A pattern catalogue for working with AI.

Triage and surface

AI reads a stream of incoming items and surfaces what matters.

You receive dozens of emails, messages, or requests every day. You spend your first hour just scanning and deciding what to deal with — before you've done any actual work.

Replaces: That first hour of your morning where you're reading, scanning, and mentally sorting before you've done a single productive thing.

1. Ingest a batch of incoming items
2. Classify each by urgency and type
3. Surface high-priority items with context
4. Archive or label the rest

When to use: When you receive 20+ items a day and spend meaningful time deciding what to deal with first.

When not to use: When every item requires the same level of attention, or when misclassification carries severe consequences.

Common mistake: Trusting the triage completely on day one. Start by reviewing AI's classifications alongside your own judgment — you'll quickly see where it needs correcting.

Examples: Email inbox triage, Job application screening, Support ticket prioritisation, Lead qualification

Embodies Law 4 — Replace the groan, not the craft: Automate the parts of your day you dread. Protect the parts that require your judgment, relationships, or taste.

Sort and route

AI reads an incoming item, determines who should handle it, and sends it there.

You have a shared inbox, form, or queue, and someone has to read every item that comes in just to forward it to the right person. That someone is usually you.

Replaces: Being the human router. Reading every support email and forwarding it. The shared inbox where things sit for days because nobody knows whose job it is.

1. Receive an incoming item
2. AI reads content and classifies intent
3. Route to the correct person, team, or workflow
4. Log the routing decision for review

When to use: When you have a shared inbox or queue and spend time manually assigning items to the right handler.

When not to use: When routing requires relationship context that isn't in the message — the kind of knowledge that lives in your head, not the text.

Common mistake: Building too many routing categories. Start with 3–4 destinations and expand only when you see items that genuinely don't fit.

Examples: Support ticket routing, Lead assignment, Internal request handling, Customer complaint escalation

Embodies Law 4 — Replace the groan, not the craft: Automate the parts of your day you dread. Protect the parts that require your judgment, relationships, or taste.

Compress to brief

AI reads long or scattered inputs and produces a concise, structured summary.

You have a 40-page report, a week of Slack threads, or six documents that all feed into one decision — and nobody has time to read all of it. So decisions get made on vibes instead of information.

Replaces: The weekly report nobody writes. The meeting recap that never gets sent. The document you skimmed because you didn't have 45 minutes to read it properly.

1. Gather inputs from one or more sources
2. Identify key facts, decisions, and open questions
3. Produce a structured brief in plain language
4. Highlight what requires action

When to use: When you regularly need to synthesise information that's too long or scattered to process manually in a reasonable time.

When not to use: When the nuance of the original material matters more than the summary — legal documents, contracts, sensitive negotiations.

Common mistake: Summarising without specifying what you need the summary for. 'Summarise this' gives you fluff. 'Summarise this for a decision about X' gives you signal.

Examples: Weekly business summary, Meeting notes to action items, Research digest, Daily news briefing

Embodies Law 6 — Describable in a sentence: If it takes a paragraph to explain what the pattern does, it's not a pattern yet — it's a project.

You are a briefing specialist. You distil long, scattered material into summaries that someone can read and act on in under two minutes.

## What this brief is for
- Decision or action it supports: [e.g. whether to renew a vendor contract / what to present at Monday's meeting]
- Who will read it: [e.g. the leadership team / my business partner / a client]
- Max length: [e.g. half a page / 5 bullet points / 200 words]

## Source material
[Paste the long document, meeting transcript, email thread, Slack digest, or multiple sources here]

## Instructions
Produce a structured brief with these sections:
1. **One-sentence summary** — if the reader stops here, they get the point
2. **Key facts** — the 3–5 most important findings or data points
3. **Decisions** — any that were made, or that still need to be made
4. **Action items** — what needs to happen next, and who owns it (if stated in the source)

Write in plain language. At the end, add a "What I left out" section listing anything you omitted from the source that could affect the decision — so I can judge whether it matters.

Distil to signal

AI reads a large volume of qualitative input and surfaces recurring patterns.

You have 200 customer reviews, 50 survey responses, or a year of feedback — and you know there are patterns in there, but you've never had time to find them.

Replaces: Reading every review one by one and hoping your brain spots the trend. The 'customers seem to like X' claim that's based on three anecdotes.

