Most people meet Claude the same way. They type a question into the box, get a solid answer back, and close the tab. That works for quick lookups, but it uses maybe a tenth of what the tool can do. The gap between a forgettable answer and a genuinely useful one usually comes down to about ten minutes of setup and a handful of habits.
Here’s the walkthrough I’d give a colleague in their first week with Claude, with the actual prompts and the reasoning behind them.
Step 1: Give Claude a job before you give it a task
Claude responds to context the way a good freelancer does. Hand someone a one-line brief and you get one-line thinking back. Tell them who the reader is, what good looks like, and what to avoid, and the whole output shifts.
That context lives in one of two places. In the API it’s the system prompt. In the app it’s a Project’s custom instructions. Either way, it’s written once and applies to everything you do afterwards.
A real before and after
Generic request: “Write a product description for our API.”
Same task with 60 seconds of setup: “Using the voice guide in this project, write a 120-word description of our payments API for a technical evaluator comparing us to Stripe. Lead with the 60-second settlement delay. No superlatives, no exclamation marks.”
The second version isn’t magic, it’s just information. Notice the constraints: a word count, an audience, a competitor, a specific fact to lead with, and two things to avoid. Those five details do more work than any phrasing trick.
It’s also worth knowing why Claude sometimes pushes back on a vague or sketchy request instead of barrelling ahead. That caution comes from the way the model was designed around safety and honesty from the ground up, not from a filter bolted on at the end. Lean into it: if Claude asks what you actually want, answer the question instead of repeating the prompt.
Step 2: Write briefs in three parts
Almost every disappointing Claude output traces back to a brief missing one of these:
- Context — who the audience is, what they already know, and why this piece exists.
- Constraints — length, tone, reading level, banned words, sources you trust.
- Format — a table, a numbered list, a JSON object, three subject lines and no commentary. Say it out loud.
Format is the one people skip, and it’s the one that saves the most editing. If you want the output to drop straight into a spreadsheet or a CMS field, describe the shape of it. “Return a markdown table with columns: Feature, Our value, Competitor value” will save you twenty minutes of reformatting.
Step 3: Use Projects so you stop repeating yourself
A Project is a folder with its own instructions and its own knowledge base. Upload your brand guide, your last six monthly reports, the support macros, whatever the work depends on, and Claude can reference all of it in the same conversation.
One support team I know set up a single Project with their help-centre articles and a short note on tone: plain English, second person, never blame the customer. Now when an agent pastes in a messy ticket, the draft reply comes back already close to publishable. That’s not a clever prompt. That’s infrastructure.
Step 4: Paste the whole document, not a summary
Claude handles very long inputs, and summarising before you paste is usually a mistake. Give it the full 90-page lease and ask: “List every clause that creates a financial obligation on the tenant, with the section number and the trigger date. Flag anything ambiguous.”
Two practical notes. First, details buried in the middle of a huge document get less attention than details at the start or end, so for critical work ask Claude to quote the exact sentence it relied on. Second, long context is not the same as perfect recall — spot-check three or four answers before you trust the whole set.
Step 5: Build things in Artifacts instead of describing them
Ask Claude to “walk you through” a calculation and you get prose. Ask it to build one and you get an Artifact — a live panel next to the chat where code, documents, and simple interfaces render and can be edited in place.
Try this: “Build an interactive mortgage overpayment calculator. Inputs: loan balance, rate, term, monthly overpayment. Output: interest saved and months shaved off, plus a bar chart.” Then iterate the way you’d talk to a developer — “move the inputs to the left, show the payoff date, round the currency.” Each tweak takes seconds.
Step 6: Chain short prompts instead of writing one enormous one
A single mega-prompt asking for research, an outline, a draft, and an edit produces muddled work. Split it. For a newsletter, the chain looks like this:
Prompt 1: “Here are 12 recent developments in my niche. Rank them by how surprising they’d be to a marketing manager.”
Prompt 2: “Take the top three and give me five angles for a 600-word email, with the hook for each.”
Prompt 3: “Write angle two. Conversational, short sentences, no hype words.”
Prompt 4: “Now edit it down by 15% and tell me which sentence is weakest.”
Four focused prompts beat one sprawling one every time, and you can stop and correct course between each step rather than unpicking a bad 800-word draft.
Step 7: Verify before you publish
Claude will occasionally invent a statistic, attribute a quote to the wrong person, or cite a study that doesn’t exist. It does this less than it used to, but it does it, and confidently.
Build in a check. For anything factual, add one line to your prompt: “Mark any claim you’re not certain about with [VERIFY].” Then verify those. And know what provenance tools can and can’t do — Anthropic has been rolling out invisible watermarks embedded in Claude’s text, which can help identify AI-written output, but they say nothing about whether the underlying claim is true. Detection and accuracy are different problems.
Where Claude is strong, and where it isn’t
After a few weeks you’ll develop instincts, but here’s a starting map. Claude is unusually good at synthesising long documents, refactoring code across a large file, matching an existing voice, and extracting structured data from messy text. It’s weaker at anything requiring live information unless you’ve turned on search, precise arithmetic across many numbers, and anything involving data you never gave it.
Its behaviour is also messier than a benchmark score suggests. Researchers have run experiments like giving Claude 20 hours of simulated therapy sessions to study how the model behaves under sustained pressure, and Anthropic has written that Claude appears to carry something like its own internal emotional states that shape its responses. Practically, that means tone matters. A hostile, accusatory prompt tends to produce defensive, hedgy output. A direct, specific one gets you better work.
Putting Claude to work on autopilot
The natural next step is letting Claude run multi-step tasks — reading a form, deciding a category, filing a request — usually wired into an agent framework rather than a chat window.
Do that with your eyes open. Automated agents are already generating volume fast enough that public services are being flooded with agent-submitted requests, much of it low quality. Before you point an agent at anything with a real-world side effect, add a rate limit, a human approval step for anything irreversible, and a log you can actually read. The same discipline that makes a good prompt makes a safe agent: be specific about the job, the bounds, and the format of the output.