Ai Prompt Engineering

Prompt Engineering Foundations

Doctoral work examines how instructions shape the behaviour of language models. Instruction design substantially changes the quality of what models produce.

Instruction Following Research

Research examines how reliably models carry out the instructions they receive. Following capability differs greatly between models and between tasks.

Prompt Sensitivity Research

Doctoral study examines outputs changing with small changes to wording. Sensitivity undermines claims that any single measured result is stable.

Prompt Robustness Research

Research examines instructions producing consistent results despite variation. Robustness matters because deployed inputs are never perfectly controlled.

Wording Variation Research

Doctoral work examines how paraphrasing an instruction changes model behaviour. Paraphrase effects reveal that models attend to surface form heavily.

Prompt Format Research

Research examines how instruction layout influences the resulting output. Format effects are large and frequently exceed differences in content.

Delimiter And Structure Research

Doctoral study examines markers separating sections within an instruction. Clear separation helps models distinguish instruction from supplied content.

System Message Research

Research examines persistent instructions establishing model behaviour throughout. System level instruction carries differing weight than user requests.

Role Assignment Research

Doctoral work examines assigning a described role within an instruction. Role framing measurably changes both style and substantive content.

Persona Prompting Research

Research examines instructing models to adopt described characters or expertise. Persona effects on genuine capability remain contested in the literature.

Task Framing Research

Doctoral study examines how a task is presented affecting model performance. Identical problems framed differently produce measurably differing results.

Goal Specification Research

Research examines stating clearly what an instruction is intended to achieve. Vague goals are among the commonest causes of unsatisfactory output.

Constraint Specification Research

Doctoral work examines expressing the limits an output must respect. Constraint adherence weakens noticeably as the number of constraints increases.

Output Format Specification

Research examines instructing models to produce particular output shapes. Format instruction is necessary wherever outputs feed automated systems.

Structured Output Research

Doctoral study examines producing machine readable rather than prose output. Structured generation is essential for integrating models into software.

Schema Guided Generation

Research examines constraining output to conform to a defined structure. Schema enforcement prevents the parsing failures that break pipelines.

Zero Shot Prompting Research

Doctoral work examines task performance without any worked examples supplied. Zero shot capability has improved substantially with model scale.

Few Shot Prompting Research

Research examines supplying worked examples within the instruction itself. Examples convey requirements that description alone conveys poorly.

Example Selection Research

Doctoral study examines choosing which demonstrations to include in a prompt. Selection strongly affects performance and is frequently done arbitrarily.

Example Ordering Research

Research examines how the sequence of supplied examples affects results. Ordering effects are substantial and were unexpected when first reported.

Demonstration Quality Research

Doctoral work examines whether example correctness actually matters for performance. Some findings suggest format matters more than example accuracy.

In Context Learning Research

Research examines models adapting behaviour from information within the prompt. This capability emerged unexpectedly and remains poorly understood.

Chain Of Thought Research

Doctoral study examines prompting models to reason before producing answers. Explicit reasoning substantially improves performance on multistep problems.

Step By Step Reasoning Research

Research examines decomposing problems into sequential reasoning steps. Sequential working exposes errors that single answers entirely conceal.

Self Consistency Research

Doctoral work examines sampling several reasoning attempts and comparing them. Agreement across attempts improves accuracy at additional computational cost.

Reasoning Path Research

Research examines the routes models take toward their eventual answers. Stated reasoning does not always reflect the actual computation performed.

Working Space Prompting

Doctoral study examines giving models room to work before answering. Intermediate working improves accuracy on computational and logical tasks.

Decomposition Prompting Research

Research examines breaking complex requests into simpler component parts. Decomposition suits problems that overwhelm any single written instruction.

Subtask Prompting Research

Doctoral work examines handling components of a task through separate requests. Separation permits differing approaches for differing components.

Progressive Prompting Research

Research examines solving simpler subproblems before addressing harder ones. Progressive approaches let earlier solutions inform later reasoning.

Analogical Prompting Research

Doctoral study examines prompting models to recall similar solved problems. Self generated analogies can substitute for supplied worked examples.

Retrieval Augmented Prompting

Research examines supplying retrieved documents alongside the instruction. Retrieval grounds outputs in sources rather than model recollection.

Context Provision Research

Doctoral work examines what background information an instruction should include. Provided context substantially reduces fabricated content in outputs.

