Research: Intent Specification — How Practitioners Translate Fuzzy Direction into Productive GenAI Collaboration

Date: 2026-03-10 Search queries used:

• intent specification GenAI collaboration practitioners translate direction
• prompt engineering intent alignment design practice LLM collaboration
• translating fuzzy requirements AI specification UX design strategy
• intent specification human AI collaboration design thinking sensemaking
• vague direction GenAI prompt clarification techniques practitioners UX product design 2024 2025
• intent alignment specification LLM prompting system design 2024 2025 research HCI
• “intent specification” OR “intent clarification” GenAI design strategy
• practitioner strategies making intent legible AI prompting context specification design product 2025
• context engineering intent specification human AI collaboration practitioners 2025
• UX professionals perceptions AI-assisted design vibe coding intent expression natural language 2025

Executive Summary

Intent specification is the practice of translating fuzzy human direction—strategic goals, creative visions, design constraints—into forms legible to GenAI systems. Research across HCI, design practice, and software engineering reveals a fundamental tension: AI systems excel at executing precisely stated intent but struggle with ambiguity, yet human practitioners often resist upfront specification because it feels unnatural and negates the apparent efficiency gains of AI. This creates a skill gap that is now framing a new field practitioners are variously calling “intent engineering,” “context engineering,” or “specification-driven development.” The core debate is whether intent should be articulated before AI interaction (top-down specification) or discovered through it (emergent iteration). Both have costs and benefits that vary by task complexity, practitioner expertise, and phase of work.

Key Sources

Intent-Based User Interfaces (Ding & Chan, 2024)

• URL: https://arxiv.org/html/2404.18196v2
• Type: Academic paper (Google PhD Fellowship research proposal, ACM 2024)
• Key points:
  • Proposes a spectrum of intent expression: one-off expression → intent iteration → intent exploration/sensemaking
  • Three task categories: fixed-scope content curation, atomic creative tasks, complex/interdependent tasks
  • Experienced data analysts use detailed command-like prompts (“Correlation plot of data - (+) correlation coefficient in red”) → high quality but limits exploration
  • Novice analysts use high-level intents (“What is the relationship between density and birth rate?”) → lower quality but more exploratory
  • Simple chatbox interface sufficient for fixed-scope tasks; complex tasks need IUI with curation, iteration, and sensemaking affordances
• Tenet alignment: Strong alignment with Human Intent First; reveals how intent expression differs by expertise level
• Quote: “The spectrum of human intent expression ranges from one-off intent expression for fixed-scope content curation tasks…to intent exploration, iteration and sensemaking for complex and interdependent tasks.”

Vibe Coding for UX Design (Feng et al., 2025)

• URL: https://arxiv.org/html/2509.10652v1
• Type: Academic paper, ACM CHI submission (20 practitioner interviews)
• Key points:
  • Vibe coding = practitioners express intent in natural language, AI translates into functional prototypes and code
  • Four-stage workflow: ideation → AI generation → debugging → review
  • Core tension: “intending the right design” (efficiency-driven, delegate intent execution) vs. “designing the right intention” (reflection-driven, use AI to explore what you actually want)
  • AI described as “amplifier of human intent” that redistributes epistemic labor
  • Ownership is now tied to intention-setting, not code execution—new distinction between authorship of ideas and execution labor
  • Key failure mode: AI “drifts from original design intent”—“I have to keep pulling it back on track” (P15)
  • Introduces concept of “intention-to-UX ecosystem” where AI mediates between human goals and executable outputs
  • Risks: deskilling, over-reliance, homogenization of outputs
• Tenet alignment: Strong alignment with Symbiotic Intelligence (tension between efficiency and reflection); raises questions about Human Intent First (intent drift); touches Pluralism risk
• Quote: “We find tensions between efficiency-driven prototyping (‘intending the right design’) and reflection (‘designing the right intention’), introducing new asymmetries in trust, responsibility, and social stigma within teams.”

