# How I add new skills to Claude Code without breaking things

Most people just type what they want and let Claude rip. That works for one-off tasks. It does not work when you are adding a new skill, a new agent, or a new piece of infrastructure to a workspace you actually use.

Two habits fix that. Both take seconds. Both compound over months.

## Habit 1. Always be in plan mode for new builds

Plan mode tells Claude to think the work through before it touches a file. You get a plan back. You read it. You catch the wrong assumption, the missing file, the part where Claude was going to do twice as much as you asked. Then you approve.

Hit `Shift+Tab` until the bottom bar says `plan mode`. Then paste the build prompt.

```text
I have a new skill saved at <path>. Read the SKILL.md and the scripts.
Plan how to install it into my Claude Code workspace cleanly. Match the
existing skill folder conventions. Flag anything Nic-specific that needs
adapting. Do not touch any files yet.
```

You will get a structured plan back. Read it. If Claude asks follow-up questions (it usually does for non-trivial builds), answer them carefully. If a question looks like it does not understand, click "Other" and tell it what you actually want.

When the plan looks right, click `Auto Accept` and then switch to `Auto mode`. Now it runs.

## Habit 2. Keep a build log at the root of your workspace

Every meaningful build in my workspace gets one row in `BUILD_LOG.md`. Status emoji, slug, one-line description. That is it.

```text
- 🟢 /library-build - 2026-05-22 - Website library replaces Drive
- 🟡 /infographic-builder - 2026-05-22 - In phase 2 of 2
- 🔴 /weekly-sync - blocked on Linear OAuth
- ⚪ /follow-up retainer wrap - queued
```

Plain markdown. Sorted by status. When I open the workspace I see what is live, what is in flight, what is stuck, what is next. Friction is zero, so the log actually stays current.

Claude does not remember what you built last week. The build log is how the workspace remembers for you.

## Habit 3 (bonus). Phase your big builds

When Claude makes the plan, ask it to build in phases if the scope is bigger than one sitting. After each phase Claude stops, shows you what it built, and waits for you to test before moving on. Bugs surface early instead of three phases deep.

```text
This build is big. Break it into 2 or 3 phases. After each phase, stop
and show me what shipped so I can test before you start the next phase.
Log each phase to BUILD_LOG.md as you go.
```

## Why this matters

Without plan mode, Claude will sometimes do the wrong thing fast. You waste tokens, you waste time, you have to revert.

Without a build log, you forget what you built. The half-finished skill from three weeks ago sits there. The blocked one never unblocks. You end up rebuilding things you already built.

The workspace that knows what it is doing beats the one that does not, every single time.

## Open this in Claude Code

The prompts above are copy-paste ready. Save them to a snippet manager or drop them into a `.claude/prompts/` folder. Then every new build starts with the same two habits.

If you want the full `/builds` skill that ships with this pattern (idea queue, status flips, "next" ranker, automatic Recent ships from ship-recap hook), it is part of my Skills Library.