Grounding Research

Research examines tying generated content to supplied source material. Grounding is the principal defence against confidently stated fabrication.

Citation Prompting Research

Doctoral study examines instructing models to attribute claims to sources. Attribution permits readers to verify what has actually been asserted.

Tool Use Prompting Research

Research examines instructing models to invoke external tools and services. Tool access lets models perform actions their weights cannot support.

Function Calling Research

Doctoral work examines models producing structured calls to defined functions. Reliable calling underpins integration with existing software systems.

Agentic Prompting Research

Research examines instructions supporting multistep autonomous task completion. Autonomy accumulates small errors across long sequences of actions.

Planning Prompt Research

Doctoral study examines prompting models to plan before acting or answering. Explicit planning improves coherence across extended task sequences.

Reflection Prompting Research

Research examines prompting models to review their own previous output. Reflection helps for some tasks and can introduce new errors elsewhere.

Self Critique Research

Doctoral work examines models identifying weaknesses within their own responses. Self identified faults are unreliable without external reference.

Self Correction Research

Research examines models revising answers after reviewing their own reasoning. Correction without external feedback frequently fails to improve accuracy.

Verification Prompting Research

Doctoral study examines separate checking of a previously produced answer. Independent verification detects errors that self review consistently misses.

Ensemble Prompting Research

Research examines combining outputs from several differing instructions. Combination reduces the variance that single prompts inherently carry.

Debate Prompting Research

Doctoral work examines opposing positions argued before reaching a conclusion. Structured disagreement surfaces considerations single responses omit.

Multi Model Prompting

Research examines coordinating several differing models within one workflow. Differing models bring complementary strengths to a shared task.

Prompt Chaining Research

Doctoral study examines passing output from one instruction into the next. Chaining permits complexity that any single instruction cannot manage.

Pipeline Design Research

Research examines architecting multistage systems built from model calls. Pipeline design determines both reliability and total operating cost.

Conditional Prompting Research

Doctoral work examines instructions branching according to intermediate results. Branching permits differing handling for differing input types.

Iterative Refinement Research

Research examines improving outputs across successive rounds of revision. Refinement gains diminish and can reverse after only a few iterations.

Clarification Prompting Research

Doctoral study examines models asking questions before attempting a task. Asking prevents confident work upon a fundamentally misunderstood request.

Negative Instruction Research

Research examines instructions stating what should not appear in output. Prohibitions are followed less reliably than positive instructions.

Emphasis And Repetition Research

Doctoral work examines restating requirements to improve adherence to them. Repetition helps and can also skew output toward the repeated element.

Length Control Research

Research examines instructing models to produce output of a target length. Length adherence remains unreliable across most currently available models.

Style Control Research

Doctoral study examines directing tone, register and voice within outputs. Style instruction is effective and can conflict with content requirements.

Prompt Optimisation Research

Research examines systematically improving instructions rather than guessing. Optimisation converts prompting from craft into measurable engineering.

Automatic Prompt Generation

Doctoral work examines models producing instructions for other model tasks. Automated generation frequently outperforms handwritten human instructions.

Prompt Search Research

Research examines exploring instruction space to find effective formulations. Search requires an evaluation signal to guide it toward improvement.

Gradient Free Optimisation

Doctoral study examines improving instructions without any model internal access. These methods suit models available only through interfaces.

Evolutionary Prompt Research

Research examines iteratively mutating and selecting candidate instructions. Evolutionary search explores formulations that people would not attempt.

Reinforcement Based Optimisation

Doctoral work examines improving instructions using reward signals. Reward design determines what the optimisation actually converges toward.

Model Assisted Prompting

Research examines using models to help people write better instructions. Assistance lowers the expertise barrier for effective prompt construction.

Meta Prompting Research

Doctoral study examines instructions that generate or improve other instructions. Meta approaches automate the iteration that people perform manually.

Prompt Compression Research

Research examines shortening instructions while preserving their effectiveness. Compression reduces both computational cost and processing latency.

Token Efficiency Research

Doctoral work examines achieving results using fewer processed text units. Efficiency matters greatly at production scale and for long documents.

Context Length Research

Research examines how much material a model can usefully consider at once. Stated capacity frequently exceeds the length models handle well.

Context Trimming Research

Doctoral study examines removing material when supplied content exceeds capacity. Trimming choices determine what information the model actually sees.

