Playbook · Prompting for individuals

Six patterns I use weekly.

Most "AI doesn't work for me" is a prompting problem, not a model problem. Plain English, in the right shape, with the right context — that's almost all of it. Six patterns below, each one earned in real work. Templates you can copy, examples that show the difference, and the anti-pattern that kills each one.

// Who it's for

Anyone using Claude in real work

// Time investment

20 minutes to read · weeks to master

// You'll need

One real task from your work

// You'll walk away with

Six reusable templates

01 —

// The working example

One task. Six ways through.

Every pattern below is shown against the same task: summarising a long document — the kind of 40-page board paper, strategy doc, or research report you don't have time to read in full but can't pretend you didn't get. Same input, six different prompts, six different shapes of output. Pick the pattern that matches what you need.

// PATTERN 01

The brief.

Default workhorse. Role + task + constraint + format.

// PATTERN 02

The exemplar.

Show, don't tell. Two examples of "good", let Claude pattern-match.

// PATTERN 03

The critic.

Draft, critique, revise. Three prompts, one cleaner output.

// PATTERN 04

The structured output.

A defined template. For anything you'll process further.

// PATTERN 05

The system prompt.

Set the rules once. Stop repeating yourself every chat.

// PATTERN 06

The chain.

Hard task → smaller prompts in sequence. Each one Claude can nail.

The default. Four parts in order: role, task, constraint, format. If you remember nothing else from this playbook, remember this shape. It works for 80% of what you'll ever ask Claude to do.

// The template

Copy thisPattern 01

You are {role}. Read the document below. {task}. Constraints: - {constraint 1 — what to focus on} - {constraint 2 — what to ignore} - {constraint 3 — length, tone, etc.} Format: {format — bullets, prose, table, etc.} Document: {paste document here}

// Before and after

// Before — vague

Summarise this document for me. [40-page PDF attached]

// After — briefed

You are a senior advisor preparing a board member for a meeting. Read the document below. Summarise it in a way the board member can absorb in two minutes. Constraints: - Lead with the decision being asked for, if any - Ignore detailed methodology — focus on findings and implications - 500 words maximum Format: three short sections — Decision, Findings, Risks Document: [40-page PDF attached]

// The anti-pattern

Treating Claude like a search engine. "Summarise this" is a search-engine prompt — Claude will guess at role, audience, length, and shape. The output will be generic because the brief was generic. Five extra lines of context is the cheapest quality lever you have.

Why this works

Claude is trained to be helpful, but helpful-without-direction defaults to safe-and-generic. Role tells it who's asking. Task tells it what success looks like. Constraints tell it what to leave out. Format tells it how to land. Each one cuts the space of possible outputs by an order of magnitude.

Some things are hard to describe but easy to recognise. Voice. House style. The way your team writes summaries. Instead of explaining the rules, paste two examples of "good" and let Claude pattern-match. Often more powerful than any number of constraints.

// The template

Copy thisPattern 02

I need to {task}. Here are two examples of the kind of output I want: --- EXAMPLE 1 --- {paste a good example} --- EXAMPLE 2 --- {paste another good example, different topic, same shape} Now do the same for this: {the actual input}

// Before and after

// Before — describing the style

Summarise this strategy doc. Use our house style — crisp, declarative, short sentences, no buzzwords, lead with the headline. [strategy doc attached]

// After — showing the style

I need to summarise a strategy doc in our house style. Here are two examples of how we do it: --- EXAMPLE 1 --- "Customer acquisition is the bottleneck. Three things would unblock it: a faster onboarding flow, clearer pricing, and partnerships with two named distributors. Cost: ~$400k. Timeline: Q3." --- EXAMPLE 2 --- "Margins are eroding faster than volume can compensate. The fix isn't pricing — it's mix. Shift 20% of the portfolio toward the higher-margin SKUs over six months. Risk: cannibalisation of the volume products." Now do the same for this: [strategy doc attached]

// The anti-pattern

One example is a coincidence; two is a pattern. A single example tells Claude "do exactly this." Two examples — same shape, different content — tell Claude "extract the pattern between these and apply it." Skip the second example and you'll get a near-clone of the first.

First drafts are draft-shaped. Three prompts in sequence beats one prompt every time: ask Claude to draft, then critique its own draft against named criteria, then revise. The middle step is where the gain comes from — Claude is a much better editor than its first-pass writer.

