025

  9. [9]

    A guide to prompt design: foundations and applications for healthcare simulationists.Frontiers in Medicine, 11:1504532, 2025

    Sara Maaz, Janice C Palaganas, Gerry Palaganas, and Maria Bajwa. A guide to prompt design: foundations and applications for healthcare simulationists.Frontiers in Medicine, 11:1504532, 2025

    2025

  10. [10]

    A survey of automatic prompt engineering: An optimization perspective.arXiv preprint arXiv:2502.11560, 2025

    Wenwu Li, Xiangfeng Wang, Wenhao Li, and Bo Jin. A survey of automatic prompt engineering: An optimization perspective.arXiv preprint arXiv:2502.11560, 2025

    arXiv 2025

  11. [11]

    Till J Adam, Salma AS Abosabie, Max Dittmer, Elise Wolf, Sara A Abosabie, Clara Behnke, Felix Baier, Annabelle Weickmann, Ludwig Köser, Christoph U Correll, et al. Prompt engineering of large language models for paper screening in medical meta-analyses and systematic reviews: A prospective comparative study.Research Synthesis Methods, pages 1–18, 2026

    2026

  12. [12]

    Prompt engineering competence, knowledge management, and technology fit as drivers of educational sustainability through generative ai.Scientific Reports, 2026

    Omer Gibreel, Kasım Karata¸ s, and Ibrahim Arpaci. Prompt engineering competence, knowledge management, and technology fit as drivers of educational sustainability through generative ai.Scientific Reports, 2026

    2026

  13. [13]

    Memento: Teaching llms to manage their own context

    Vasilis Kontonis, Yuchen Zeng, Shivam Garg, Lingjiao Chen, Hao Tang, Ziyan Wang, Ahmed Awadallah, Eric Horvitz, John Langford, and Dimitris Papailiopoulos. Memento: Teaching llms to manage their own context. arXiv preprint arXiv:2604.09852, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Agentic ai on the rise: Keys to unlocking value

    Amazon Web Services. Agentic ai on the rise: Keys to unlocking value. AWS Marketplace eBook, 2025. Amazon Web Services, Inc

    2025

  15. [15]

    A practical guide for designing, developing, and deploying production-grade agentic ai workflows.arXiv preprint arXiv:2512.08769, 2025

    Eranga Bandara, Ross Gore, Peter Foytik, Sachin Shetty, Ravi Mukkamala, Abdul Rahman, Xueping Liang, Safdar H Bouk, Amin Hass, Sachini Rajapakse, et al. A practical guide for designing, developing, and deploying production-grade agentic ai workflows.arXiv preprint arXiv:2512.08769, 2025

    arXiv 2025

  16. [16]

    Agentic ai patterns and workflows on aws

    Amazon Web Services. Agentic ai patterns and workflows on aws. Aws prescriptive guidance, Amazon Web Services, Inc., 2026

    2026

  17. [17]

    Llm-based agentic systems for software engineering: Challenges and opportunities.arXiv preprint arXiv:2601.09822, 2026

    Yongjian Tang and Thomas Runkler. Llm-based agentic systems for software engineering: Challenges and opportunities.arXiv preprint arXiv:2601.09822, 2026. 11 PromptMN: A Pseudo-Prompting DSL

    arXiv 2026

  18. [18]

    React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Pith/arXiv arXiv 2022

  19. [19]

    Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks

    Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, et al. Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. InProceedings of the 2022 conference on empirical methods in natural language processing,...

    2022

  20. [20]

    Prompting with pseudo-code instructions

    Mayank Mishra, Prince Kumar, Riyaz Bhat, Rudra Murthy, Danish Contractor, and Srikanth Tamilselvam. Prompting with pseudo-code instructions. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15178–15197, 2023

    2023

  21. [21]

    Training with pseudo-code for instruction following.arXiv preprint arXiv:2505.18011, 2025

    Prince Kumar, Rudra Murthy, Riyaz Bhat, and Danish Contractor. Training with pseudo-code for instruction following.arXiv preprint arXiv:2505.18011, 2025

    arXiv 2025

  22. [22]

    Alice"; %out

    Sebastian Femepid, Lachlan Hatherleigh, and William Kensington. Gradual improvement of contextual under- standing in large language models via reverse prompt engineering.Authorea Preprints, 2024. 12 PromptMN: A Pseudo-Prompting DSL A Appendix. The specification of the PromptMN PromptMN is a pseudo-prompting language (domain-specific language) designed to ...

    2024

  23. [23]

    Walker— patrols platforms, turns at edges/walls; damages on contact; 1 HP; can be defeated by jump-stomp, slide, or snowball

  24. [24]

    Flyer— hovers in sine-wave or circular patterns; periodically swoops toward the player; 1 HP; immune to slide (must be shot or stomped)

  25. [25]

    23 PromptMN: A Pseudo-Prompting DSL

    Shooter— stationary or slow-moving turret-like enemy (e.g., a walrus with an ice cannon); fires projectiles at intervals or aimed at the player; projectiles must be visible and dodgeable; 2 HP. 23 PromptMN: A Pseudo-Prompting DSL

  26. [26]

    Press Start

    Charger— idles until the player enters its line of sight, telegraphs (shake/flash ∼0.5 s), then charges fast in a straight line until hitting a wall (briefly stunned, vulnerable); 2 HP. Include simple spawning/placement via level data, and scale enemy density with difficulty. Level Design & Progression • At least5 handcrafted levelsdefined as tile maps in...