Long Context Prompting

Research examines instructing models supplied with very extensive material. Performance degrades on long inputs in ways stated limits do not reveal.

Position Effect Research

Doctoral work examines where within an input material is placed mattering. Material in the middle of long inputs is attended to least reliably.

Attention Allocation Research

Research examines which parts of an input models actually rely upon. Allocation analysis explains why some instructions are quietly ignored.

Prompt Caching Research

Doctoral study examines reusing computation across instructions sharing content. Caching substantially reduces cost for repeated system instructions.

Latency Optimisation Research

Research examines reducing the delay between request and completed response. Latency determines whether interactive applications feel usable.

Operating Efficiency Research

Doctoral work examines the resources consumed by prompting at production scale. Efficiency choices determine whether deployment remains viable.

Soft Prompt Research

Research examines learned continuous vectors replacing written instructions. Soft prompts are effective and entirely uninterpretable to people.

Prefix Tuning Research

Doctoral study examines learned parameters prepended to model inputs. Prefix methods adapt behaviour without modifying model weights themselves.

Prompt Tuning Research

Research examines optimising a small learned instruction representation. Tuning bridges written prompting and full model adaptation methods.

Parameter Efficient Adaptation

Doctoral work examines adapting models by changing very few parameters. Efficient adaptation is frequently compared against careful prompting alone.

Instruction Tuning Interaction

Research examines how training on instructions shapes later prompt behaviour. Training history explains much of what works for a given model.

Model Family Difference Research

Doctoral study examines instructions performing differently across model families. Prompts tuned for one family transfer poorly to another family.

Model Version Sensitivity

Research examines instruction performance changing across successive model releases. Deployed prompts can silently degrade when models are replaced.

Prompt Portability Research

Doctoral work examines instructions working across differing model providers. Portability protects deployments from dependence on a single supplier.

Prompt Transfer Research

Research examines adapting effective instructions to new tasks or models. Transfer reduces the effort required for each additional application.

Multilingual Prompting Research

Doctoral study examines instructing models in languages beyond English. Performance differs substantially between languages for identical tasks.

Cross Lingual Transfer Research

Research examines instructions in one language supporting tasks in another. Cross lingual approaches extend capability to underserved languages.

Low Resource Language Prompting

Doctoral work examines prompting for languages with limited training material. Performance gaps between languages are large and persistently understudied.

Code Prompting Research

Research examines instructions for generating and modifying software code. Code tasks permit automatic verification that prose tasks entirely lack.

Mathematical Prompting Research

Doctoral study examines instructions for quantitative and symbolic reasoning. Mathematical tasks have verifiable answers supporting rigorous evaluation.

Multimodal Prompting Research

Research examines instructions combining text with other input types. Combined inputs raise questions text only prompting never encountered.

Image Prompting Research

Doctoral work examines instructing models that accept visual inputs. Visual grounding introduces distinctive failure modes requiring separate study.

Audio Prompting Research

Research examines instructions for models processing speech and other sound. Audio inputs carry information that transcription entirely discards.

Video Prompting Research

Doctoral study examines instructing models supplied with moving image material. Video inputs are large and demand aggressive selection of content.

Document Prompting Research

Research examines instructions operating over lengthy structured documents. Document structure carries meaning that flat text conversion loses.

Table And Data Prompting

Doctoral work examines instructions operating over tabular and numeric data. Models handle tables inconsistently and errors are easily overlooked.

Prompt Template Research

Research examines reusable instruction patterns with substitutable content. Templates permit consistency across many similar repeated requests.

Template Library Research

Doctoral study examines curated collections of validated instruction patterns. Shared libraries prevent repeated rediscovery of the same approaches.

Prompt Versioning Research

Research examines tracking successive revisions of deployed instructions. Version tracking is essential for diagnosing behaviour changes later.

Prompt Management Research

Doctoral work examines organising instructions across a whole organisation. Management practice determines whether prompts remain maintainable.

Prompt Repository Research

Research examines storage systems holding instructions and their history. Repositories treat prompts as artefacts deserving proper engineering.

Collaboration Workflow Research

Doctoral study examines teams working jointly on shared instruction sets. Collaboration requires review practices that text editing alone lacks.

Prompt Documentation Research

Research examines recording the intent and constraints behind instructions. Undocumented prompts become unmaintainable once authors have moved on.

