AI-DLC: Personal AI-Assisted Engineering Workflow
- Source
- Personal Workflow
- Service
- AI-DLC
- Role
- Workflow Owner
- Scope
- MarkdownCodexClaude CodeHuman-in-the-loop
Overview
AI-DLC is a personal engineering workflow I defined to integrate AI into the development lifecycle, from work intake and analysis to verification, communication, and knowledge updates.
Instead of repeatedly re-explaining recurring integration issues, test failures, or data reflection problems, the workflow turns work context, decision rationale, and verification steps into Markdown-based knowledge that Codex and Claude Code can reference consistently.
The core principle is human-in-the-loop engineering: AI proposes root-cause candidates and verification order, while the final judgment is validated through code, queries, batch flow, and test results.
Problem
Real-world engineering issues often span specifications, code paths, data state, batch jobs, and external communication, so the same type of issue can be interpreted differently across tools and sessions.
At the same time, sensitive operational details cannot be exposed directly. The workflow needed a clear boundary between AI-readable work context and internal evidence that must be verified by the engineer.
My Role
- Structured integration specs, code flow, test cases, and error patterns as AI-readable Markdown knowledge
- Organized entry routing, agent rules, domain documents, and access boundaries so Codex and Claude Code can start from the same work context
- Re-verified AI-suggested root-cause candidates against controller flow, service logic, batch jobs, SQL, database state, and test results
- Separated public-facing explanations from internal implementation details across partner emails, Jira comments, and wiki drafts
Key Contributions
- Context: organized work requests, related specs, code flow, and data state into a single traceable working context
- Analyze: grouped root-cause candidates by layer and narrowed the evidence and verification order before touching code
- Verify: converted AI suggestions into falsifiable checks against code, queries, and tests instead of treating them as conclusions
- Communicate: rewrote AI-generated drafts into external-safe messages focused on reproduction details, confirmation requests, and next actions
- Learn: fed new error cases and AI misjudgments back into the Markdown knowledge base as reusable decision rationale
Context Routing
- Defined entry routing so AI tools can locate the most relevant work context first based on the question type
- Used agent rules to keep read/write behavior, pre-change checks, and knowledge-update conditions consistent across tools
- Separated specification documents from domain knowledge so fields, response codes, test cases, and system flow can be referenced by purpose
- Maintained access boundaries that explain work flow and decision rationale without exposing connection values, permissions, or raw internal logs
Human-in-the-loop Verification
- Used AI responses only as root-cause candidates and verification order, then validated request values, connection conditions, and data reflection conditions through code, queries, and tests before adding confirmed checks to the knowledge base
- In a data-reflection issue, AI initially pointed to a connection condition; batch reprocessing flow and data-state checks disproved that candidate, identified the actual failure layer, and turned the verification order into a reusable checklist
- Classified whether an issue belonged to API, batch, data, network, or test procedure first, so the investigation scope did not expand unnecessarily
- Defined communication rules that separate partner-safe explanations from internal implementation notes
Results
- Reduced initial root-cause candidate discovery from hours to minutes for data issues that require tracing database procedures and Java code flow together, using error-data-driven AI candidate analysis while validating final judgment through code, queries, and tests
- Narrowed the initial analysis scope for data-consistency issues that require manual code and data comparison by organizing comparison points and verification paths first
- Turned AI-suggested causes into code, query, and test checkpoints to narrow the investigation scope and reuse the verification flow in similar cases
- Accumulated AI misjudgments and disproven candidates in the knowledge base as conditions to check earlier in future analysis
- Improved the speed and quality of external communication drafts by structuring them around reproduction details, confirmation requests, and next actions