From Prompt Engineering to Context Engineering: Why Precision Oncology Needs Context

Introduction

Precision oncology strives to tailor treatment to the unique characteristics of each patient’s cancer—the right therapy for the right patient at the right time. Paradoxically, clinical practice is also guided by standardized protocols such as the National Comprehensive Cancer Network (NCCN) guidelines. These evidence-based guidelines provide a crucial baseline, but they are not monolithic rules; in fact, multiple valid treatment strategies exist for many oncologic scenarios. ¹ Oncologists frequently face choices among several guideline-concordant options, and they make decisions by weighing clinical evidence alongside real-world factors. The “art” of oncology lies in this personalized decision making. Oncology practice is heavily informed by protocols yet still “tailored to the individual patient,” and chemotherapy prescribing habits can vary markedly between different care settings. ²

This variability reflects the many contextual factors that guidelines alone cannot capture. Patient comorbidities and preferences, drug availability, physician experience, institutional treatment philosophies, and even financial or logistical considerations can all influence the chosen therapy. In other words, oncology isn’t a one-size-fits-all domain—10 oncologists might legitimately choose several different (but evidence-supported) treatment plans for the same case, each informed by subtle nuances. This inherent personalization is a cornerstone of high-quality cancer care. However, it poses a challenge for artificial intelligence (AI)-driven decision support: how can an AI tool provide useful recommendations without oversimplifying such a nuanced landscape?

The Risk of One-Size-Fits-All AI

Early forays into AI for oncology decision making revealed the pitfalls of ignoring context. A prominent example was IBM’s Watson for Oncology, an AI system trained primarily on expert inputs from a single cancer center. Watson often produced recommendations that, while based on guidelines, were perceived as overly generic, biased toward its training institution’s preferences, and not well adapted to other care settings. ³ For instance, its U.S.-centric guideline approach proved problematic in regions with different available treatments or medical practices. Doctors observing these outputs found that the AI frequently suggested options they already knew from guidelines, adding little new insight or nuance. ⁴ In effect, the system functioned like a high-powered literature and guidelines summarizer—useful as a reference, but falling short as a truly adaptive clinical partner.

If an oncology AI simply “spits out” the same answer for every patient based on average standards, it risks replacing clinical personalization with homogenization. What’s worse, an overly standardized AI might give the illusion of optimal care while in reality overlooking important patient-specific considerations that a human would catch. There is also a latent danger that busy clinicians might defer to an AI recommendation that appears authoritative, even when it doesn’t fit the patient’s context. ⁵ Current evidence strongly cautions that today’s AI models (especially general large language models [LLMs]) are far from ready to replace clinicians’ final judgment and at best require expert oversight as decision support tools. A naive implementation of AI could indeed make oncology more efficient in delivering information, yet less effective in individualizing care—the opposite of precision medicine’s promise.

Context in Oncology: Why One Guideline Doesn’t Fit All

To harness AI for true precision oncology, we must appreciate why experts often deviate from or customize guidelines. Oncology decisions are made in context, and several contextual dimensions can lead to equally valid variations in care. For instance, academic cancer centers often lean toward cutting-edge therapies and early-phase clinical trial data, whereas community clinics may prioritize treatments that minimize patient travel, reduce toxicity, and fit local resource constraints. Individual oncologists bring their own clinical experience and an institution’s philosophy into the equation, leading to variations that formal guidelines can’t fully capture.

Even within the bounds of evidence-based care, these factors mean that what is “optimal” can differ from one setting to another. An academic center with clinical trials on site might recommend an experimental targeted therapy supported by promising Phase II data, while a community practice without trial access might favor a more established regimen that aligns with a patient’s practical needs. Neither choice is wrong—both can be consistent with national guidelines, which often list multiple options—but each is optimized for a different context. Financial considerations can play a role as well; for example, studies have found differences in chemotherapy prescribing between practice settings, such as higher usage of costly biologic drugs in some settings under certain reimbursement models. ² Likewise, regional availability of drugs or cultural preferences might sway decisions. All these nuances represent “friction” in oncology care—the real-world constraints and preferences that guidelines abstract away. Seasoned clinicians inherently account for these factors (“contextualize” the guidelines) when treating patients.

