AI Best Practices

Don't just say "summarize this" — say "summarize this in 3 bullet points for a non-technical audience." The more constraints you give, the better the output.

Starting a prompt with "You are a senior software architect reviewing a PR" produces measurably better results than asking the same question cold. Give Claude a persona with relevant expertise.

Include an example of the output you want. One good example is worth a paragraph of instructions. "Format it like this: [example]" almost always outperforms describing the format in words.

If the first response misses, correct it in the same conversation rather than starting over. Claude uses conversation context — a quick "actually, make it more concise and skip the intro" is faster than re-prompting from scratch.

Running tests, formatting code, linting — these don't need AI. Use hooks or scripts for predictable tasks, and save Claude's context for reasoning and generation. This is the single highest-leverage cost reduction.

Opus for architecture decisions and complex reasoning; Sonnet for coding, execution, and repetitive generation. A common pattern: Opus plans → Sonnet executes → Opus reviews. Don't pay Opus prices for Sonnet work.

Each sub-agent starts with a fresh context — no accumulated history. For multi-step workflows, sub-agents are cleaner and cheaper than one long conversation that grows context rot over time.

When a Claude Code session gets long, use context compacting to summarize history. You keep the key knowledge without carrying every message forward.

Use /context to see what’s consuming your window, /compact to compress history mid-task, and /clear to wipe it entirely when switching tasks. Starting complex work with a full context is a recipe for degraded output and wasted tokens.

If you’re typing a prompt yourself in real time, use Claude.ai Team — it doesn’t hit the shared API quota. Reserve the API for programmatic use: automated pipelines and production features. Use Claude.ai Team’s Projects to store persistent context across conversations at no extra token cost. (From: Claude Usage Guidelines, Feb 2026)

Don't rely on Claude to self-police dangerous operations. Write hooks that deterministically block DROP TABLE commands, destructive file operations, or cloud infrastructure changes before they execute. Hooks run scripts, not AI — they're fast and reliable.

AI can hallucinate confidently. Treat AI-generated content like a smart intern's first draft — review before sending, especially for facts, code in production, and anything a client will read.

For important automated workflows, add a dedicated validator sub-agent that reviews the writer's output before it's finalized. Cheap insurance against confident errors.

For experimental tools touching UI or client workflows — as long as no protected data is involved — a brief email asking for client sign-off is usually all you need. Keeps the relationship transparent and protects both sides. (Hat tip: Paul Tidwell, AI Office Hours Mar 2026)

Put standing instructions, project context, coding standards, and team norms in Claude.md — Claude reads it every session. You write it once; Claude never needs to be re-briefed again.

If you've typed the same prompt more than three times, make it a Command (.claude/commands/). Standup summaries, PR descriptions, code review templates — all great candidates.

Researcher → Writer → Validator is a proven pattern. Each agent gets a narrow, well-defined role, which produces better results than asking one agent to do everything in a single pass.

For complex tasks, engage Plan Mode before execution. Claude reasons through the approach first, uses sub-agents for execution, and produces significantly more coherent results than diving straight in.

Skills go a step beyond Commands — they trigger automatically based on task context, no slash command needed. Build a Skill for any output type you produce repeatedly (client reports, presentations, code reviews) so the right domain instructions load every time.

For recurring processes that combine multiple variable inputs — a résumé + role requirements + tech stack for interview prep — a purpose-built agent produces far more targeted results than a generic prompt.

Don’t assume a slide deck is the right output. When you have rich source material, ask Claude to produce an interactive HTML dashboard instead. A self-contained, navigable artifact can serve clients better than a static deck, and costs no more effort to create.

Layer 1 is a plain-English YAML file anyone can write. Layer 2 adds step-by-step instructions and Python scripts. Layer 3 is a full reference library with automated API calls. Non-technical team members author the first layer; engineers take it from there. (Hat tip: Max Oberbrunner, LLMunch & Learn, Feb 2026)

Instead of generating code and hoping it’s right, define API contracts and executable tests upfront — then let Claude iterate until every contract passes. This approach hit 100% compliance on a competitive programming benchmark vs. 79% single-shot. Especially powerful for legacy modernization. (Hat tip: Max Oberbrunner, AI Office Hours Mar 2026)

When producing many variations of a component, write a Skill capturing the architecture and styling — then add a deterministic hook that verifies every output shares required structure and naming. Claude fills the gaps; the gate enforces consistency. (From: AI Office Hours Mar 2026)

AI hallucinations drop dramatically when Claude is working from actual documents rather than general knowledge. Connect Claude to Guru, internal docs, and client content via RAG before asking it to answer domain-specific questions.

MCP connectors let Claude interact with Asana, Slack, Google Calendar, Guru, and more — in a single conversation. Instead of manually copying context between tools, let Claude work directly in the systems you already use.

n8n or similar tools are great for the "pipe data from A to B" parts of a workflow. Save AI for the steps that require judgment — synthesis, generation, decision-making. Hybrid pipelines (script + AI + script) are often better than pure AI pipelines.

Pipe your transcript tool into an N8N workflow to auto-generate summaries, scoring, and action items within minutes of a session ending. Praxent’s recruiting team used this pattern and saw interview pass rates double (25% → 50%). (From: LLMunch & Learn, Mar 2026)

Instead of asking Claude to help with a recurring chore, ask it to generate the script that eliminates the chore entirely — then let the script run on a schedule without AI involvement. The goal: delete the task from your calendar, not just do it faster. (From: AI Innovation Challenge, Feb 2026)

Feed Stitch your design system (colors, fonts, spacing) and describe the screen you need — it generates responsive layouts across mobile, tablet, and desktop. Output code feeds directly into Claude for further development and exports to Figma. Great for one-off prototypes without committing to a new workflow. (From: AI Office Hours Mar 2026)

Meeting summaries, status updates, email drafts — pick the thing you do most often and try AI on it first. Quick wins build confidence for more complex use cases.

The best AI workflow isn't "AI writes it, I send it." It's "AI drafts it, I improve and approve it." You bring judgment, context, and client knowledge that AI doesn't have.

If you're unsure whether to trust an output, ask "why did you say that?" or "what assumptions did you make?" Claude will surface its reasoning, which helps you spot where to verify independently.