Prompt to Design Interfaces: Why Vague Prompts Fail (Wang, NNG, Dec 2025)

• URL: https://www.nngroup.com/articles/vague-prototyping/
• Type: Practitioner research article (Nielsen Norman Group)
• Key points:
  • Vague prompts produce “Frankenstein layouts”—randomly pieced together elements lacking hierarchy and flow
  • AI has surface-level understanding of individual components but fails at integrating them purposefully
  • “Verbosity without precision hurts clarity”—concise, well-chosen keywords outperform long vague descriptions
  • Five strategies for intent specification in design: (1) precise visual keywords referencing established design styles, (2) lightweight visual references (moodboards, screenshots), (3) AI-assisted visual analysis to generate structured prompt descriptions, (4) mock data / JSON to ground layout decisions, (5) code snippets as the most direct form of context
  • Key: providing realistic data guides AI to generate better layouts (content-focused design principle)
  • Using AI to critique and iterate prompts is itself a productive strategy
• Tenet alignment: Aligns with Context as Infrastructure; visual references and mock data are forms of context that function as infrastructure
• Quote: “A skilled human designer can take a broad design statement and identify and prioritize information, but AI struggles with ambiguity and is unable to deliver thoughtful results within a broad context.”

The Intent Stack: A New Design Space for Human-AI Collaboration (Lockie, 2026)

• URL: https://www.linkedin.com/pulse/intent-stack-new-design-space-human-ai-collaboration-david-lockie-ljgue
• Type: Practitioner framework article (LinkedIn)
• Key points:
  • Intent Stack is a hierarchical framework for organizing human intention for AI consumption
  • Five layers: Lifetime Intent → 5-Year Intent → Annual Intent → Operational Intents (3-6 active) → Project-Specific Intents
  • Lower layers inherit context from higher layers automatically
  • Paired with a Personal Context Document (PCD)—“where PCD says who I am, Intent Stack says what I’m trying to do and why”
  • Tools like claude.md handle preferences/how-to; Intent Stack handles purpose/why—complementary, not competing
  • Three practical uses: (1) AI context layer (agents consume the stack), (2) self-authoring tool (updated through reflection, therapy, assessments), (3) decision filter for new commitments
  • Living document: new sub-intents added as insight surfaces (“Before abandoning a project, ask: Am I bored or is this actually not serving my goals?”)
• Tenet alignment: Directly aligned with Human Intent First (making intent legible) and Context as Infrastructure (intent as persistent infrastructure)
• Quote: “The tools handle preferences. The Intent Stack handles purpose. They’re complementary — the stack provides the conceptual frame, tools like claude.md provide the mechanism for acting on it.”

Prompting Split into Four Skills (Nate’s Newsletter, Feb 2026)

• URL: https://natesnewsletter.substack.com/p/prompting-just-split-into-4-different
• Type: Practitioner newsletter (paywalled—only intro accessible)
• Key points:
  • Distinguishes four disciplines: context engineering, intent engineering, specification engineering, constraint architecture
  • “Intent engineering sits above context engineering the way strategy sits above tactics”
  • “The Klarna trap”: excellent context + missing intent = massive customer satisfaction failure despite $40M in projected savings
  • With autonomous agents running for hours or days, specifying intent upfront is not just a quality issue—it’s a supervision issue
  • The shift from chat-based AI to autonomous agents makes intent specification a fundamentally different discipline
  • Specification engineering = structured, formal documents handed to agents; distinct from expressing intent in prose
• Tenet alignment: Directly aligned with Human Intent First; the Klarna trap is a real-world example of intent drift at scale
• Quote: “Intent engineering sits above context engineering the way strategy sits above tactics. You can have perfect context and terrible intent — and the agent will execute exactly what you didn’t mean.”

