Recoleta
GitHub
Source note

Context Engineering: A Practitioner Methodology for Structured Human-AI Collaboration

Authors
Elias Calboreanu
Source
arXiv
Published
Collection
Software Intelligence
Context EngineeringHuman AI CollaborationPrompting MethodologyWorkflow DesignLLM Operations
Open arXivMarkdown source

Summary

This paper argues that AI output quality depends more on complete context than on prompt wording. It proposes a practitioner method for packaging context and routing work through fixed stages, with observational evidence from 200 human-AI interactions.

Problem

  • Ad hoc prompting often leaves out requirements, examples, constraints, and quality criteria, which leads to extra revision cycles and weak first drafts.
  • The paper claims published prompt guides and agent frameworks do not give practitioners a clear human-side method for assembling, ranking, and sequencing context across tasks.
  • This matters because missing context wastes operator time and makes AI output less reliable in professional workflows.

Approach

  • The method defines a four-stage pipeline: Reviewer -> Design -> Builder -> Auditor. Each stage has a separate job: extract requirements, create a plan, produce the artifact, and check it against the plan.
  • It defines a five-role context package with fixed priority: Authority, Exemplar, Constraint, Rubric, Metadata. Higher-priority items resolve conflicts when instructions disagree.
  • A central rule is that the design output becomes the main Authority document for downstream stages, so builders execute against a specification instead of improvising from the latest prompt.
  • The paper also introduces Operator Authority, a versioned file of recurring user standards such as tone, formatting, and quality rules, so the model does not have to learn them through repeated corrections.
  • The evidence is an observational study of 200 documented interactions across Claude, ChatGPT, Cowork, and Codex over four months, plus mention of a companion production automation system with 2,132 classified tickets.

Results

  • Incomplete context was associated with 72% of iteration cycles in the 200-interaction dataset.
  • Structured context assembly was associated with a drop in average iteration cycles from 3.8 to 2.0 per task.
  • First-pass acceptance improved from 32% to 55%.
  • Among structured interactions, 110 of 200 were accepted on the first pass, compared with 16 of 50 baseline interactions.
  • With iteration allowed, final success reached 91.5% (183 of 200).
  • The paper states these results are observational, come from a single-operator dataset, and do not establish a controlled causal comparison.

Link