Reproducibility Research

Doctoral work examines whether reported prompting results can be repeated. Model changes and sampling variation both obstruct any reproduction.

Determinism Research

Research examines obtaining identical outputs from identical inputs. Complete determinism is difficult even with sampling entirely disabled.

Sampling Parameter Research

Doctoral study examines settings governing randomness in generated output. Sampling settings interact with prompting in poorly documented ways.

Decoding Strategy Research

Research examines algorithms selecting each successive generated output element. Decoding choice affects quality as much as instruction wording does.

Stopping Criterion Research

Doctoral work examines determining when generation should actually finish. Poor stopping produces both incomplete and needlessly extended output.

Prompt Evaluation Research

Research examines assessing whether an instruction achieves its intended purpose. Evaluation rigour distinguishes engineering from anecdotal impression.

Benchmark Design Research

Doctoral study examines constructing tasks that meaningfully test capability. Benchmark design determines what the field believes has improved.

Task Suite Research

Research examines collections of tasks assessing performance broadly. Suites reveal capability differences that single tasks entirely conceal.

Automatic Evaluation Research

Doctoral work examines scoring outputs without any human involvement. Automatic scoring scales well and correlates imperfectly with judgement.

Model Based Evaluation

Research examines using models to assess the outputs of other models. Model judges carry their own biases toward particular styles of output.

Human Evaluation Research

Doctoral study examines people assessing model output quality directly. Human assessment remains the reference standard and is genuinely expensive.

Annotation Agreement Research

Research examines consistency between people judging the same model outputs. Low agreement indicates the evaluation question is poorly defined.

Evaluation Validity Research

Doctoral work examines whether measurements capture the capability intended. Many popular measures test something other than what they claim.

Data Contamination Research

Research examines evaluation material appearing within model training data. Contamination inflates scores without any genuine capability gain.

Benchmark Overfitting Research

Doctoral study examines instructions tuned to specific evaluation datasets. Tuned prompts perform very much worse on genuinely unseen material.

Statistical Significance Research

Research examines whether reported differences exceed random variation. Many published prompting comparisons lack any statistical treatment.

Variance Estimation Research

Doctoral work examines quantifying run to run variation in measured results. Single run results are unreliable given the variation actually present.

Ablation Study Research

Research examines removing prompt elements to identify what actually matters. Ablation reveals that many included elements contribute nothing.

Comparative Study Research

Doctoral study examines fairly comparing differing prompting approaches. Unequal tuning effort makes most published comparisons largely unreliable.

Reporting Standard Research

Research examines what must be disclosed about any prompting experiment. Incomplete reporting prevents independent appraisal of published claims.

Failure Mode Research

Doctoral work examines characteristic ways prompted systems go wrong. Failure taxonomy supports systematic rather than reactive improvement.

Fabrication Research

Research examines models producing confident but entirely invented content. Fabrication is the most consequential failure for factual applications.

Factuality Research

Doctoral study examines whether generated claims correspond to actual reality. Factual accuracy varies enormously across topics and across phrasings.

Faithfulness Research

Research examines outputs remaining true to the supplied source material. Unfaithful summaries introduce claims the source never actually made.

Output Consistency Research

Doctoral work examines models giving compatible answers across related questions. Inconsistency reveals that apparent understanding is shallow.

Contradiction Research

Research examines models asserting incompatible claims within one response. Internal contradiction is common within longer generated passages.

Conflicting Requirement Research

Doctoral study examines behaviour when a request demands incompatible things. Resolution behaviour is rarely specified and rarely tested at all.

Ambiguity Handling Research

Research examines responses to requests admitting several interpretations. Models typically guess rather than surfacing the ambiguity present.

Underspecification Research

Doctoral work examines requests omitting information needed to answer well. Underspecified requests are extremely common within real world usage.

Refusal Behaviour Research

Research examines models declining to carry out particular requests. Refusal behaviour should track genuine risk rather than surface phrasing.

Overrefusal Research

Doctoral study examines models declining requests that are entirely benign. Excessive refusal undermines usefulness without improving safety.

Sycophancy Research

Research examines models agreeing with users rather than maintaining accuracy. Agreement seeking makes models unreliable for genuine verification.

Position Bias Research

Doctoral work examines models favouring options by their placement order. Position effects distort comparison and evaluation tasks substantially.

Anchoring Effect Research

Research examines earlier content disproportionately shaping later output. Anchoring means initial framing constrains everything that follows.