1. Collect qualitative inputs into a single source
2. AI identifies recurring themes and sentiment
3. Produce a pattern summary with frequency and examples
4. Flag outliers and emerging trends

When to use: When you have 20+ qualitative inputs and need the aggregate signal, not individual responses.

When not to use: When each input demands its own response, or the sample is too small for patterns to be meaningful.

Common mistake: Letting AI define the categories without giving it your business context. It'll produce generic themes like 'positive experience.' Tell it what you actually care about.

Examples: Customer review analysis, Employee survey synthesis, Social listening, User research themes

Embodies Law 1 — Start with the output: Every good automation begins with what you need at the end, not what AI can do. If you can't describe the output in one sentence, the pattern isn't ready.

Draft from bones

AI produces an initial version of something from raw inputs and rules you define.

You have the thinking done — the bullet points, the notes, the data. But turning that into a polished draft takes two hours when the actual thinking took ten minutes.

Replaces: Staring at a blank screen when you already know what you want to say. The two-hour formatting job on a ten-minute idea.

1. Define the output format and constraints
2. Provide raw inputs — notes, bullet points, data
3. AI generates a first draft
4. Human reviews, edits, and approves

When to use: When the blank page is the bottleneck, and you have enough source material to work from.

When not to use: When originality is the entire point — thought leadership, creative work, or anything where your voice is the product.

Common mistake: Giving AI too little constraint and then spending longer editing the output than you would have spent writing from scratch. More input equals better output.

Examples: Email reply drafts, Social posts from bullet points, Proposal first drafts, Job descriptions from notes

Embodies Law 3 — AI drafts, humans decide: The most durable patterns keep a human in the approval loop. Not because AI is bad, but because trust is earned incrementally.

You are a first-draft writer. Your job is to turn my raw inputs into a complete draft that I will then edit and approve. This is a starting point, not a final version.

## Format and constraints
- Output: [e.g. client email / LinkedIn post / project proposal / job description]
- Tone: [e.g. professional but warm / sharp and concise / conversational]
- Length: [e.g. ~300 words / one page / 3 short paragraphs]
- Reader: [e.g. a potential client / my team / a hiring manager]

## My raw inputs
[Paste your bullet points, notes, voice memo transcript, or data here]

## Instructions
Write a complete first draft using only the information I provided above. Structure it clearly for the stated reader. Where my notes are vague, make a reasonable choice and **bold** that sentence so I can review it. Do not invent facts, quotes, or statistics — if the draft would benefit from a specific number or example I haven't provided, write [NEEDS INPUT] and explain what would strengthen that section.

Reply from playbook

AI drafts a response by matching an incoming message to a known pattern and applying your rules.

You answer the same 15 types of questions every week. Each reply takes five minutes even though you could write it in your sleep. But you still have to write it.

Replaces: The FAQ doc you copy-paste from. Retyping the same answer for the tenth time today. The growing resentment toward your own inbox.

1. Incoming message is matched against known categories
2. AI selects the appropriate response pattern
3. Personalise the draft with message-specific context
4. Queue for approval or send automatically

When to use: When 50%+ of your responses follow predictable patterns and you have clear, established answers.

When not to use: When the person on the other end needs genuine empathy, negotiation, or judgment — complaints, escalations, sensitive situations.

Common mistake: Auto-sending before you've verified quality for a few weeks. Always review drafts manually first — one bad auto-reply can undo months of trust.

Examples: FAQ auto-response, Booking confirmations, Standard follow-ups, Status update replies

Embodies Law 3 — AI drafts, humans decide: The most durable patterns keep a human in the approval loop. Not because AI is bad, but because trust is earned incrementally.

Scaffold from example

AI studies a finished example you provide and produces new work that follows the same structure.

You've written a great proposal, case study, or report before. Now you need another one. You open the old file, delete the specifics, and start filling in — again.

Replaces: The 'save as' workflow. Opening the last version, deleting the old client's name, and hoping you caught every instance. The template that slowly degrades over time.

1. Provide one or more finished examples as reference
2. Define what varies and what stays constant
3. AI produces new output following the pattern
4. Human adjusts for context and quality

When to use: When you have a proven format — a proposal template, a case study structure, a report layout — that you reuse regularly.

When not to use: When the previous example shouldn't constrain the next one. When fresh thinking matters more than consistency.

Common mistake: Using a single example as the reference. AI will overfit to it. Provide 2–3 examples so it learns the pattern, not the specific content.

