AI Guides › Workbench
By Nigel Guy · 7 min read
When the output keeps coming back slightly wrong, the usual fix is to write more instructions. Another paragraph on tone, a list of banned words, a capitalised NEVER. It feels like progress because the prompt keeps getting longer, but you are describing the shape you want in words when you could simply show it. Three short, well-chosen examples will usually pin down format, tone and length more firmly than three paragraphs of adjectives.
The rule: show three examples, each wrapped in tags and each different from the others, and keep instructions for the things an example can't show.
Anthropic's current prompting guide calls examples "one of the most reliable ways" to steer Claude's format, tone and structure, and recommends three to five of them. The reason is simple. "Warm but brief" means something different to everyone who reads it. An example of a warm, brief reply means one thing. Format, length, register, how much detail goes into each part: these are hard to put into words and easy to copy from a sample.
Instructions still matter: the same guide says explaining why you want something helps Claude generalise. So split the work. Instructions carry the goal, the reasons and the hard limits; examples carry the shape.
| Item | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| 1. Tagged examples prompt | Puts three input-and-output pairs into a single chat prompt | Free on any Claude plan | One-off jobs and testing an idea | You paste it again every time |
| 2. Example auditor | Gets Claude to check your examples for overlap and copyable patterns | Free | Before you reuse a set | Its verdict is a second opinion, not a test |
| 3. Example filler | Writes extra examples to cover gaps in your set | Free | Moving from three examples to five | Synthetic examples can drift from your real voice |
| 4. Project instructions | Keeps your examples on hand for every chat in a Claude Project | Free plan allows up to five projects; paid plans have more | Tasks you repeat each week | Applies only to chats inside that project |
| 5. Skill with an examples file | Packages the examples so Claude loads them when the task comes up | Depends on your plan and settings (see below) | Work you, your team or Claude Code repeat often | More to set up and maintain |
On cost: Anthropic's pricing page lists the paid plans in US dollars (Pro at $20 a month, roughly £15 at time of writing). We couldn't confirm the exact pound price, so check what you're shown at checkout.
Anthropic's guide says to wrap each example in <example> tags and the full set in <examples> tags, so Claude can tell the examples apart from your instructions. Pick three real cases: one typical, one awkward and one edge case.
You are drafting [TYPE_OF_OUTPUT] for [AUDIENCE]. The aim is [GOAL], and a good result reads as [WHAT_GOOD_LOOKS_LIKE].
Hard limits: [CONSTRAINTS, e.g. under 80 words, British spelling, no discount offers].
Why these limits matter: [ONE_LINE_REASON].
Match the shape of the examples below: length, structure and register. Do not reuse their wording or their specific details.
<examples>
<example>
<input>[TYPICAL_CASE_INPUT]</input>
<output>[TYPICAL_CASE_OUTPUT]</output>
</example>
<example>
<input>[AWKWARD_CASE_INPUT]</input>
<output>[AWKWARD_CASE_OUTPUT]</output>
</example>
<example>
<input>[EDGE_CASE_INPUT]</input>
<output>[EDGE_CASE_OUTPUT]</output>
</example>
</examples>
Now do the same for this:
<input>[NEW_INPUT]</input>
If the new input is missing something you need, ask me before writing. Do not guess.
Before you answer, check: is it within the limits, does it match the examples in shape but not in wording, and does it rely on any fact I didn't give you?
Fill in: the output type, audience, goal, limits, three real input-and-output pairs, and the new case.
The guide suggests asking Claude to check your examples for relevance and variety. That matters because the main risk with examples is that Claude copies the wrong thing from them.
You are reviewing a set of examples I plan to use in a prompt. The task they are for: [TASK_DESCRIPTION].
<examples>
[PASTE_YOUR_TAGGED_EXAMPLES]
</examples>
Work through these in order:
1. List any features all the examples share that I may not have meant to teach, e.g. the same opening word, the same length, the same sign-off or the same product.
2. Say which real situations for this task none of the examples covers.
3. Point out any example that contradicts my instructions: [PASTE_INSTRUCTIONS].
4. Suggest one change to each example that would make the set more varied without changing the format.
Present this as a table with columns: issue, which example, suggested fix.
If you don't know what situations this task meets in practice, ask me rather than inventing them.
Before you finish, check that every issue you list points to specific text in the examples.
Fill in: the task, your examples and your current instructions.
The guide also says Claude can write more examples from your starting set. Use this to fill the gaps the auditor found. Don't use it to replace real examples.
Here are three examples of [TASK] written to my standard:
<examples>
[PASTE_YOUR_THREE_EXAMPLES]
</examples>
Write [NUMBER, 1 or 2] more examples covering these situations, which my set leaves out: [GAPS_FROM_AUDIT].
Keep the format, length range and register of my examples. Make the details clearly made up: use placeholder names and no real companies, people or figures.
Give each one in the same <example> structure.
If a gap is too vague to write an example for, ask me to narrow it down first.
Before you answer, check that no new example repeats an opening, closing or phrase from mine.
Fill in: the task, your three examples, and the gaps you want covered.
For examples you use every week, the help centre's route is: go to claude.ai/projects, click + New Project, then Set project instructions, paste in item 1 without the new input, and click Save instructions. They then apply to every chat in that project. Free accounts get up to five projects. Keep the examples in the instructions rather than an uploaded knowledge file, so they are always in front of Claude.
Anthropic's skill-writing guide describes an "examples pattern": input-and-output pairs inside a skill, used as in an ordinary prompt. A skill is a folder with a SKILL.md file whose frontmatter needs a name (lowercase letters, numbers and hyphens, 64 characters at most) and a description saying what it does and when to use it. Keep SKILL.md under 500 lines and move longer material into a linked examples.md, which Claude reads only when needed. Which plans and surfaces run custom skills changes, so check Anthropic's skills documentation before you build one.
Claude copies everything your examples share, including what you didn't mean. If all three open with "Thanks for getting in touch", every reply will. If all three run to about 60 words, a case that needs 150 gets cut short. Run the auditor (item 2) and vary on purpose: lengths, openings, subjects.
The second trap is an example that breaks your own instructions. If the limit is 80 words and an example runs to 110, you can't be sure which Claude will follow. Fix the example.