Format Bias Research

Doctoral study examines preferences for particular output shapes regardless of merit. Format preference distorts model based evaluation considerably.

Verbosity Bias Research

Research examines longer responses being judged better irrespective of quality. Length preference affects both human and model based judging.

Social Bias In Prompting

Doctoral work examines instructions eliciting biased treatment of social groups. Prompt phrasing can substantially amplify or reduce measured bias.

Stereotype Amplification Research

Research examines outputs reinforcing stereotypes present in training material. Amplification occurs even where instructions request neutrality.

Fairness Evaluation Research

Doctoral study examines measuring whether outputs treat groups equitably. Measurement choice determines which disparities become visible at all.

Demographic Sensitivity Research

Research examines output changing when demographic details are varied. Sensitivity testing exposes differential treatment that averages conceal.

Calibration Research

Doctoral work examines whether expressed confidence matches actual accuracy. Models frequently express certainty that their accuracy cannot support.

Confidence Expression Research

Research examines how models communicate certainty within their responses. Expressed confidence strongly influences how readers treat the output.

Uncertainty Communication

Doctoral study examines conveying genuine uncertainty to those relying on output. Honest uncertainty is more useful than confident approximation.

Abstention Research

Research examines models declining to answer when they lack sufficient basis. Appropriate abstention prevents confident fabrication in factual tasks.

Prompt Injection Research

Doctoral work examines instructions embedded in content overriding intended behaviour. Understanding this failure is essential for building defences.

Indirect Injection Research

Research examines instructions reaching models through retrieved external content. Retrieval and tool use expand this exposure surface considerably.

Injection Defence Research

Doctoral study examines protecting systems from instructions hidden in content. No current defence provides complete protection against this problem.

Instruction Hierarchy Research

Research examines models prioritising instructions by their source and authority. Hierarchy helps models distinguish operator intent from supplied content.

Adversarial Robustness Research

Doctoral work examines systems resisting deliberately constructed problem inputs. Robustness evaluation must assume genuinely motivated adversaries.

Red Teaming Research

Research examines structured adversarial testing of deployed model systems. Red teaming finds weaknesses that ordinary evaluation entirely misses.

Safety Prompting Research

Doctoral study examines instructions supporting safe and appropriate behaviour. Instruction based safety complements rather than replaces training.

Guardrail Design Research

Research examines system components checking inputs and outputs for problems. Guardrails catch cases that instruction alone does not reliably handle.

Content Policy Research

Doctoral work examines translating policy requirements into system behaviour. Policy translation requires judgement that rules alone cannot capture.

Misuse Prevention Research

Research examines reducing opportunity for deployed systems to cause harm. Prevention must balance protection against ordinary legitimate use.

Security Evaluation Research

Doctoral study examines systematically assessing security of prompted systems. Security assessment differs substantially from capability evaluation.

Domain Specific Prompting

Research examines instructions tailored to specialist professional fields. Domain knowledge in prompts substantially improves specialist task quality.

Healthcare Prompting Research

Doctoral work examines prompting within clinical and health related tasks. Errors carry consequences demanding far stricter evaluation standards.

Legal Prompting Research

Research examines prompting applied to legal analysis and documentation. Fabricated citations in legal use have caused documented professional harm.

Scientific Prompting Research

Doctoral study examines prompting supporting scientific reasoning and writing. Scientific use demands traceability to genuine underlying sources.

Educational Prompting Research

Research examines instructions supporting teaching and student learning. Educational use must support understanding rather than merely supplying answers.

Software Engineering Prompting

Doctoral work examines prompting within software development workflows. Generated code requires review as carefully as any human contribution.

Data Analysis Prompting

Research examines instructions supporting statistical and analytical work. Analytical errors are difficult to detect within fluent explanation.

Creative Writing Prompting

Doctoral study examines prompting for narrative and creative production. Creative quality resists the measurement other tasks readily permit.

Translation Prompting Research

Research examines instructions supporting translation between languages. Prompting permits control over register that earlier systems lacked.

Summarisation Prompting Research

Doctoral work examines instructions producing condensed versions of material. Summaries frequently introduce claims the source never actually made.

Question Answering Prompting

Research examines instructions for answering questions from supplied material. Answer quality depends heavily on how the question is expressed.

