Claude Code-only features — Skills, Subagents, Hooks, Plan Mode explained
Series part 6
Ever wondered what those extra menu items in Claude Code actually do — and quietly never used them?
Part 5 covered five instruction techniques. This post is about Claude Code–specific features, and how to decide which one to reach for when.
Feature list
From the official docs:
| Feature | What it does |
|---|---|
| Skills | Inject reusable how-to guides / domain knowledge |
| Subagents | Spawn a separate agent to run a subtask in isolation |
| Hooks | Auto-run scripts on tool invocations |
| Plan Mode | Force Claude to plan without writing code |
| Auto Mode | Run non-interactively |
| Slash Commands | /clear, /compact, /rewind, etc. |
Let Claude interview you — the reverse-interview trick
Surprisingly underused, and explicitly recommended by Anthropic:
When starting a complex task, ask Claude to ask you clarifying questions before implementation.
Instead of front-loading all context, ask Claude to list its questions first:
I want to implement {{goal}}.
Before we start, please list up to 5 questions you'd want me to answer first.I use this every time I design a new feature. It surfaces assumptions I never spelled out.
Offload research to Subagents
For heavy exploration, delegate to a subagent so the main session’s context stays focused.
Delegate exploratory research to subagents so the main agent’s context stays focused on execution.
“Find every instance of pattern X in the codebase” is the perfect Subagent task — you only want the summary back in your main session.
The Writer / Reviewer pattern
Officially recommended collaboration mode:
Have one Claude session write code, then start a fresh Claude session to review it.
Separate the session that writes from the session that reviews.
| Role | Session |
|---|---|
| Writer | Dedicated to generation; done when tests pass |
| Reviewer | A fresh /cleared session; reviews the code from the outside |
An outsider’s eye catches things the Writer couldn’t.
Non-interactive automation
With the --print flag, Claude Code runs without interaction — ideal for CI.
Use
I’m designing a CI flow where “Claude auto-reviews every PR and posts comments” based on this. If it ships, human review load drops another tier.
Anti-patterns the docs explicitly warn against
| ❌ Anti-pattern | Why |
|---|---|
| Working for hours in the same session | Context pollution → /clear |
| Vague “make it nice” asks | Violates the Golden Rule |
| Marking tasks done without verification | Hallucination breeding ground |
| Skipping Plan Mode | Right solution to the wrong problem |
Giant CLAUDE.md | Ironically, more context pressure |
The giant CLAUDE.md pitfall is easy to fall into. The official guidance: keep it around 150 lines, and push details to Skills.
How I combine features on HomePage
| Tool | How I use it |
|---|---|
CLAUDE.md (rules only) | ~150 lines, project-wide guardrails |
.claude/skills/ | ”Add a blog post” / “SEO check” procedures |
| Hooks | Stop-hook runs tests + pre-commit checks automatically |
| Plan Mode | Mandatory for any multi-file change |
/clear | Always after each task |
The net effect: “repeating the same instructions” has disappeared from my workflow.
Cost-reduction priority (my interpretation)
Ranked by impact on Claude’s token usage:
| Rank | Lever | Effect |
|---|---|---|
| 1 | Disciplined /clear | Kills context pollution at the source |
| 2 | Plan Mode | Prevents wasted rewrites |
| 3 | Well-maintained CLAUDE.md | Eliminates repeat instructions |
| 4 | Subagents for research | Keeps the main agent focused |
| 5 | Upfront verification criteria | Fewer iteration cycles |
Which feature for which problem
| Problem | Feature |
|---|---|
| Repeating the same instructions | CLAUDE.md |
| Complex tasks going off the rails | Plan Mode |
| Session feels like it’s lost its edge | /clear |
| Huge investigation task | Subagents |
| Want to review what Claude wrote | Writer / Reviewer |
| Need to run inside CI | Auto Mode |
If you’ve been working in a single, ever-growing session for weeks, you might try just one disciplined /clear tomorrow morning. The contrast can be surprisingly stark.
Related posts
- Writing instructions Claude gets right on the first try — instruction techniques
- Never let Claude jump straight into code — the EPCC workflow — Plan Mode in depth
- Reference-grade Qiita version: Qiita profile