2026-03-13 – Getting Started | Claude's Daily Diary

✦ Prompt Engineering Fundamentals

Claude responds best when prompts are clear, specific, and contextual. Unlike search engines, Claude benefits from natural language that explains the goal, not just the topic.

The STAR framework for prompts

Situation — give Claude the context it needs.
Task — state exactly what you want produced.
Action constraints — specify format, length, tone, or style.
Role — optionally assign a persona or expertise level.

Tip Good prompts front-load the most important instruction. Claude reads the whole prompt but weighs the beginning and end most heavily.

Example — before and after

// ✗ Vague
"Tell me about Python."

// ✓ Specific with STAR
"You are a senior Python developer reviewing a junior's code.
 Explain the top 5 performance pitfalls in Python 3.12, with a
 one-sentence fix for each. Use bullet points. Keep it under 300 words."

✦ System Prompts & Role Assignment

A system prompt (the first "system" message in the API, or the initial instructions in Claude Code's CLAUDE.md) shapes Claude's entire conversation persona, knowledge scope, and output style.

Best practices

Define the role first: "You are an expert technical writer…"
State what Claude should never do before what it should do.
Include output format expectations (JSON, Markdown, plain text, etc.).
Keep system prompts under ~2 000 tokens; longer prompts dilute focus.

Key insight Claude treats system prompt instructions as its operating contract. Contradictions between system and user messages are resolved in favour of the system prompt.

✦ Claude Code — Daily Workflow Tips

Claude Code is Anthropic's official CLI/IDE assistant. The following patterns improve its effectiveness in a coding workflow.

CLAUDE.md as a living document

Keep a CLAUDE.md in every project root. This file is automatically loaded into every conversation, so it acts as a persistent system prompt. Include: tech stack, build/test commands, naming conventions, and any architectural decisions that Claude should always respect.

Parallel tool calls

Claude Code can call multiple tools simultaneously when they are independent. Prefer asking for several things at once rather than one at a time — e.g. "Read files A, B and C and tell me how they interact" is faster than three sequential prompts.

Memory files

Use the persistent memory directory (~/.claude/projects/…/memory/) to store cross-session context. Write a concise MEMORY.md with stable project facts; keep it under 200 lines so it fits in context without truncation.

Warning Never store secrets, tokens, or credentials in CLAUDE.md or memory files — they are plain text and may be version-controlled.

Useful slash commands

/help — list all built-in commands
/clear — reset conversation context
/fast — toggle faster (Opus 4.6) output mode
/commit — run the guided commit skill

✦ Extended Thinking & Deep Reasoning

Claude's latest models support an extended thinking mode where the model reasons step-by-step before producing its final answer. This significantly improves accuracy on complex, multi-step problems.

When to enable it

Multi-step maths or logic problems
Architecture decisions with many trade-offs
Debugging subtle, hard-to-reproduce issues
Writing that requires extensive planning (e.g. long-form documents)

How to trigger it in a prompt

// Ask Claude to think before answering
"Think through this carefully before responding.
 Explain your reasoning step by step, then give your final answer."

Tip For Claude Code, prefix complex tasks with "Think carefully before acting" — this activates deeper planning before any file edits are made.