Claude Certified Architect – Foundations

Skills

• — Design agentic loops for autonomous execution
• — Orchestrate coordinator–subagent systems and configure spawning and context passing
• — Enforce multi-step workflows with handoff patterns
• — Use Agent SDK hooks to intercept tool calls and normalise data
• — Manage session state, resumption and forking

In practice

• A research system where a coordinator plans, three subagents search in isolation, and a synthesiser writes the report keeps each context small and lets the coordinator retry one branch without restarting the whole job.
• A hook that rewrites every date a tool returns into ISO format before it enters context removes a class of downstream parsing failures.

Skills

• — Write tool interfaces with clear descriptions and boundaries
• — Return structured errors from MCP tools
• — Distribute tools across agents and configure tool choice
• — Integrate MCP servers into Claude Code and agent workflows
• — Use Read, Write, Edit, Bash, Grep and Glob appropriately

In practice

• A support agent with 40 tools uses them worse than one with 8 scoped tools; giving the refund tool only to the escalation subagent is both a quality and a safety decision.
• An MCP tool that returns `{error: 'RATE_LIMITED', retry_after: 30}` lets the agent wait; one that returns a 500 string makes it guess.

Skills

• — Structure CLAUDE.md with hierarchy, scoping and modular organisation
• — Create custom slash commands and skills
• — Apply path-specific rules
• — Decide when to use plan mode
• — Integrate Claude Code into CI/CD pipelines

In practice

• A `/release-notes` command that reads the changelog and opens a PR encodes a team process once instead of relying on each engineer's prompt.
• Running Claude Code headless in CI to review every PR against a repository's conventions file is a Domain 3 pattern; deciding what it may and may not change is a Domain 5 one.

Skills

• — Design prompts with explicit criteria to cut false positives
• — Use few-shot examples for consistency
• — Enforce structured output with tool use and JSON schemas
• — Build validation, retry and feedback loops
• — Design batch strategies and multi-instance, multi-pass review

In practice

• An extraction pipeline that runs two independent passes and flags disagreements for human review catches errors a single confident pass never surfaces.
• Defining 'high-risk clause' with five explicit criteria and three labelled examples halves the false-positive rate compared with 'flag anything risky'.

Skills

• — Preserve critical information across long interactions
• — Design escalation and ambiguity-resolution patterns
• — Propagate errors across multi-agent systems
• — Manage context in large-codebase exploration
• — Design human review with confidence calibration
• — Preserve provenance and handle uncertainty in multi-source synthesis

In practice

• A subagent that fails should return a typed failure the coordinator can reason about, not an empty string that looks like 'nothing found'.
• A synthesis report that attributes every claim to a source and marks conflicts between sources is what a reviewer can actually check.