.png){kind=icon}
.png){kind=icon}
Return
A Skill-Mediated LLM Workflow for Migrating a.NET Monolith to Cloud-Deployed Microservices: a Controlled Case Study
.png){kind=icon}
OA
AI
DOI：10.1109/access.2026.3712587.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Migrating a monolithic application to cloud-deployed microservices requires orchestrating a full sequence of dependent steps - decomposition, testing, review, and deployment - that current LLM-based coding assistants support only in isolation. We report a controlled, multi-case study of one composition pattern addressing this gap: ten Claude Code skills chained into a single end-to-end workflow for migrating.NET monoliths, evaluated on PetRescue (primary case) and Microsoft’s eShopOnWeb (transfer case, scored against the active dotnet/eShop reference). On PetRescue, nine of ten attempted skill-mediated runs completed, scoring 12.11 of 20 (60.6%, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sigma =3.9\%$ </tex-math></inline-formula>) on a pre-registered twenty-dimension structural rubric; the tenth exhibited a skill-invocation failure. A raw LLM with a fully-specified prompt scored 14.80 of 20 (74.0%, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sigma =3.2\%$ </tex-math></inline-formula>). All scores measure agreement with one independently-engineered expert reference, not absolute migration quality. Under a matched-specificity baseline with a deliberately vague one-sentence prompt, four of five attempted runs completed, scoring 12.75 of 20 (63.8%, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sigma =2.5\%$ </tex-math></inline-formula>); the fifth exited prematurely. We detected no significant difference between the skill-mediated and matched-specificity conditions (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$t=-1.77$ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$p \approx 0.11$ </tex-math></inline-formula>; <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$N=9$ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$N=4$ </tex-math></inline-formula> successful runs). The 13.4-percentage-point skill-vs-baseline gap therefore decomposes into a significant 10.3-point prompt-specificity effect (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$t=-6.40$ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$p\lt 0.001$ </tex-math></inline-formula>) and a 3.2-point skill-abstraction effect that does not reach significance at <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\alpha =0.05$ </tex-math></inline-formula>. The transfer case reproduces the gap (−13.3 vs. −13.4 points) on a substantially different codebase. The workflow’s automated security check shows perfect recall (10/10) but 52.6% precision and 10% specificity on a matched 20-file corpus, consistent with automated DevOps hygiene rather than formal security validation. The two short-prompt conditions fail in qualitatively distinct ways (skill-invocation vs. premature exit). The contribution is not higher single-shot LLM output but reproducible workflow engineering: auditable SDLC orchestration, artefact-mediated state, mechanical pre-commit gating, and lower prompt-engineering burden. We publish skill scripts, both codebases, run logs, scoring code, the verifier changelog, and the injection corpus through GitHub and Zenodo.
Keywords:
LLM-assisted software engineering
microservices migration
monolith decomposition
reproducible case study
ablation study
empirical evaluation
.NET
Oracle Cloud Infrastructure