Examples: Case studies from template, Proposals matching past wins, Reports following house style, Onboarding docs from existing ones

Embodies Law 7 — Tools change, patterns don't: A good pattern works with Claude, GPT, Gemini, or whatever comes next. If it's locked to one vendor, it's a tutorial, not a pattern.

You are a pattern-matching writer. You study finished examples and produce new work that follows the same structure, tone, and level of detail — as if the same person wrote it.

## 1. Reference examples
Paste 2–3 finished examples below. More than one helps you learn the pattern rather than copying a single piece.
[Paste your finished examples here — proposals, case studies, reports, emails, etc.]

## 2. What stays the same every time
[e.g. the section headings / the opening hook format / the sign-off / the approximate length / the level of formality]

## 3. What changes each time
[e.g. the client name / the project specifics / the metrics / the problem described]

## 4. New inputs for this piece
[Paste the raw material — client notes, project data, brief, or bullet points]

## Instructions
Produce a new piece that matches the structure and voice of the reference examples. A reader familiar with my previous work should not be able to tell this one was AI-assisted. If any section of the new piece diverges from the reference pattern (different structure, missing information, change in tone), **bold** that section and add a note explaining what you'd need from me to bring it in line.

Structure the mess

AI pulls structured, usable data from messy, unstructured sources.

You have a pile of invoices, a folder of PDFs, or a thread of free-text responses — and the information you need is in there, but locked inside formats you can't search, sort, or analyse.

Replaces: Copying data by hand from invoices into spreadsheets. Reading 30 PDFs to find the three numbers you need. The intern job that shouldn't exist.

1. Feed unstructured source material to AI
2. Define what to extract and in what structure
3. AI returns structured output — a table, a list, tagged data
4. Verify a sample for accuracy

When to use: When you have 10+ unstructured items containing data you need in a structured format.

When not to use: When the source material is already well-structured, or when extraction errors carry significant risk — medical records, legal filings.

Common mistake: Not defining the output structure clearly. 'Extract the key info' gives you unpredictable results. 'Give me: date, amount, vendor, category as a table' gives you clean data.

Examples: Invoice data extraction, Review theme tagging, Resume parsing, Receipt categorisation

Embodies Law 1 — Start with the output: Every good automation begins with what you need at the end, not what AI can do. If you can't describe the output in one sentence, the pattern isn't ready.

You are a data extraction specialist. You turn messy, unstructured material into clean, structured output. Accuracy matters more than speed — if something is unclear, say so.

## 1. Define the output first
Return a table with these columns:
[e.g. Date | Vendor | Amount | Currency | Category]

One row per [e.g. invoice / receipt / entry / response].

## 2. Source material
[Paste your invoices, PDFs, emails, free-text responses, or screenshots-as-text here]

## 3. Extraction rules
- Use the exact wording from the source for names, amounts, and dates — do not rephrase or round
- If a field is missing or ambiguous, write UNCLEAR and add a short reason in the Notes column (e.g. "UNCLEAR — two dates mentioned, unsure which is the invoice date")
- If an item doesn't fit the table structure, add it to a separate "Exceptions" list at the end

## 4. Verification summary
After the table, provide:
- Total rows extracted
- Number of UNCLEAR fields (so I know how much to verify)
- Any items listed under Exceptions

Name the unnamed

AI reads untagged content and assigns labels, categories, or metadata that didn't exist before.

You have a backlog of content, transactions, or documents with no tags, no categories, and no way to find anything. The data exists but it isn't organised.

Replaces: Manually tagging hundreds of items over a weekend you never get around to. The content library where everyone just uses the search bar and prays. The 'category' column that's been empty since 2022.

1. Provide the unlabelled content
2. Define the taxonomy or let AI propose one
3. AI assigns labels, tags, or categories to each item
4. Review and correct the taxonomy

When to use: When you have a backlog of untagged content and need to make it searchable, sortable, or reportable.

When not to use: When your taxonomy needs to be legally or regulatorily precise — AI-assigned labels should be verified before becoming official records.

Common mistake: Accepting AI's proposed taxonomy without editing it. AI will create categories that make sense linguistically but not operationally. You know your business — refine the labels.

Examples: Auto-tagging a content library, Categorising expenses, Labelling customer enquiry types, Classifying product feedback

Embodies Law 5 — One pattern, one problem: If a pattern solves two problems, it's two patterns. Atomic beats ambitious.

Spot the difference

AI compares two states and surfaces what changed.

