This paper, by an author at LinearB and published for VibeX 2026 (Glasgow, June 2026), argues that the bottleneck in agentic coding is not code generation but alignment between developer intent and agent output. It traces that alignment problem to Vibe Coding, which it describes as the dominant workflow pattern: the developer states an intent, the agent produces code, and misalignments are resolved through iterative correction. The author’s claim is that the alignment problem is fundamentally a preparation problem, and the paper’s answer is Mise en Place (Agentic Coding Methodology) — a preparation-first methodology named after the professional-kitchen practice of having every ingredient measured and every tool positioned before cooking begins.
The paper makes four contributions: the three-phase methodology, grounded in backward design and in the externalization of tacit knowledge; a case study of applying it at a competitive hackathon; the concept of Context Fluency; and a research agenda of five open questions. It is explicit that the ingredients of the methodology are individually well established and that its contribution is their integration into one phase-gated sequence completed before implementation.
Key Points
- The methodology has three sequential phases whose artifacts feed the next: contextual grounding, which externalizes domain expertise and tacit knowledge into briefing documents agents consult; collaborative specification, in which human-agent dialogue produces a design document that records value judgments as constraints and captures why as well as what; and task decomposition, which turns the specification into dependency-aware task records so that several agents can work in parallel without coordination overhead, followed by integration verification against the specification.
- The paper positions the methodology against neighbouring approaches: against Spec-driven development it adds a phase for tacit, value-laden knowledge that specifications alone do not capture; against Prompt engineering it works at the scope of the workflow rather than the single invocation; and against iterative or vibe-coding workflows it front-loads alignment work that those flows pay incrementally as rework.
- In the hackathon case study, held in January 2026 with about 12 teams and a five-hour window, the author spent roughly two hours preparing while most teams began coding within fifteen minutes. Preparation produced 10 planning documents totalling 9,386 words and 64 task records with dependencies; four parallel subagents then worked across distinct feature areas, closing task records at a median of 5.9 minutes each, and the final codebase of 8,496 lines of TypeScript was deployed to production as a full-stack educational platform.
- The author reports that no structural refactoring was needed during deployment — the bugs that emerged were integration and styling issues — and that bug-type task records resolved in a median 1.2 minutes against 9.7 minutes for implementation tasks, and reads near-zero architectural rework as consistent with the hypothesis that rich upfront context reduces agent misalignment, while stating that causation cannot be established from a single case.
- Across the twelve-team field, workflow style split roughly in half between vibe-coding pitches and teams that opened with explicit planning, and only the author’s team fanned out parallel subagents from a decomposed plan; the paper does not link these patterns to outcomes.
- The research agenda asks whether deliberate preparation reduces agent misalignment compared with iterative development, what counts as sufficient context and when to stop preparing, how context fluency varies across practitioners, domains and models, whether the method scales beyond prototyping, and how preparation time relates to implementation quality.
Notes
For task decomposition the author used Beads, which the paper describes as lightweight Git-backed JSON records carrying priorities, dependencies and acceptance criteria (see Beads), while noting that the principle generalizes. It also cites the Research-Plan-Implement methodology as the closest prior work, differing from it by formalizing tacit-knowledge externalization and by requiring all three phases to finish before implementation begins.
The author states three limitations: the evidence is a single five-hour hackathon with one practitioner, which is illustrative rather than generalizable; there was no control group or iteration-first comparison and the other eleven teams were not instrumented; and the author’s own background in both education and software engineering is a confound that cannot be separated from the method’s contribution. The paper releases an artifact bundle with the anonymized hackathon data, prompts, agent configurations and context-engineering scaffolds on Zenodo. The version extracted here is arXiv:2605.05400v1.