Spec-Driven Development (Liu Shangqi, Thoughtworks, Dec 2025)

• URL: https://www.thoughtworks.com/en-us/insights/blog/agile-engineering-practices/spec-driven-development-unpacking-2025-new-engineering-practices
• Type: Industry blog (Thoughtworks)
• Key points:
  • Spec-driven development (SDD) = well-crafted software requirement specifications as prompts → AI generates executable code
  • A specification is more than a PRD: includes external behavior definitions (input/output mappings, preconditions/postconditions, invariants, interface types, integration contracts, state machines)
  • Domain-oriented ubiquitous language for business intent (not implementation)
  • Given/When/Then structure (from BDD); completeness + conciseness; clear structure for determinism
  • Semi-structured inputs significantly improve reasoning performance and reduce hallucinations
  • Separates planning/spec phase from implementation phase—human in the loop during spec review
  • “Spec drift and hallucination are inherently difficult to avoid”
  • Distinguishes prompt engineering (optimizes human-LLM interaction) from context engineering (optimizes agent-LLM interaction)
  • Specs compressed into context reduce the context engineering problem
• Tenet alignment: Aligns with Context as Infrastructure (specs as structured context); Human Intent First; raises Scalability tenet question (overhead of maintaining specs)
• Quote: “Specifications should use domain-oriented ubiquitous language to describe business intent rather than specific tech-bound implementations.”

Helping Users Update Intent Specifications for AI Memory at Scale (ACM CHI 2025)

• URL: https://dl.acm.org/doi/full/10.1145/3746059.3747778
• Type: Academic paper (ACM CHI)
• Key points:
  • As AI interactions continue across sessions, teams must “commit” new information to the intent specification
  • Intent specifications accumulate outdated or inconsistent dependencies that must be resolved
  • At scale, managing intent specifications becomes a knowledge management problem—not just a prompt problem
  • Compares to version control for human intent
• Tenet alignment: Directly aligned with Context as Infrastructure; raises Human Intent First question of how intent degrades or drifts over time

Examining LLMs’ Impact on Design (Zhou & Chen, 2025)

• URL: https://www.sciencedirect.com/science/article/pii/S3050741325000175
• Type: Systematic literature review (118 publications), Design and Artificial Intelligence journal
• Key points:
  • LLMs support all four design phases: Discover, Define, Develop, Deliver
  • Strengths: generating survey questionnaires, synthesizing evaluations, generating diverse ideas
  • Limitations: lack of nuanced design judgment, context misunderstanding, hallucination
  • Key finding: LLMs are advancing design through “context awareness” beyond discriminative AI predecessors
  • Highlights need to maintain designer competencies even as LLMs augment them
• Tenet alignment: Aligns with Symbiotic Intelligence over Automation; raises competency preservation concern
• Quote: “This research highlights opportunities for future studies to enable LLMs to complement designers rather than replace them.”

Major Positions

Position 1: Intent as Pre-Specification (Top-Down)

• Proponents: Thoughtworks (Liu Shangqi), Spec-Driven Development community, Nate’s Newsletter framework
• Core claim: Productive GenAI collaboration requires formalizing intent into structured specifications before AI engagement. The planning phase and implementation phase must be separated.
• Key arguments: Ambiguity compounds in autonomous systems; upfront specification reduces hallucination; domain-oriented ubiquitous language creates precision without verbosity; BDD patterns (Given/When/Then) transfer naturally
• Relation to tenets: Strong Human Intent First alignment; supports Context as Infrastructure; risk of over-specification constraining Pluralism; potential scalability concern

Position 2: Intent as Emergent Through Iteration (Bottom-Up)

• Proponents: Ding & Chan (IUI research), early vibe coding practitioners
• Core claim: For complex and creative tasks, intent cannot be fully pre-specified—it is discovered and refined through AI interaction. The model’s outputs themselves help practitioners articulate what they actually wanted.
• Key arguments: Novices especially cannot specify upfront what they don’t yet know they want; iterative refinement is cheaper than upfront specification; AI outputs serve as a thinking surface
• Relation to tenets: Supports Symbiotic Intelligence (AI as thinking partner); risks Human Intent First if users follow model suggestions uncritically

