Sonith ManakkatGetting good with Claude is less about secret phrases and more about habits. The advice here follows Anthropic's own prompting guidance, translated into plain language, with my own illustrative examples. The before-and-after prompts are mine, not quotes. Every habit has a quick picture, then an Expert view with the details.
Habit 1: Be clear and direct
Anthropic's guidance suggests thinking of Claude as a brilliant but new employee who does not know your norms or workflow. The more precisely you explain what you want, the better the result.
Before
- "Write something about our new product."
- Claude has to guess length, audience and goal
After
- "Write a 120-word announcement email for existing customers about our invoice-reminder feature. Friendly, no jargon, one call to action: try it from the Billing page."
- Claude knows exactly what good looks like
Expert viewThe colleague test
Anthropic's "golden rule" for clarity: show your prompt to a colleague with minimal context and ask them to follow it. If they would be confused, Claude will be too. It also advises giving instructions as numbered steps when order or completeness matters, and being explicit if you want "above and beyond" behaviour instead of expecting Claude to infer it.
Habit 2: Say who it is for, and why
Context improves answers. Anthropic's example: instead of "never use ellipses", explain that the reply will be read aloud by a text-to-speech engine that cannot pronounce them. Claude generalises from the reason.
- AudienceWho will read this?
- PurposeWhat should it achieve?
- FormatBullets, table, 100 words?
- ConstraintsWhat must it avoid?
Habit 3: Show examples
Examples are one of the most reliable ways to steer format, tone and structure.
Rewrite customer complaints into a calm one-line summary.
<example>
Input: "I've been waiting THREE WEEKS and nobody replies!!!"
Output: Customer has waited three weeks without a reply.
</example>
Now do this one: ...Expert viewWhat good examples look like
Anthropic recommends examples that are relevant (mirroring your real use), diverse (covering edge cases so Claude does not latch onto an accidental pattern) and clearly marked, wrapped in tags so Claude can tell example from instruction. It suggests three to five examples for best results, and says you can even ask Claude to judge your examples for relevance and diversity or to generate more from your first set.
Habit 4: Structure long prompts
When a prompt mixes instructions, documents and examples, labelling each part helps.
Expert viewTags and placement
Anthropic recommends consistent, descriptive tags such as <instructions>, <context> and <input>, nested when content has a natural hierarchy. For long inputs of about 20k tokens or more, its guidance says to put the longform data near the top, above your query, instructions and examples, and reports that queries at the end can improve response quality by up to 30 percent in its tests, especially for complex multi-document inputs.
Habit 5: Ask for a plan before a big job
For anything with several steps, ask first: "Before writing anything, list the steps you would take and the questions you have." Correcting a plan takes seconds. Correcting a finished piece of work takes much longer.
- 1Describe the jobState audience, purpose, format and constraints.
- 2Ask for a plan"List the steps and any questions first." Then adjust it: correcting a plan takes seconds.
- 3Claude draftsThe first answer is a draft.
- 4You check itFacts, numbers and code each need their own check.
- 5Give specific feedback"Shorter. Keep the second paragraph." Then go round again.
- 6Save what workedKeep the prompt in a note, or a standing rule in your instructions.
Takeaway. Short rounds with a check in each beat one giant prompt.
Habit 6: Work in rounds
Treat the first answer as a draft. Reply with specific feedback: "Shorter. Keep the second paragraph. Make the tone less formal." Short rounds beat one giant prompt.
Habit 7: Ask Claude to check itself, then check it yourself
Ask: "What assumptions did you make? What might be wrong here?" It often finds its own gaps. The final check is yours.
- FactsAsk for sources, then open them
- NumbersRecalculate a few by hand
- CodeRun the tests
Expert viewGrounding and investigation
Anthropic's coding guidance includes a sample instruction that tells Claude to investigate before answering: read the relevant files first and never make claims about code it has not opened. The same principle works outside code: give Claude the source material, ask it to quote the passages it relied on, and make "not stated in the material" an allowed answer.
Habit 8: Give it the right material, not all of it
Because quality can drop as the context grows (part 6), give Claude the relevant pages rather than a whole folder. If a conversation has become long and tangled, ask for a summary of decisions so far and start a fresh chat with that summary.
Habit 9: Use Projects for repeated work
Anthropic's support pages describe Projects as self-contained workspaces with their own chat histories and a knowledge base: files you add are reused across every chat in the project, and project instructions apply to every chat. Anthropic's pricing page lists Projects among Pro features. Chats inside a project do not share context with each other unless the information is in the project's knowledge. Part 9 covers sharing with a team.
Habit 10: Write standing instructions for coding work
In Claude Code, a file called CLAUDE.md holds instructions that Claude reads at the start of every session. Part 7 explains where it lives and what to put in it.
# CLAUDE.md (example)
- Use TypeScript. Keep functions small.
- Run the tests before saying a task is done.
- Never edit files under /generated.- 1Would a new colleague with no context be confused by it?Yes Add the missing background, then ask again.No keep going
- 2Is the audience, purpose or format missing?Yes Say who it is for, what it should achieve and how long it should be.No keep going
- 3Is a long document placed after your question?Yes Move long material to the top and put your question last.No keep going
- 4Is it a big multi-step job in one go?Yes Ask for a plan first.No keep going
- 5OtherwiseSend it, then check the answer and reply with specific feedback.
Takeaway. Most weak answers come from a missing detail you can add in a minute.
Four common mistakes
Longer prompts are always better
Length is not the goal. Clarity is. Remove material that does not help.
You need secret prompt tricks
Clear writing, good examples and careful checking cover most of it.
- Vague requests. "Make it better" gives Claude nothing to aim at. Say what better means.
- Assuming it knows your situation. It only has the page.
- Accepting the first answer. One round of feedback usually makes a visible difference.
- Skipping the check. A fluent answer is not a correct answer.
A weekly routine that works
- DraftPick one real task and do the first draft with Claude
- SaveKeep your best prompts in a note and improve them
- LearnAdd a line to your standing instructions for each repeated mistake
Mastery is a loop: ask clearly, look carefully, correct, and save what worked.
Frequently asked questions
Do I need to learn prompt engineering?
Not as a special skill. Clear writing, good examples and careful checking cover most of it.
Should I be polite to Claude?
Tone does not hurt, but clarity matters far more than politeness.
Longer prompts or shorter?
As long as needed to be clear, and no longer.