  27. [27]

    Opens and runs from a double-clicked single file with no installation, no internet, no console errors

  28. [28]

    All four enemy types behave as specified and are each defeatable by at least one method

  29. [29]

    Player movement feels tight: coyote time, jump buffering, and variable jump verified

  30. [30]

    All 5+ levels are completable; checkpoints and lives work; game can be finished start to end

  31. [31]

    Audio plays (after first user interaction, per browser autoplay rules), and mute works

  32. [32]

    High score persists across page reloads

  33. [33]

    Shall” denotes a mandatory requirement; “should

    Stable 60 FPS on a mid-range laptop with 50+ entities on screen. Deliver the complete code, followed by a short README textbf covering controls, how to run it, and a brief architecture overview. 25 PromptMN: A Pseudo-Prompting DSL E Appendix. Reverse Prompt: Software Requirements Specification Version:1.0Date:June 12, 2026Status:Draft for AI-Assisted Impl...

    2020

  34. [34]

    Run a fixed-timestep game loop with decoupled rendering

  35. [35]

    26 PromptMN: A Pseudo-Prompting DSL

    Render procedurally drawn retro graphics on a virtual low-resolution canvas. 26 PromptMN: A Pseudo-Prompting DSL

  36. [36]

    Accept keyboard input (primary), with optional gamepad and touch input

  37. [37]

    Simulate 2D platformer physics: gravity, jumping, ground/wall collision

  38. [38]

    Control a penguin player character with movement, jump, and two attacks

  39. [39]

    Spawn and simulate four enemy archetypes with distinct AI behaviors

  40. [40]

    Manage combat: hit detection, damage, knockback, invincibility frames, enemy death

  41. [41]

    Present at least three handcrafted levels plus a boss-free final gauntlet, with a defined level format permitting easy addition of new levels

  42. [42]

    Track score, lives, health, and collectibles; display them on a HUD

  43. [43]

    Manage game states: title, playing, paused, level complete, game over, victory

  44. [44]

    Generate all sound effects and music procedurally via synthesized audio

  45. [45]

    spiral of death

    Persist the local high score where the platform allows (e.g.,localStorage), degrading gracefully if unavail- able. User Characteristics Casual players aged roughly 8+, familiar with classic platformer conventions (arrow keys / W ASD, jump on a button). No tutorial text beyond a single controls screen is required. Constraints C-1: Zero external dependencie...

    2018

  46. [46]

    Opening the single file in each major browser shows the title screen; Enter starts Level 1

  47. [47]

    The penguin runs, jumps with variable height, coyote time, and jump buffering; slides under a 1-tile gap; throws at most 3 concurrent snowballs

  48. [48]

    Each enemy demonstrably exhibits its specified behavior, telegraphs, vulnerabilities, and immunities (notably: stomping a charging Bull Seal hurts the player; snowballs cancel icicles)

  49. [49]

    Spikes, falling off the level, and enemy contact damage all reduce HP/lives correctly, with knockback and i-frames; checkpoints respawn correctly

  50. [50]

    All three levels are completable; the victory screen shows after Level 3; high score persists across reloads where storage permits

  51. [51]

    Pausing freezes simulation and audio; muting silences everything instantly

  52. [52]

    Disconnecting from the network and reopening the file changes nothing (full offline operation)

  53. [53]

    Frame-rate test: gameplay speed is identical on a 60 Hz and a 144 Hz monitor

  54. [54]

    E.6 Suggested Implementation Order (for the AI Coding Session)

    The file contains no<script src>,fetch,import,@font-faceURL, or any other external reference. E.6 Suggested Implementation Order (for the AI Coding Session)

  55. [55]

    Canvas setup, integer scaling, fixed-timestep loop

  56. [56]

    Input layer (just-pressed/held abstraction)

  57. [57]

    Tile map parsing + rendering + AABB collision

  58. [58]

    Player movement, jump feel (buffer/coyote/variable height), camera

  59. [59]

    Snowball + slide attacks, projectile pooling

  60. [60]

    Enemy framework, then Walker→Flyer→Shooter→Charger

  61. [61]

    Damage system, i-frames, lives, checkpoints, pickups

  62. [62]

    31 PromptMN: A Pseudo-Prompting DSL

    Levels 1–3 authored in ASCII; goal/checkpoint logic. 31 PromptMN: A Pseudo-Prompting DSL

  63. [63]

    State machine + HUD + menus

  64. [64]

    Procedural audio (SFX first, then music sequencer)

  65. [65]

    Particles, screen shake, polish, high-score persistence

  66. [66]

    Run the acceptance checklist (Subsection A.5). 32