Position 3: Intent as Mediated Process (Relational)

• Proponents: Feng et al. (vibe coding research), Meske et al.
• Core claim: Intent specification is not a one-time act but an ongoing mediation—a redistribution of epistemic labor between human and machine across the entire workflow. “Designing the right intention” is at least as important as “intending the right design.”
• Key arguments: Vibe coding reconfigures who decides what, not just how fast execution happens; ownership attaches to intention-setting; AI drifts from intent requiring continuous alignment; the process reveals what you actually want
• Relation to tenets: Deepest alignment with Symbiotic Intelligence; Human Intent First is the generative constraint, not a solved precondition

Position 4: Intent as Persistent Infrastructure (Architectural)

• Proponents: Lockie (Intent Stack); ACM paper on AI memory
• Core claim: Intent should be treated as architecture, not per-session instruction. A hierarchical, living document of human intent provides persistent context across all AI interactions, eliminating the need to re-specify in each session.
• Key arguments: Per-session intent specification is repetitive and inconsistent; agents need durable purpose not just per-task instructions; intent at higher levels (why) liberates intent at lower levels (what/how); intent accumulates and requires version management
• Relation to tenets: Most direct expression of Context as Infrastructure tenet; natural home for Human Intent First as foundational principle

Key Debates

Debate 1: Upfront Specification vs. Iterative Refinement

• Sides: Spec-driven development (specify first) vs. IUI research and vibe coding (discover through iteration)
• Core disagreement: Can intent be fully articulated before the AI interaction begins, or does the AI interaction itself constitute part of the intent-formation process?
• Current state: Emerging consensus that task type determines approach—fixed-scope tasks benefit from upfront spec; complex creative tasks benefit from iteration; hybrid approaches are emerging

Debate 2: Precision vs. Productive Ambiguity

• Sides: NNG (precision is essential) vs. design practice (some ambiguity is generative)
• Core disagreement: Does making intent fully legible constrain the creative outputs that make AI collaboration valuable? Does precision reduce the serendipitous outputs that break fixation?
• Current state: Unresolved. Ding’s research shows novices benefit from broad intents for exploration; NNG shows precision improves quality. The tradeoff may be exploration-quality.

Debate 3: Intent at What Level of Abstraction?

• Sides: “Why” level (Intent Stack, strategy) vs. “What” level (BDD specs) vs. “How” level (code snippets, NNG)
• Core disagreement: At which level of abstraction should intent be made legible to AI? Higher abstraction = more durable but less actionable. Lower abstraction = more precise but over-constraining.
• Current state: The four-skill framework (Nate) and Intent Stack suggest these levels are complementary, not competing—practitioners need to operate at all levels

Debate 4: Who Holds Intent? (Individual vs. Team)

• Sides: Individual tools (PCD, Intent Stack) vs. team-level specifications (SDD, BDD)
• Core disagreement: Is intent specification a personal cognitive practice or an organizational governance practice?
• Current state: Both are needed; tension emerges at handoff points where individual intent must be legible to team AI systems

Historical Timeline

Potential Article Angles

Based on this research, an article could:

1. “The Intent Gap: Why AI Executes What You Say, Not What You Mean” — The gap between stated intent and actual intent is the central practitioner problem. Frames intent specification as a new professional competency for designers, product owners, and strategists. Aligns strongly with Human Intent First tenet; introduces the “Klarna trap” and intent drift as real-world risks. Could draw on the expertise gradient finding (experts over-specify, novices under-specify) to argue for deliberate middle-ground practice.

2. “Intent Specification as Design Material” — Following on from the “prompt as design material” angle, this piece treats intent itself (not the prompt) as the design artifact. Intent is not found but constructed—through reflection, constraints, examples, and iteration. Aligns with Symbiotic Intelligence: the act of specifying intent is itself a cognitively valuable process that AI can facilitate.

