The Claude Code prompt playbook was hiding in the official docs — kicking off a new series
Are you using Claude Code “by feel”?
Ever had that moment where Claude nails a hard task one minute and then completely whiffs an easy one the next — and you start blaming the weather?
I run three personal products powered by AI-driven development, and I lean on Claude Code heavily for all of them. For a long time I was using it purely on instinct — “Claude feels smart today” or “Claude is off today” — as if the model’s performance was the daily forecast.
What pulled me out of that mode was sitting down and actually reading Anthropic’s official best practices, front to back.
My honest first reaction:
“Oh. The answer was here the whole time.”
What I read
Only two sources:
- Best Practices for Claude Code — Anthropic official
- Prompting best practices for Claude 4 — Anthropic official
They’re in English, but they’re also short. Half a day alongside your normal coding and you can finish both. Despite that, few Japanese write-ups cover them comprehensively, which is part of what inspired this series.
What changed in my workflow
1. I always attach a way for Claude to verify itself
Anthropic says this is “the single highest-leverage thing you can do.” Always ship a test, a screenshot, or an expected output alongside your prompt. That one habit pushed my generated-code quality up a visible notch.
2. I stopped being afraid of /clear
The official rule: if you’ve had to correct Claude more than twice on the same issue in a session, run /clear. Performance degrades as context fills. The myth that “longer sessions make Claude smarter” is just that — a myth.
3. I stopped going straight to code
Explore → Plan → Code → Commit. Skipping Plan produces correct solutions to the wrong problem. For anything non-trivial I now always enter Plan Mode first.
What the series will cover
Over the next week or so I’ll publish a series called “Claude Prompt Mastery: The Complete Guide” on this site.
| Part | Topic |
|---|---|
| 1 | Three core principles (Golden Rule / context management / verification) |
| 2 | The Explore → Plan → Code → Commit workflow |
| 3 | 11 development prompt patterns |
| 4 | Five techniques for writing clearer instructions (CLAUDE.md / XML / Few-shot etc.) |
| 5 | How to mix and match Claude Code-only features |
| 6 | 13 thinking-mode prompt patterns |
| 7 | 18 everyday-life prompt patterns |
| 8 | Series wrap-up |
Notation for trust level
To separate official statements from my own opinion, every post uses the following labels.
| Label | Meaning |
|---|---|
| 🟢 Official quote | Direct quote or translation from the docs |
| 🟡 Official content | Described in the docs (mechanism / feature / step) |
| 🔵 My interpretation | My reorganization or prioritization of official info |
A deeper reference lives on Qiita
The homepage version focuses on personal takeaways and readability. The Qiita version is a reference-grade writeup with full official quotations — if you want the deep dive, visit my Qiita profile.
If you’ve been using Claude Code “by feel” too, you might try opening just one of the two links above today. Half a day. That’s all it took to change how I work.
Related posts
- “The testing phase disappeared” — comparing AI-driven and traditional development across the full lifecycle — what AI-driven development actually looks like
- From “built but never used” to running three products at zero monthly cost — how Claude Code changed personal development for me