If you're using Claude Code the same way you use ChatGPT — typing conversational prompts and hoping for the best — you're leaving 80% of its power on the table.
I've been deep in Claude Code for 90+ days building a production AWS learning platform. What I've learned: the difference between average and power-user results comes down to knowing what's under the hood.
The March 2026 source code leak gave us that knowledge. Here's the complete, tested guide.
What Got Leaked (Quick Summary)
On March 31, 2026, Anthropic published Claude Code version 2.1.88 to npm with a .map source file included by accident. 512,000 lines of TypeScript. 1,900 files. 44 feature flags. The complete system prompt. Everything.
The key takeaways for power users:
- Claude Code has 40+ built-in tools — most people use 3-4
- The system prompt explicitly tells Claude to be direct — conversational prompts work against this
- Parallel tool execution is built-in but only triggers with properly structured prompts
- CLAUDE.md is a first-class feature that loads automatically every session
- 20+ unreleased features are built and tested but flagged off in public builds
Let's turn each of these into actionable techniques.
1. The Prompt Frameworks
/ghost — Kill the AI Voice
Use /ghost when Claude Code's output will be read by humans (docs, READMEs, PR descriptions, emails).
# Without /ghost
/ghost rewrite this README to sound like a senior engineer wrote it, not an AI
The output drops the hedging ("It's worth noting that..."), the filler ("In order to..."), and the corporate tone. You get clean, direct prose.
When to use: Documentation, blog posts, commit messages, Slack messages, code comments.
OODA — Force a Decision
OODA (Observe-Orient-Decide-Act) is a military decision framework. In Claude Code, it prevents the "here are the pros and cons" non-answer.
OODA: Our API currently uses REST. We're adding real-time features.
Should we add WebSockets, use SSE, or switch to GraphQL subscriptions?
We have 3 engineers and ship weekly.
Claude will observe your constraints, orient around your team size and shipping cadence, decide on one approach, and act by providing the implementation plan.
No more "it depends." You get a decision with reasoning.
L99 — Expert Depth
L99: Design a DynamoDB single-table schema for a multi-tenant SaaS
with per-tenant billing, usage metering, and team permissions
Without L99, you get a tutorial answer. With L99, you get GSI overloading patterns, composite sort keys for time-series queries, TTL-based cleanup strategies, and the exact access patterns with their key conditions.
SCAFFOLD — Full Project Setup
SCAFFOLD: Production Next.js 15 app
- Prisma + PostgreSQL
- NextAuth with GitHub + Google
- Tailwind CSS
- Vitest + Playwright
- GitHub Actions CI/CD
- Vercel deployment
Claude generates the full directory structure, all config files, and boilerplate before touching business logic.
PERSONA — Domain Expert Mode
PERSONA: Senior AWS Solutions Architect with 10 years experience.
Review this CloudFormation template for production readiness.
This frames Claude's entire response through that expertise lens.
2. System Prompt Insights You Can Exploit
Be Terse
The system prompt tells Claude Code to be direct. Mirror that:
# ❌ This triggers verbose mode
"Hey Claude, could you take a look at this function and let me know
if there's anything that might be improved? Thanks!"
# ✅ This gets the system prompt working for you
"What's wrong with this function? Fix it."
Parallel Execution
Claude Code can run tools simultaneously, but only when you make independence explicit:
# ❌ Sequential — Claude plays it safe
"Check the frontend components and then run the API tests"
# ✅ Parallel — both fire at once
"Two independent tasks:
1. Audit src/components/ for React anti-patterns
2. Run and fix failing tests in src/api/__tests__/"
On large codebases, this cuts response time significantly.
The Bash Tool is King
The leaked tool definitions show bash is Claude Code's most capable tool. It can install packages, run tests, check git status, read files, and execute arbitrary commands.
Power users lean into this:
"Run the test suite. If anything fails, read the failing test file,
understand what it expects, then fix the source code. Repeat until green."
This creates an autonomous debugging loop.
3. CLAUDE.md — Your Project's AI Memory
The source confirms CLAUDE.md is loaded automatically. Here's a battle-tested template:
# Project: [Name]
## Stack
- [Framework] + [Database] + [Deploy target]
- Key deps: [list the non-obvious ones]
## Architecture Decisions
- [Decision 1]: [Why]
- [Decision 2]: [Why]
## Code Conventions
- [Pattern]: [Where to use]
- [Anti-pattern]: [What to avoid]
## Current Context
- Working on: [Current feature/sprint]
- Recent changes: [What changed this week]
- Known issues: [Active bugs]
## Testing
- Framework: [Vitest/Jest/Playwright]
- Run: [exact command]
- Coverage: [current %]
## Deploy
- Staging: [command]
- Production: [command + checklist]
This replaces 10 minutes of context-setting at the start of every session.
4. Unreleased Features (From the Leak)
These are built, tested, and hidden behind compile-time flags:
KAIROS — Background Agent Mode
Referenced 150+ times in the source. An autonomous daemon that runs memory consolidation while you're idle. It can plan tasks, execute them in the background, and surface results when you return. Think: "Fix all the lint warnings while I'm at lunch."
BUDDY — AI Coding Companion
A Tamagotchi-style pet that lives in your terminal. Species deterministically generated from your userId hash (salt: friend-2026-401). Cosmetic hats. Speech bubbles. Claude generates its name and "soul description" on first hatch.
Leaked internal notes suggest an April 1-7 teaser with full launch in May 2026. The /buddy command did activate on April 1 as predicted.
Voice Mode
Talk to Claude Code. Useful for pair programming. In the source but not shipped.
ULTRAPLAN
Extended planning mode with multi-step task decomposition.
Undercover Mode
Details sparse, but it's referenced in the feature flags.
5. My Daily Workflow
After 90 days, here's my exact process:
Morning:
claude --resume # pick up where I left off
/init # if new project, generate CLAUDE.md
Planning:
Shift+Tab → plan mode # architecture first
OODA for decisions # no "it depends"
L99 for deep dives # principal-engineer answers
Building:
Parallel task prompts # 2x speed on independent work
SCAFFOLD for new modules # structure before code
Autonomous bash loops # "fix until green"
Shipping:
/ghost for PR descriptions
/ghost for changelog entries
Review with PERSONA: "Senior reviewer focused on security"
Hands-On Practice
If you want to practice these techniques with real AWS infrastructure, I've been building Cloud Edventures — guided missions in sandbox environments with automated validation. The AI Navigator path specifically covers Claude + Bedrock integration where you'll use a lot of these patterns.
The /ghost and OODA frameworks work in any AI tool though — start there.
What's Your Power-User Trick?
I'm collecting Claude Code techniques for a follow-up post. If you've found something that changed your workflow, drop it in the comments.
The gap between "using Claude Code" and "being a Claude Code power user" is wider than most people think. These frameworks close it.
Top comments (0)