Agent Skills: Engineer-Like Process for AI Coders
Agent Skills encodes senior-engineer workflows into 7 markdown commands (/spec, /plan, etc.) and specialist personas, enforcing specs, testing, and review to make AI agents reliable—portable to tools like Verdent.
Embed Senior-Engineer Discipline via 7 Core Commands
AI coding agents fail by skipping specs, planning, testing, and reviews despite writing decent code—Agent Skills fixes this with a structured lifecycle mirroring careful engineers: define, plan, build incrementally, test, review, simplify, ship. Invoke via 7 slash commands: /spec for spec-driven development and idea refinement; /plan for task breakdown and sequencing; /build for incremental slices; /test for test-driven verification; /review for code quality checks; /code-simplify to prioritize simplicity over cleverness; /ship for CI/CD and deployment. Supporting skills cover API design, frontend, debugging, security, performance, and docs. Load behaviors sequentially—spec first, then plan, build/test in parallel—to avoid context overload, as dumping all skills creates noise. This pushes small, verifiable tasks, evidence-based testing, pre-merge reviews, and simplicity, reducing confident-but-sloppy outputs.
Leverage Specialist Personas to Catch Hidden Issues
Single agents miss maintainability, coverage gaps, or security flaws—use dedicated personas like code reviewer (for readability/maintainability), test engineer (for verification/coverage), and security auditor (for vulnerabilities). After building, route to these for targeted inspection: reviewer flags over-engineering; tester demands proof via tests; auditor scans exploits. This multi-perspective approach outperforms one-agent self-review, mimicking team dynamics. Start minimally with spec-driven dev, TDD, and review—these fix 80% of AI failures—then add project-specific skills like UI engineering or hardening. Free/open-source markdown format ensures portability across Claude Code, Cursor, Gemini CLI, WinSurf, OpenCode, Copilot, without extra subs; costs tie only to your LLM.
Adapt to Verdent for Native Orchestration
Port workflows into Verdent's rules: verdent.md for universal habits (spec before code, verify changes, prefer simple); agents.md for project rules (small tasks, evidence verification, focused diffs). Shape plans via plan rules to include scope, criteria, sequencing, verification, rollbacks. Map personas to custom sub-agents—reviewer, tester, security—invoked post-build. Exploit parallel workspaces: one for /build, another /test/verify, third /review/alt-impl, with isolated git trees preventing interference. Once spec exists, parallelize for speed; this leverages Verdent's orchestration over literal command ports, yielding disciplined, reliable outputs. Process trumps model: strong workflow + decent LLM > top model + slop.