Classification Prompting Research

Doctoral study examines using models to assign items to defined categories. Prompted classification competes with purpose trained smaller models.

Extraction Prompting Research

Research examines pulling structured information out of unstructured text. Extraction is among the most widely deployed practical applications.

Dialogue Prompting Research

Doctoral work examines instructions governing extended conversational exchange. Behaviour must remain stable across many conversational turns.

Customer Service Prompting

Research examines prompting within support and customer service applications. Service use demands consistency and appropriate escalation to people.

Accessibility Prompting Research

Doctoral study examines prompting supporting users with differing access needs. Accessible design determines who can benefit from these systems.

User Prompting Behaviour

Research examines how people actually write instructions in everyday practice. Real usage differs sharply from the prompts research papers study.

Novice User Research

Doctoral work examines inexperienced users interacting with these systems. Novices receive poorer results and rarely understand why they do.

Expert User Research

Research examines how domain experts adapt prompting to their own work. Expert strategies differ substantially from published general advice.

Mental Model Research

Doctoral study examines what users believe about how these systems function. Inaccurate beliefs lead directly to inappropriate reliance patterns.

Prompt Literacy Research

Research examines the skills people need to use these systems effectively. Literacy differences create a new dimension of digital inequality.

Interface Design Research

Doctoral work examines how prompting interfaces shape user behaviour. Interface design influences results as much as underlying model capability.

Prompt Suggestion Research

Research examines systems proposing improved instructions to their users. Suggestion narrows the gap between novice and expert user outcomes.

Feedback Loop Research

Doctoral study examines users learning from the responses they receive. Feedback quality determines whether users improve or repeat mistakes.

Human Model Collaboration

Research examines people and models working jointly upon shared tasks. Collaboration outcomes depend on how the work is divided between them.

Trust Calibration Research

Doctoral work examines users trusting outputs to an appropriate degree. Both excessive and insufficient trust produce poor overall outcomes.

Overreliance Research

Research examines users accepting output without adequate verification. Fluent presentation encourages acceptance regardless of actual accuracy.

Automation Bias Research

Doctoral study examines people deferring to automated output over their own judgement. This bias is well documented across many earlier technologies.

Cognitive Offloading Research

Research examines users delegating thinking to these automated systems. Offloading saves effort and may erode the capability being delegated.

Skill Effect Research

Doctoral work examines how sustained use affects underlying human capability. Effects on skill development remain genuinely uncertain currently.

Labour Impact Research

Research examines how prompted systems change the nature of human work. Impact differs greatly between tasks and between occupational groups.

Prompt Ownership Research

Doctoral study examines whether instructions constitute protectable creative work. Ownership questions remain legally unsettled across jurisdictions.

Prompt Sharing Research

Research examines communities exchanging effective instruction patterns. Sharing accelerates practice and spreads unverified claims equally.

Open Prompt Research

Doctoral work examines openly published instructions supporting reproducibility. Open prompts permit others to verify and to build upon results.

Prompt Marketplace Research

Research examines commercial exchange of instructions and their templates. Quality claims in these markets are largely unverified by anyone.

Disclosure And Provenance

Doctoral study examines indicating when content was produced with model assistance. Disclosure norms differ substantially between differing settings.

Regulatory Consideration Research

Research examines rules affecting how prompted systems may be deployed. Regulation increasingly addresses system behaviour rather than models alone.

System Auditing Research

Doctoral work examines independent examination of deployed prompted systems. Auditing requires access that system operators rarely willingly grant.

Organisational Governance Research

Research examines how organisations control the prompts that they deploy. Governance determines accountability when deployed systems cause harm.

Deployment Practice Research

Doctoral study examines moving prompted systems from prototype into production. Production imposes reliability demands that prototypes never face.

Production Monitoring Research

Research examines observing deployed system behaviour during actual use. Monitoring detects degradation that predeployment testing cannot anticipate.

Incident Response Research

Doctoral work examines responding when deployed systems behave unacceptably. Response capability determines how quickly harm is actually contained.

Theory Of Prompting Research

Research examines principled explanation of why particular instructions work. The field remains substantially empirical rather than theoretically grounded.

Mechanistic Explanation Research

Doctoral study examines internal model behaviour underlying prompt effects. Mechanistic understanding would replace trial and error with prediction.

Implementation And Adoption

Research examines why prompting advances reach practice or fail to do so. Published techniques are frequently ignored in actual deployed systems.