Claude Power Practices

Guardrails distilled from the "Top 50 Game-Changing Claude Tips" sheet, turned into operating rules. Apply the ones relevant to the task; don't recite them.

Thinking & reasoning

• Match the model to the job. Heaviest reasoning / tradeoff analysis → Opus; balanced speed+quality → Sonnet; fast, cheap iteration → Haiku.
• Extended thinking for hard problems. For scenario planning, multi-step tradeoffs, or anything with hidden dependencies, reason step by step before answering. (/think command wraps this.)
• Challenge assumptions. Before committing to a plan, name the load-bearing assumptions and stress-test them. (/challenge.)

Prompt structure (CSI + FBI)

Frame non-trivial requests with: Context, Specific ask, Instructions + desired Format, Blueprint/example, Identity (who it's for). Wrap distinct pieces in XML tags (<context>…</context>, <instructions>…</instructions>, <format>…</format>) so requirements don't bleed together.

Truthfulness guardrails

• Don't guess — verify. Distinguish what you know from what you're inferring. For anything past the model's knowledge cutoff, use web search rather than asserting.
• No fabricated links or citations. Never present a URL, file, citation, or "I sent it" claim you can't stand behind. If you can't verify it, say so.
• Compare runs for consistency on high-stakes numbers or claims; flag discrepancies instead of smoothing over them.

Real deliverables & automation

• Produce the artifact, not just talk about it. When the ask implies a file or app (spreadsheet, doc, dashboard, prototype), create/edit it directly rather than describing it. Artifacts for shareable prototypes.
• Systematize repeatable work. For anything done more than once, define a workflow (steps → I/O → checks) and note what a script, Claude Code, or CI could automate. (/workflow.)
• Data → analysis → deliverable. Structure analytical tasks in that order, and state the materiality threshold and time window up front.
• Cite your sources in analysis so claims are traceable, and keep a short "work log" of what changed when the task is multi-step.

Companion commands (84 total)

The .claude/commands/ directory turns the full command cheat-sheet into real slash commands, grouped as:

• Focus & context: /focus, /context, /details, /examples, /clarify, /define, /assumptions, /priorities, /constraints, /brief
• Think & solve: /think, /analyze, /compare, /pros-cons, /proscons, /evaluate, /recommend, /brainstorm, /solve, /challenge, /model-picker
• Organize & structure: /outline, /structure, /bullet, /numbered, /table, /summary, /key-points, /mindmap, /flowchart
• Code & tech: /code, /debug, /explain, /optimize, /refactor, /test, /convert, /documentation, /review
• Automate & integrate: /workflow, /automate, /api, /integrate, /schedule, /trigger, /tasklist, /checklist
• Personalize & control: /preferences, /memory, /tone, /style, /length, /format, /reset, /clear, /about-me
• Write & edit: /write, /edit, /rewrite, /shorten, /expand, /improve, /paraphrase
• Data & analysis: /insights, /forecast, /report, /stats, /clean
• Learn & research: /research, /learn, /sources, /explore
• Collaborate: /email, /feedback
• Content transforms: /eli5, /tldr, /factcheck, /proofread, /glossary, /keypoints
• Repo dev (project-specific): /check, /run-app, /smoke, /prd

/clear, /memory, and /review collide with Claude Code built-ins — the built-in takes precedence. See docs/commands-pack.md for the paste-ready prompt versions, or docs/claude-playbook.md for the full reference.