// The template (three prompts in one chat)

Prompt 1 — DraftPattern 03

Summarise the document below for {audience}. {length}. Lead with the headline. {document}

Prompt 2 — CritiquePattern 03

Now critique your own draft against these criteria. Be specific and direct — don't be polite: 1. Is the headline actually the most important thing, or did you bury the lede? 2. Is there anything in the summary that isn't supported by the source? 3. What did you leave out that the audience would actually want? 4. Where is the language vague or hedged when it could be direct? List the issues. Don't fix them yet.

Prompt 3 — RevisePattern 03

Now rewrite the summary, addressing every issue you raised.

// The anti-pattern

Asking for the critique and revision in one prompt. Bundling "critique your draft and rewrite it" into a single message gets you a polite critique and a barely-changed revision. The separation is the point — when the critique is its own turn, Claude takes it more seriously. When you then ask for the rewrite, it has to honour what it just said.

Why this works

This is the closest equivalent to how good human work happens: draft, step back, sharpen. Claude has the same blind spot humans do — what feels finished often isn't. Forcing a critique step before the rewrite produces a different output, every time.

When the output is going somewhere else — a spreadsheet, a tracker, a template — define the shape up front. Either name the fields explicitly, or paste a blank template. Claude is far better at filling a known shape than at inventing one.

// The template

Copy thisPattern 04

Read the document below and extract the following fields. If a field isn't present, write "Not stated" — do not guess. Fields: - {Field 1 name}: {what goes here} - {Field 2 name}: {what goes here} - {Field 3 name}: {what goes here} - {Field 4 name}: {what goes here} Format as a table with two columns: Field, Value. Document: {paste document here}

// Before and after

// Before — open-ended

Summarise the key things from this board paper. [board paper attached]

// After — structured

Read the board paper below and extract these fields. If a field isn't present, write "Not stated" — do not guess. Fields: - Decision sought: what is the board being asked to approve? - Recommendation: what does management recommend? - Top 3 risks: with mitigation noted - Cost: dollar figure if stated - Timeline: key dates - What changes if approved: in one sentence Format as a table with two columns: Field, Value. Document: [board paper attached]

// The anti-pattern

Forgetting to say what happens when a field is missing. Without the "if not present, write Not stated — do not guess" rule, Claude will fill blanks with plausible-sounding inferences. The structured shape gives you a false sense of accuracy because everything looks neat. The "don't guess" rule is what makes the table trustworthy.

If you're typing the same role, constraints, or style rules at the top of every chat, lift them into a system prompt — the persistent layer Claude reads before every message. In Claude.ai, this lives in a Project's instructions or a custom Style. Set it once, get the benefit every chat.

// The template (paste into a Project or Style)

Copy thisPattern 05

You are helping {your name}, {your role} at {your context}. When summarising documents: - Lead with the decision, recommendation, or headline finding - Use plain English — no "leverage", "synergy", "ecosystem" - Australian English spelling - Be concrete. Numbers beat adjectives. - If something is uncertain, say so. Don't smooth over gaps. When drafting writing: - Short sentences. Fragments where they land harder. - Active voice unless passive is genuinely better. - One idea per paragraph. What to avoid: - Generic openers like "I'd be happy to help" - Hedging language ("might consider perhaps possibly") - Closing summaries that repeat what you just said

// What belongs in a system prompt vs. a message

// In the system prompt

Things that don't change between tasks: • Who you are • Your writing style preferences • Spelling and grammar conventions • What to always include • What to always avoid • How to handle uncertainty

// In the message

Things specific to this task: • What you want done right now • The document or input • Length and format for this output • Audience for this output • Anything one-off about this case

// The anti-pattern

Stuffing the system prompt with everything. If the system prompt has more than ~300 words, Claude starts skimming. Keep it to genuine standing rules — the things you'd write at the top of every message anyway. Task-specific detail belongs in the message, not the system prompt.

When one prompt isn't enough, chain three or four smaller prompts in sequence. Each one Claude can handle cleanly. The output of one becomes the input of the next. This is how to get a real piece of work done, not a draft.

// The chain — four prompts for a 40-page document

Step 1 — MapPattern 06

Read the document below. Without summarising it, list the sections it contains and what each one is about in one line. Output as a numbered list. {document}

Step 2 — IdentifyPattern 06

From the section map above, which 3–4 sections matter most for {audience and decision}? Explain why each one matters in one sentence.

Step 3 — ExtractPattern 06

For each of those 3–4 sections, pull the key facts, numbers, and claims. Don't summarise yet — just extract. Use bullets, source the page if cited.