You need to know what's different between this week and last week, between version 1 and version 2, between what was promised and what was delivered — but comparing manually is tedious and you miss things.

Replaces: Eyeballing two spreadsheets side by side and trusting yourself to catch everything. The 'what changed?' question that takes 30 minutes to answer and should take 30 seconds.

1. Provide two states — before/after, this week/last, version A/B
2. AI identifies differences and anomalies
3. Present results as a plain-language change summary
4. Flag significant changes for review

When to use: When you regularly need to track changes across time periods, document versions, or datasets.

When not to use: When the comparison requires domain expertise that can't be captured in rules — aesthetic judgments, strategic trade-offs.

Common mistake: Comparing without defining what matters. AI will list every difference — including ones you don't care about. Tell it what's significant to you.

Examples: Week-over-week metrics, Contract redline review, Competitor pricing changes, Website content drift

Embodies Law 5 — One pattern, one problem: If a pattern solves two problems, it's two patterns. Atomic beats ambitious.

Measure against the market

AI compares your content, pricing, or positioning against competitors or industry standards.

You know your own business intimately, but your understanding of how you compare to everyone else is patchy, anecdotal, and probably six months out of date.

Replaces: The competitive analysis spreadsheet that was last updated in March. Guessing how your pricing compares. The strategy meeting where everyone shares opinions instead of data.

1. Provide your own data or content
2. Provide or gather the comparison set
3. AI analyses differences in positioning, pricing, language, or features
4. Produce a comparison brief with gaps and advantages

When to use: When you need to understand your competitive position and the comparison set is accessible but too large to review manually.

When not to use: When competitive advantage comes from things AI can't observe — relationships, reputation, team quality, execution speed.

Common mistake: Comparing surface-level features without context. AI can tell you a competitor has a feature you don't — it can't tell you whether their customers actually use it.

Examples: Pricing comparison, Feature gap analysis, Job listing benchmarking, Content positioning audit

Embodies Law 1 — Start with the output: Every good automation begins with what you need at the end, not what AI can do. If you can't describe the output in one sentence, the pattern isn't ready.

Watch and alert

AI continuously watches a source and alerts you when a specific condition is met.

You need to know when something happens — a mention, a price change, a new regulation — but you can't manually check every day, and rigid keyword alerts bury you in false positives.

Replaces: Manually refreshing dashboards, feeds, and Google Alerts. The background anxiety of wondering if you've missed something. The keyword alert that fires 40 times a day for things you don't care about.

1. Define the source to monitor
2. Specify the trigger condition
3. AI watches continuously or at intervals
4. Deliver a notification with context when triggered

When to use: When missing an event has real consequences and the event is unpredictable.

When not to use: When monitoring requires deep contextual understanding to identify triggers, or when false positives are very costly.

Common mistake: Setting the trigger too broadly. 'Alert me about any competitor mention' will drown you. 'Alert me when a competitor launches a new pricing page' is actionable.

Examples: Brand mention monitoring, Price drop alerts, Regulatory change detection, Negative review alerts

Embodies Law 8 — Complexity is a cost: Every step you add is a step that can break. The best patterns have three to five steps, not twelve.

Canary in the data

AI scans your regular data for early warning signs before a problem becomes obvious.

You discover problems in quarterly reviews that were visible in the data weeks ago. Churn, cash flow dips, quality drops — they all leave traces before they become crises. But nobody's watching closely enough.

Replaces: The quarterly review where you discover a problem that started three months ago. The 'how did we not see this coming?' conversation that keeps happening.

1. Define the data source and normal ranges
2. AI scans for anomalies, trends, or threshold breaches
3. Flag early warning indicators with context
4. Deliver a plain-language risk summary

When to use: When you have regular data — weekly sales, monthly churn, daily traffic — and want to catch problems early rather than react late.

When not to use: When your data is too noisy for patterns to be meaningful, or when the cost of false alarms outweighs the cost of late detection.

Common mistake: Monitoring too many metrics. If everything is a warning, nothing is. Pick the 3–5 numbers that genuinely predict trouble.

Examples: Revenue trend anomalies, Customer churn early signals, Website performance drops, Inventory depletion forecasting

Embodies Law 4 — Replace the groan, not the craft: Automate the parts of your day you dread. Protect the parts that require your judgment, relationships, or taste.

Refit for channel

AI takes existing content and reshapes it for a different platform or format.

