Agent
Use this agent when you need to review a pull request for test coverage quality and completeness. This agent should be invoked after a PR is created or updated to ensure tests adequately cover new functionality and edge cases. Examples:\n\n \nContext: A developer has just created a pull request with new…
Agent
QA engineer. Runs tests, lint, and build checks. Validates PRs before merge. Reports failures to team lead. Does NOT modify code.
Agent
For a decision (not code), create an agent team whose members hold distinct personas, skeptic, customer, competitor, domain expert, have each research independently, then make them debate until they reach consensus or name their core disagreement.
Agent
File conflicts between agents are the primary failure mode of agent teams; design team roles around domain isolation, not component ownership.
Agent
Teams enable agent-to-agent messaging; they cost 4-7x more tokens than a single session, reserve for genuinely collaborative multi-domain work.
Agent
Claude Code removed the interactive /agents wizard; create or edit subagents by writing files in .claude/agents/ directly or by asking Claude to do it.
Agent
Naming subagents 'designer' / 'backend' / 'tester' creates coordination overhead via the main agent; use agent teams for multi-role work instead.
Agent
TL;DR: When building CLI tools that Claude Code will drive, design the --help text to be self-contained so the agent can learn the whole tool from running it.
Agent
A subagent that hasn't seen your existing code generates more distinct designs; the main session's code context biases toward what already exists.
Agent
Forked subagents inherit the full conversation history (and share its prompt cache), so they preserve nuance that normal subagents compress away.
Agent
When an agent hits a malformed state, its default is to pile on a tolerant parser, then a fallback, then a migration, then debug output. The correct fix is usually to enforce an invariant upstream so the bad state can never occur.
Agent
For questions with second- and third-order consequences (a change → its implications → the blast radius of each implication), have layer-1 subagents each spawn layer-2 subagents, so every level of analysis runs in its own isolated context and only clean conclusions roll back up.
Agent
When you want a coding agent to fix slow code, don't dump raw profiler output, give it a single tool that aggregates samples (frequencies, percentiles) into an analyzed view, because LLMs are bad at doing those computations themselves.
Agent
Give teammates small, regularly-delivered tasks and include a devil's-advocate / security agent that critiques peers in real time.
Agent
Split generation from verification across two agents: give the evaluator browser automation to click through the live app like a user and concrete runtime pass-thresholds, rather than letting the agent that wrote the code grade itself.
Agent
When ending an agent-team session, send each teammate a 'save your work and confirm when ready' message and wait for the confirmation, rather than force-killing them mid-task.
Agent
The TeamCreate/TeamDelete tools have been removed; with CLAUDECODEEXPERIMENTALAGENTTEAMS=1 every session has one implicit team and you spawn teammates simply by passing a name to the Agent tool.
Agent
Subagents keep main context lean by handling research, log analysis, and exploration; they are poor at large edits due to compression loss.
Command
Prefer giving agents small CLI tools over MCPs: a CLI is read once then driven by command, is deterministic, self-documents via --help, and costs far less context than an MCP's loaded tool schemas.
Command
Package the "ask before you build" habit into a reusable custom slash command: it captures the task, asks 5 clarifying questions, drafts a plan, and waits for approval before executing.
Command
/effort sets how deeply the model reasons: low is fast and shallow, the default is high, and extra-high/max are slower but markedly more careful, switch up for risky or complex changes, back down for routine ones, and check the cost with /usage afterwards.
Command
The environment selector at the bottom of Claude Code switches where work executes: local for fast interactive iteration, cloud (Anthropic servers) for routines that must run with your machine off, SSH for driving a remote box like a VPS.
Command
TL;DR: Embed bash in a slash command to pre-compute state (e.g. git status) so the model gets the context for free instead of round-tripping for it.
Skill Claude CodeCodex
Audit and improve CLAUDE.md files. Use when asked to "check CLAUDE.md", "audit project instructions", "improve CLAUDE.md", "fix CLAUDE.md", or "CLAUDE.md maintenance". Scans all CLAUDE.md files, scores quality, outputs report, then applies targeted updates.