3. “From Why to What to How: Intent Architecture for AI Collaboration” — A practitioner-facing framework drawing on Intent Stack, SDD, and the four-discipline model. Proposes that practitioners need to operate at three levels (purpose/why, specification/what, implementation/how) and that conflating them produces both over-engineering and under-performance. Aligns with Context as Infrastructure and Scalability tenets.

When writing the article, follow obsidian/project/writing-style.md for:

• Named-anchor summary technique for forward references
• Background vs. novelty decisions (what to include/omit)
• Tenet alignment requirements
• LLM optimization (front-load important information)

Gaps in Research

• No philosophical/phenomenological treatment: No sources from philosophy of action, speech act theory, or philosophy of mind on what “intent” means in the context of probabilistic systems. Wittgenstein on meaning-in-use, Searle on speech acts, and Bratman on planning theory could inform a richer account.
• Role and strategy variation: No systematic study comparing how Systemic Designers, UX Strategists, and Product Owners specify intent differently. The vibe coding research (Feng et al.) covers UX practitioners but not strategic/systems roles.
• Long-horizon intent degradation: How does intent specified at session start degrade over a multi-hour or multi-session collaboration? The ACM memory paper points to this as a problem but offers no empirical study of degradation patterns.
• Cross-cultural and epistemological pluralism: All sources are from Western/English-language practice communities. No research on how different epistemic traditions approach intent articulation or whether Western-centric specification frameworks disadvantage non-Western practitioners.
• Non-verbal and embodied intent: All frameworks treat intent as linguistic. No research on how gesture, sketching, or other embodied channels could communicate intent that resists verbalization.
• Intent specification and power: Who gets to specify intent in organizational settings? The ACM memory paper raises version control but not governance or politics of intent ownership.

Citations

1. Ding, Z. (2024). Towards Intent-based User Interfaces: Charting the Design Space of Intent-AI Interactions Across Task Types. arXiv:2404.18196. https://arxiv.org/html/2404.18196v2

2. Feng, S. et al. (2025). Vibe Coding for UX Design: Understanding UX Professionals’ Perceptions of AI-Assisted Design and Development. arXiv:2509.10652. https://arxiv.org/html/2509.10652v1

3. Wang, H. (2025, December 5). Prompt to Design Interfaces: Why Vague Prompts Fail and How to Fix Them. Nielsen Norman Group. https://www.nngroup.com/articles/vague-prototyping/

4. Lockie, D. (2026, February). The Intent Stack: A New Design Space for Human-AI Collaboration. LinkedIn. https://www.linkedin.com/pulse/intent-stack-new-design-space-human-ai-collaboration-david-lockie-ljgue

5. [Nate] (2026, February 27). Prompting just split into 4 different skills. Nate’s Newsletter (paywalled). https://natesnewsletter.substack.com/p/prompting-just-split-into-4-different

6. Liu, S. (2025, December 4). Spec-driven development: Unpacking one of 2025’s key new AI-assisted engineering practices. Thoughtworks. https://www.thoughtworks.com/en-us/insights/blog/agile-engineering-practices/spec-driven-development-unpacking-2025-new-engineering-practices

7. [Authors] (2025). Helping Users Update Intent Specifications for AI Memory at Scale. ACM CHI 2025. https://dl.acm.org/doi/full/10.1145/3746059.3747778

8. Zhou, Y. & Chen, C.-H. (2025). Examining the impact of large language models on design: Functions, strengths, limitations, and roles. Design and Artificial Intelligence, 1(2), 100017. https://doi.org/10.1016/j.daai.2025.100017

9. Meske, C. et al. (2025). Vibe Coding as a Reconfiguration of Intent Mediation in Software Development: Definition, Implications, and Research Agenda. arXiv:2507.21928.