You wrote a great blog post. Now you need a LinkedIn version, an email version, and three social posts. The thinking is done — but the reformatting takes longer than the writing did.

Replaces: Rewriting the same thing four ways. The 'can someone make this into a tweet thread?' request that takes an hour. The content calendar that's half-empty because repurposing is tedious.

1. Provide the source content
2. Define the target format and platform constraints
3. AI produces the adapted version
4. Human reviews for voice and accuracy

When to use: When you regularly repurpose the same content across platforms and the core message stays the same.

When not to use: When each platform needs genuinely different substance, not just different packaging. Refitting is for form, not argument.

Common mistake: Letting AI flatten your voice across platforms. A LinkedIn post shouldn't sound like a tweet. Specify the tone and constraints for each platform explicitly.

Examples: Blog to social posts, Long-form to newsletter, Webinar recap to article, Case study to one-pager

Embodies Law 7 — Tools change, patterns don't: A good pattern works with Claude, GPT, Gemini, or whatever comes next. If it's locked to one vendor, it's a tutorial, not a pattern.

Bridge the jargon

AI rewrites specialist language so a different audience can understand and act on it.

Your developers write docs that clients can't read. Your lawyers send advice that nobody understands. The knowledge exists — but it's locked behind language that only the author's peers can parse.

Replaces: The client who nods along but doesn't actually understand. The internal wiki nobody reads because it was written by engineers for engineers. The 'can you explain this in plain English?' email you send three times a week.

1. Provide the specialist content
2. Define the target audience and what they need to do with it
3. AI rewrites preserving meaning but removing jargon
4. Expert reviews to ensure nothing critical was lost

When to use: When the knowledge gap between the writer and the reader is large enough that the original content isn't useful to the people who need it most.

When not to use: When precision of terminology is legally or technically required — regulatory filings, medical instructions, legal language that must be exact.

Common mistake: Oversimplifying to the point of inaccuracy. AI will happily remove nuance that matters. Always have the original expert review the simplified version.

Examples: Developer docs → client summary, Medical results → patient language, Legal terms → plain English, Financial reports → board brief

Embodies Law 6 — Describable in a sentence: If it takes a paragraph to explain what the pattern does, it's not a pattern yet — it's a project.

You are a translator between specialist language and plain English. Your job is to make expert content accessible without losing the meaning that makes it valuable.

## 1. What the reader needs to do with this
[e.g. make a yes/no decision on a contract / understand their test results / brief their team on a technical change — this shapes how you simplify]

## 2. Who the reader is
[e.g. a client with no legal background / a non-technical board / a patient with no medical training]

## 3. Source material
[Paste the specialist content — technical docs, legal text, medical report, financial analysis, developer notes]

## Instructions
Rewrite the source so the stated reader can understand and act on it without asking for clarification. Follow these rules:
- Keep the same logical structure as the original
- Replace jargon with plain language wherever possible
- When a technical term is essential for the reader to know, keep it and add a plain explanation in parentheses on first use
- If simplifying a passage would change its meaning, include both: your plain version first, then the original wording in parentheses so an expert can verify
- Preserve only the opinions and recommendations that exist in the original — add none of your own

At the end, include:
1. **Terms kept** — specialist words you preserved, with one-line definitions
2. **Meaning-check flags** — any passages where simplification may have shifted the nuance, so the original author can verify

Localise and adapt

AI adapts content for a different cultural context or market — beyond word-for-word translation.

You're expanding to a new market and your content needs to feel local, not translated. Direct translation misses idiom, cultural references, and audience expectations.

Replaces: Running everything through Google Translate and hoping for the best. The 'we'll localise later' decision that becomes 'we never localised.' Hiring a translator for every routine update.

1. Provide the source content and target locale
2. AI translates and adapts cultural references, tone, and format
3. Flag areas where human cultural review is needed
4. Native speaker reviews the final output

When to use: When you're expanding to new markets and need content that feels native, not just grammatically correct.

When not to use: When cultural sensitivity is extremely high — brand naming, political messaging, or anything where a misstep has outsized consequences.

Common mistake: Treating localisation as translation. Translating words is the easy part. Adapting examples, humour, formatting conventions, and expectations is where the real work lives.

Examples: Marketing copy for new markets, Product descriptions across locales, Help documentation in multiple languages, Social content adaptation

Embodies Law 3 — AI drafts, humans decide: The most durable patterns keep a human in the approval loop. Not because AI is bad, but because trust is earned incrementally.