A Lean and Spec-Driven AI-Assisted Software Development Lifecycle for Applied AI Education: The AI-SDLC Approach
Andreas Martin, Sandro Schwander
cs.SE · cs.AI
Quick verdict
AI-SDLC: a six-phase spec-driven lifecycle for agentic coding, evaluated by a small single-cohort survey (n=13); perceived value high, comprehension gap open.
Summary
The paper introduces AI-SDLC, a lightweight, spec-driven software development lifecycle designed for AI coding agents, developed for the FHNW bachelor course 'AI-Assisted Software Development'. The lifecycle organizes work into six phases — Bootstrap, Specify, Design, Develop, Validate, Deploy — and combines established practices (TDD, testing pyramid, Clean Architecture, CI/CD, Kanban, selected SAFe concepts) with repository-local guidance: a compact AGENTS.md, phase-specific SKILL.md files, docs/PROJECT.md, docs/TASKS.md, use case specifications, and optional ADRs, all version-controlled so agents operate with bounded autonomy under reviewable artifacts. Evaluation is a cross-sectional survey of one course cohort: of 34 invited students, 13 completed the questionnaire; closed rating items plus open questions were analyzed descriptively with reverse-worded control items and no inferential statistics. Reception was positive: all 13 agreed Specify sharpened requirements, 12 of 13 would reuse the process, median recommendation 8 of 10 with no detractors, and phase usefulness means ranged 3.77-4.42. The most valuable finding is a comprehension gap: students trusted test-verified output while rating their ability to follow the code lowest, and 11 of 13 described adopting generated code without understanding it. Key limitations — single cohort, n=13, self-report only, single-coder thematic analysis, instructor-designed evaluation — are largely acknowledged by the authors; the lifecycle template is available on Zenodo.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- AI-SDLC — six phases (Bootstrap, Specify, Design, Develop, Validate, Deploy) with repository-local AGENTS.md, phase-specific SKILL.md files, and persistent spec/task/state artifacts — enables AI coding agents to operate with bounded autonomy inside an explicit, reviewable, test-oriented lifecycle (Section 2, Fig. 1).
- In a single-cohort survey, students reported strong perceived value: 13/13 agreed Specify sharpened requirements, 12/13 would reuse the process, median recommendation 8/10 with no detractors, and phase usefulness means 3.77-4.42 (Section 4).
- Repository-local instruction files were judged effective (0/13 agreed they had little noticeable effect), but free-text accounts show they work neither automatically nor without maintenance (Section 5).
- Students' verification shifted from reading code to trusting tests, yielding confidence without comprehension: 11/13 described adopting generated code without fully understanding it (Sections 4-5).
Strengths
Candid reconciliation of closed and open-ended data
Section 4-5 explicitly surface rating/free-text discrepancies: 6/13 admitted un-understood adoption on the closed item but 11/13 described such situations in free text; PR review was agreed by 7 on closed items but mentioned by only 1 in free text; the conclusion concedes free text contained more instruction failures than successes — honesty that exceeds typical workshop-paper practice.
Methodologically proportionate analysis
Section 3.2 restricts itself to descriptive frequencies and absolute counts given n=13 ordinal items; no composite scores; at least one reverse-worded item per block counters acquiescence (citing Weijters & Baumgartner); Section D separates adherence from perceived usefulness following Carroll et al. and Davis.
Versioned, inspectable artifact
Transferable core finding
Transparent survey reporting
Weaknesses
Single-cohort, self-selected n=13 with no comparator
Single-coder thematic analysis without reliability checks
Perception-only outcome measures
Instructor-designed evaluation of the instructors' own method
Framework novelty is compositional and unattributed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: AI-SDLC (Martin & Schwander, arXiv:2609.24348)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is AI-SDLC?
AI-SDLC is a lightweight, spec-driven software development lifecycle for AI coding agents, described in a 2026 FHNW paper by Martin and Schwander. It defines six phases — Bootstrap, Specify, Design, Develop, Validate, Deploy — and stores all guidance in version-controlled repository files: a compact AGENTS.md as entry point, per-phase SKILL.md skill files, project docs, task lists, and use case specifications. This lets AI agents work with substantial autonomy inside a phase while humans keep control by reviewing and editing the persistent artifacts. The template is archived on Zenodo (v1.0.0, September 2026).
What did the student evaluation of AI-SDLC find?
A survey of 13 students from one FHNW cohort (of 34 invited) found strong positive perceptions: all 13 agreed the Specify phase forced sharper requirements before AI implementation, 12 of 13 would use the process again, the median recommendation score was 8 of 10 with no detractors, and every phase was rated useful (means 3.77-4.42). Downsides: perceived overhead on small tasks, and a comprehension gap — students trusted test-verified code without understanding it; 11 of 13 admitted adopting generated code they could not fully follow.
What are the main limitations of the AI-SDLC paper?
It evaluates one cohort of 13 self-selected respondents from a single course, most with little prior software engineering experience; all outcomes are self-reported perceptions with no objective code-quality or repository measures; open answers were coded by a single author without reliability checks; and the course instructors evaluated their own method. The authors state the results are descriptive and exploratory, not generalizable.
How does AI-SDLC compare to GitHub spec-kit or Kiro?
AI-SDLC belongs to the same spec-driven-development family — the paper cites GitHub spec-kit, Kiro, and Tessl as related approaches — but combines specifications with an explicit six-phase lifecycle, TDD-first development, per-phase agent skill files, and persistent task/spec/state documents (AGENTS.md, TASKS.md, PROJECT.md, ADRs). The paper presents no empirical comparison against those tools, so any practical superiority claim would be unsupported.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures appear in the extraction as raw number/label sequences (Figs. 1-3); exact per-item distributions could not be fully re-verified from text alone.
- The survey instrument (German questionnaire) is not stated as archived; this was judged from the text, which is otherwise complete through the references.
- codeAvailable=true refers to the Zenodo-archived lifecycle artifacts (two DOIs); survey data and analysis code availability are not stated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.