Step 4 — ComposePattern 06

Now write the final summary for {audience}. {length}. Use only the facts you extracted in step 3. Lead with the decision being asked for, if there is one.

// The anti-pattern

Chaining when a single prompt would do. The chain is for genuinely hard tasks — long documents, multi-step reasoning, anything where a one-shot output is unreliable. For everyday work, the brief (Pattern 01) is faster and just as good. Reach for the chain when you've tried Pattern 01 and the output is too shallow.

Insight from the field

The chain is what separates "I asked Claude and it gave me something OK" from "I worked with Claude and it gave me something I'd put my name on". Same model, same input — different result, because the second person treated it like a working session, not a vending machine.

Plain English, in the right shape, with the right context. Almost all of it.

02 —

// Common pitfalls

Four habits that kill the output.

Even with the right pattern, prompts go wrong in predictable ways. These four are the ones I see most often — in my own prompts and in everyone else's. Worth pinning to the monitor.

// Pitfall 01

The vibe brief.

"Make it punchier." "Sound more strategic." "Less corporate." These are vibes, not instructions. Claude can't see the vibe in your head, only the words on the page.

The fix Translate the vibe into rules. "Punchier" → "shorter sentences, lead with verbs, cut hedges". "Strategic" → "name the trade-off being made". "Less corporate" → "no 'leverage', 'synergy', 'ecosystem'".

// Pitfall 02

The polite ask.

"Could you possibly help me by perhaps maybe summarising this if you don't mind?" Claude mirrors register. Vague, polite prompts produce vague, polite outputs.

The fix Be direct. Imperatives are fine — "Summarise this in 200 words" is better than "Could you summarise this for me?". Politeness isn't rudeness in reverse; it's clarity.

// Pitfall 03

No negative instructions.

You said what you wanted but not what to avoid. Claude defaults to including everything that might be relevant — which means filler, hedges, and restatements you didn't ask for.

The fix Tell Claude what to leave out. "Don't include a closing summary." "Don't add disclaimers." "Skip the methodology." Negative instructions are often higher leverage than positive ones.

// Pitfall 04

The blind retry.

First output was wrong, so you typed "no, do it better" and got something equally wrong. You haven't told Claude what changed.

The fix Name what was wrong. "Too long — cut to 200 words." "Lead was buried — open with the decision." "Tone was too cautious — be more direct." Specific feedback beats general displeasure every time.

03 —

// Your starter toolkit

Six patterns. Pick one for the task in front of you.

A quick guide to which pattern fits which kind of work. When in doubt, start with Pattern 01 — the brief covers most ground. Reach for the others when the output isn't where it needs to be.

The brief

Reach for: most things. The default workhorse — role, task, constraint, format. Solves 80% of prompts. If you're not sure where to start, start here.

The exemplar

Reach for: house style, voice, taste. When the rules are hard to write but the examples are easy to find. Paste two, let Claude pattern-match.

The critic

Reach for: quality matters. Three prompts: draft, critique, revise. The middle step is the magic. Worth the extra two turns when the output has to be good.

The structured output

Reach for: it goes somewhere else. When the output feeds a spreadsheet, a tracker, or a template. Define the fields. Always say "don't guess".

The system prompt

Reach for: you keep repeating yourself. Lift standing rules into a Project's instructions or a Style. Stop typing the same setup in every chat.

The chain

Reach for: the hard tasks. Long documents, multi-step reasoning, anything where Pattern 01 gave you something too shallow. Break it into steps. Each one Claude can nail.

Using this in practice?

The patterns above aren't a theory of prompting. They're the six shapes I reach for in real work, week in, week out. Three of them — the brief, the structured output, the system prompt — cover almost everything. The other three are there for when the first three aren't enough.

Plain English is enough. Almost everything that makes a prompt better is also what makes a brief to a junior colleague better — clear role, named constraints, defined shape, examples where you have them. If you can write a good email, you can write a good prompt.

This playbook is the foundation for the rest of the series. Playbook 02 (Memory and context) picks up where Pattern 05 ends — what Claude knows about you and your work across conversations. Playbook 03 (Skills) is what happens when you've found a prompt worth keeping.

Six patterns I use weekly.

One task. Six ways through.

The brief.

The exemplar.

The critic.

The structured output.

The system prompt.

The chain.

Four habits that kill the output.

The vibe brief.

The polite ask.

No negative instructions.

The blind retry.

Six patterns. Pick one for the task in front of you.

Using this in practice?