From Prompt Engineering to Context Engineering

Modern AI systems, especially LLMs, have shown remarkable ability to parse vast medical knowledge. With the right query, a general LLM like ChatGPT can summarize NCCN guidelines for Stage IV lung cancer in seconds. However, the future of oncology AI isn’t about asking AI better questions—it’s about equipping AI with better context. Prompt engineering (crafting an ideal query ⁶ ) alone cannot yield a truly personalized answer if the AI lacks information about the setting and patient. This is where context engineering comes in.

In practical terms, context engineering means designing AI tools that incorporate the who, where, and why behind clinical decisions—not just the what. Rather than treating guidelines as one-size-fits-all, a context-aware system would adapt its recommendations based on additional inputs and training: the specific patient’s attributes, the treating institution’s patterns, and the cumulative experience of clinicians. Imagine an LLM-based assistant that doesn’t just generically list all guideline-recommended options but actually highlights the option most aligned with your clinic’s past practice for a patient with similar characteristics—and explains the reasoning. By embedding local protocol preferences, resource considerations, and patient-specific factors into its decision-making process, such an AI would provide adaptive decision support instead of prescriptive instructions.

In order to achieve this, data integration is key—AI models must be trained and updated with diverse real-world datasets that reflect different practice settings, including academic centers, community hospitals, and global oncology contexts. This will help the AI appreciate a variety of “right answers” under different circumstances and avoid overreliance on any single institution’s approach. Ensuring diversity in the training data (both in terms of patient demographics and practice environments) also addresses fairness and generalizability, so that recommendations hold up across underserved populations and varying resource settings. Second, AI tools should allow clinician feedback loops: if a recommendation was off-base due to missing context, the system should learn from that input. This could be as simple as clinicians rating or annotating AI suggestions, or as complex as iterative model refinement using local outcomes (a learning health system approach). Over time, this feedback-driven tuning embeds the collective wisdom of the oncology team into the AI’s knowledge base.

Contextualization is not meant to override medical evidence but to select the best evidence-supported option for a given scenario. An AI can be both adaptive and evidence-aligned—for example, it might know that both Drug A and Drug B are NCCN-approved for a condition, but learn that patients at your clinic with certain comorbidities tend to tolerate Drug B better, and thus recommend B (with an explanation). Such a system would preserve the integrity of national standards while fine-tuning their application. It’s a delicate balance: too little context, and the AI is an impersonal textbook; too much unchecked context, and the AI could encode local biases or outdated habits. Achieving the right balance will require careful validation and possibly regulatory oversight to ensure that adding context improves care quality and doesn’t drift into unwarranted variation.

Conclusions

The next era of AI in precision oncology will be defined by its ability to embrace the very complexity that makes oncology challenging. Rather than forcing conformity to generic recommendations, successful AI systems will acknowledge that the optimal cancer treatment is not only tumor-specific but context-specific. By moving beyond prompt engineering to context engineering, we can develop AI that respects institutional practices, adapts to patient circumstances, and ultimately delivers truly personalized recommendations. Such AI will act as a “true partner”—one that enhances oncologists’ capabilities and insight while preserving the nuanced decision making that each patient deserves. Ultimately, AI in oncology should remain augmentative—a tool to support, not replace, clinician decision making. The goal of context engineering is not to create an autonomous oncologist in a box but to off-load cognitive burden (like sifting through literature and records) and surface relevant insights that might otherwise be overlooked. With a context-aware assistant handling data synthesis, clinicians are freed to focus on the human elements of care: communication, empathy, and judgment in the face of uncertainty. In this partnership model, the AI continuously learns from the experts even as it informs them, creating a virtuous cycle of improvement.