Tips & Best Practices
Practical advice from building 76 skills and running 10 autonomous agents
Essential Tips
Lessons learned from real-world usage.
Context is King
The AI is only as good as its context. CLAUDE.md, rules, and skill
instructions shape quality more than any prompt engineering trick. Invest time in
these files — they compound in value across every session.
Start Small
Don’t run /dx-agent-all on day one. Start with the building blocks:
/dx-req <id> # full requirements pipeline /dx-plan # generate a plan /dx-step # execute a step
Review, Don't Trust Blindly
AI makes mistakes. Code review exists for a reason. /dx-step-verify runs
6 verification phases for exactly this reason — build check, lint check, code review,
test verification, accessibility scan, and visual comparison.
Config Over Code
Use .ai/config.yaml for project-specific values. Never hardcode build
commands, branch names, or org URLs in skills.
build: command: "cd ui.frontend && npm run build" scm: base-branch: development
Model Tiering
Use the cheapest model that can do the job well. Cost differs 10x between tiers.
Opus ($$$)
Deep reasoning — code review, implementation planning, root cause analysis. Use when quality of judgment matters most.
Sonnet ($$)
Execution — step implementation, PR review, AEM inspection. Good judgment at moderate cost.
Haiku ($)
Lookups — file resolution, doc search, test running. Speed matters more than depth.
Tool-Specific Tips
Small details that save hours of debugging.
Shell Env for Copilot CLI
MCP secrets must be in ~/.bashrc (or ~/.zshrc). Copilot CLI
does not read .claude/settings.local.json. If your MCP connections fail
in Copilot but work in Claude Code, check your shell environment.
Dual Frontmatter in Rules
Rules need both paths: and applyTo: to work in both tools.
Forgetting one means the rule silently doesn’t load in that tool.
--- paths: ["**/*.js"] applyTo: "**/*.js" ---
allowed-tools in Skills
Add allowed-tools to skill frontmatter to avoid excessive permission
prompts in Copilot CLI. Without it, every tool call asks for approval.
--- allowed-tools: [read, edit, grep, bash, mcp] ---
Pre-flight Checks
Before running multi-step pipelines, verify: (1) all inputs exist (URLs, ticket IDs), (2) MCP servers respond, (3) target branch exists. Fail fast rather than burning through steps and tokens.