Artifacts Are All You Need
A Methodological Framework for Artifact-Oriented Prompt Micro-Systems
Christopher Luna Canessa · System Engineer · AI + Governance + Data + Strategy · San José, Costa Rica
Professional uses of large language models increasingly require more than isolated prompts. They require repeatable, validated, and renderable artifacts that preserve intent, evidence, quality, and lifecycle continuity. This paper proposes Artifact-Oriented Prompt Micro-Systems as a methodological framework for designing prompt-based systems whose unit of control is the artifact rather than the prompt. Drawing inspiration from the architectural clarity of the Transformer and the statement “Attention Is All You Need,” this paper argues that, in professional prompt workflows, artifacts are the operational unit through which attention, memory, validation, and continuation are governed. The proposed framework is grounded in the development of StrategyForge AI, a workplace strategy micro-system built through iterative testing, regression analysis, and lifecycle correction.
A prompt asks; an artifact governs.
I. Introduction
Large language models can produce high-quality text, but professional workflows require more than high-quality text. They require controlled progression from problem understanding to artifact delivery. The StrategyForge AI project began as an executive advisory prompt system and evolved into a lifecycle-based micro-system after repeated failures revealed that the core issue was not writing quality but state control: direct drafting, approval without export, render without continuation, and delivery instability.
II. Background and discovery process
The initial need was practical: transform workplace situations into HR-safe documentation, strategic follow-up notes, Workday performance inputs, and portfolio artifacts. The opportunity area expanded as the system began to reveal reusable behavior patterns. Each failure exposed a missing control layer. The process moved from prompt refinement to product specification, methodology design, artifact profiles, runtime control, response scaffolding, and regression memory.
- Prompt refinement
- Product specification
- Methodology design
- Artifact profiles
- Runtime control
- Response scaffolding
- Regression memory
III. Related work
The proposed framework connects to several streams of research. Transformers formalized attention as a mechanism for weighting value vectors based on query-key similarity [1]. Prompting taxonomies show that prompting already includes methods for structuring reasoning, roles, and multi-step behavior [2]. DPO and related preference optimization work illustrate how model behavior can be aligned during post-training, while this work focuses on external runtime alignment [3]. Memory-augmented systems such as MemGPT demonstrate the value of managing external memory and context tiers [4]. Context engineering further supports the claim that inference behavior depends heavily on information payload structure [6], [7].
IV. Mathematical reference: attention and QKV
The scaled dot-product attention function is commonly expressed as:
In the Transformer, Q, K, and V are learned projections of token representations. This paper does not claim that prompt micro-systems modify internal QKV computations. Instead, it argues that artifact-oriented prompts shape the external context that becomes available to the model during inference. The analogy is operational: user intent acts as an external query, project files and case evidence act as external keys, and artifact candidates act as values.
V. Memory: parametric, contextual, operational, and external
Memory in LLM systems should be understood as more than stored information. It includes the ability to retain, retrieve, use, inhibit, and update information. StrategyForge AI primarily uses external and procedural memory: project files, methodology, artifact profiles, templates, and regression libraries. These components do not change the model’s parameters, but they influence the context and procedural constraints that guide generation.
VI. Artifacts as the unit of control
The central claim is that professional prompt systems should be designed around artifacts. A prompt asks; an artifact governs. Artifacts provide structure, acceptance criteria, validation logic, rendering expectations, and continuation pathways. This shift changes the design problem from generating responses to governing lifecycle execution.
VII. Cognitive skills and behavior activation
The system does not create new cognitive skills inside the model. Rather, it discovers and stabilizes prompt patterns that reliably activate useful behaviors already present in the model after pretraining and alignment. Post-training methods such as RLHF and DPO shape preferences and response tendencies; StrategyForge AI operates externally by arranging context, lifecycle rules, and artifacts so that beneficial behavior becomes more likely at runtime.
VIII. StrategyForge AI lifecycle
The lifecycle contains nine stages: intake, analysis, preliminary notes, expansion, validation, render, deepen invitation, user choice, and new cycle or closure. The fourth stage is the governing expansion phase because it determines which future subcases are legitimate continuations. Render is not the terminal state; it is a transition into user-directed continuation.
IX. Failure analysis
The system development process revealed recurring failures: direct-draft collapse, approval without export, render without delivery, render without continuation, and opportunity narrowing. These failures were not failures of writing quality. They were failures of lifecycle control and external state salience.
X. Discussion
The findings suggest that prompt systems should be evaluated not only by output quality but also by lifecycle adherence. A polished artifact can still be a system failure if it skipped intake, omitted validation, failed to export, or did not continue into justified subcases. This is analogous to mode confusion: the user expects one lifecycle state while the system behaves as if another state is active.
XI. Limitations
The framework remains an applied methodological proposal. It is based on iterative prototyping and scenario-based testing in one domain. Further validation is needed across models, organizations, artifact families, and workflow settings. The QKV framing is an external analogy and should not be interpreted as a claim of mechanistic modification inside the model.
XII. Conclusion
Artifacts are all you need is not a claim that prompts are irrelevant. It is a claim that, in professional LLM workflows, prompts become reliable only when organized around artifacts, lifecycle control, validation, rendering, and regression memory. StrategyForge AI demonstrates a practical path for transforming prompting from ad hoc instruction writing into artifact-oriented micro-system engineering. Its contribution lies in making attention and memory operationally governable from outside the model by structuring what the model sees, what state it must follow, what artifact it must produce, and what failure patterns it must avoid.
References
- A. Vaswani et al., “Attention is all you need,” in Advances in Neural Information Processing Systems, 2017.
- S. Schulhoff et al., “The Prompt Report: A systematic survey of prompting techniques,” arXiv, 2024.
- R. Rafailov et al., “Direct Preference Optimization: Your language model is secretly a reward model,” arXiv, 2023.
- C. Packer et al., “MemGPT: Towards LLMs as operating systems,” arXiv, 2024.
- Y. Wu et al., “A survey on memory mechanisms in the era of LLMs,” arXiv, 2025.
- Anthropic, “Effective context engineering for AI agents,” Anthropic Engineering, 2025.
- L. Mei et al., “A survey of context engineering for large language models,” arXiv, 2025.