A T I M B E R & I N K F I E L D G U I D E
A field guide to writing good prompts and using Claude.ai well.
If the workshop was about building the workshop, this is about doing the work inside it. The shop is built. The tools are arranged. Now we sharpen the chisel and learn how to read the grain of the wood.
The workshop made one big claim: AI gets dramatically more useful when you stop treating it like a search engine and start treating it like a workspace. That claim is true, but it isn't enough on its own. A workshop is only as useful as the work that happens inside it — and the quality of that work depends on two things you do every single time you sit down: how you ask, and which tools you reach for.
This guide covers both. The first half is about the craft of the prompt itself — how to write something worth answering, regardless of which platform you're using. The second half is a practical field guide to the features inside Claude.ai specifically — when to use which model, what styles and skills are actually for, what Adaptive Thinking does, and when Cowork earns its keep.
You don't need to read it all at once. Skim the headings, find the part you need, come back later for the rest. That's how field guides are supposed to work.
Part One
Draft outside the AI first.
The single biggest upgrade to anyone's AI use — bigger than choosing the right platform, bigger than learning fancy techniques — is this: stop typing your first thought into the chat window.
Most people sit down to a blank chat and start writing the prompt straight into the box. The cursor blinks. The pressure is on. You half-think, half-type, and what comes out is roughly what was rattling around in your head five seconds before you opened the tab. The AI does its best with the half-formed question. You get a half-formed answer. You re-explain. You correct. You re-roll the response. Twenty minutes in, you've had three rounds of clarification just to land where you wanted to start.
Here is the simple discipline that solves this: before you open the AI, open a note.
Apple Notes. Google Keep. A scrap of paper on the desk. A markdown file. It does not matter which one. What matters is that you draft your prompt in a space that isn't pressuring you to send it.
A note has three quiet gifts. It lets you slow down — you can pause mid-sentence, walk away for coffee, and the cursor is still there waiting, no clock running. It lets you see the whole thought at once, the way you see the whole shape of a letter before you put it in the envelope. And it gives you a record. The next time you need a similar prompt, you don't start from scratch — you copy, edit, refine. Over weeks, you build a small library of prompts that have already proven themselves.
This single habit will change your AI use more than any other. If you take nothing else from this guide, take this.
A draft is not a polished essay. It's a worked-through note. Here is the shape of one:
A Drafted Prompt
I'm writing a midweek devotional for our Wednesday night student group — 7th–12th graders, about 30 of them, mixed church-and-curious. The topic is forgiveness. Not the cheap "let it go" kind, but the real, costly, restorative kind. I want to ground it in Joseph and his brothers (Genesis 50). 10–12 minutes spoken. I want one good question to send them out on. I do NOT want a full lesson plan with games and breakouts — just the talk.
That's a draft. It's not formatted prettily. There are dashes and parentheses and an emphasized word. But every important thing is there: who you are, who the audience is, what the topic is, what the form is, what the boundaries are, and what the deliverable looks like. Drop that into Claude or ChatGPT and you will get back something genuinely usable on the first try.
Compare it to what most people would actually type into the box: "Write me a devotional on forgiveness for teenagers."
Same task. Wildly different result.
What to put in the draft.
A strong prompt has four layers stacked underneath it. You don't have to think of them as a checklist when you're drafting — but if your prompts ever feel like they're producing weak answers, run through these four and see which one is missing.
The AI cannot read you. It does not know if you are a hospital nurse, a homeschool mom, a small-business owner, a youth pastor, or a writer working on a novel. It does not know your level of expertise, your tone preferences, your faith tradition, your industry, or the specific vocabulary your field uses. It will guess, and the guess will be average.
A short sentence at the top of the prompt fixes this. "I'm a clinical informatics specialist supporting Cerner." "I'm a 7th–12th grade volunteer student minister at a Baptist church in Tulsa." "I'm planning a 10-year anniversary trip and we have two kids under five." The AI is now no longer working with the average person — it is working with you.
If you have a Custom Instructions or Project Instructions set up — and the workshop strongly recommended that you do — much of this happens automatically and you don't have to repeat it. But for one-off chats outside a project, this layer is what the project would have given you for free.
State what you want, plainly. Not "help me with" — what specifically. Help me draft, outline, review, edit, brainstorm, compare, summarize, explain, translate, rewrite for a different audience, plan, decide between.
Verbs do work that adjectives cannot. "Help me with my email" is a vague gesture. "Rewrite this email to sound less defensive" is a job description. The first asks the AI to figure out what you mean. The second tells it.
If you don't know what verb fits, that's often a sign the task itself isn't clear enough yet — and that's actually useful information. The note is the place to figure that out before you send anything.
This is the layer most people forget exists, and it is one of the most important. The AI, in any given chat, only knows what is in that chat. It does not remember the brilliant analysis you got in a different thread last week. It does not remember the decisions you and the AI worked through three days ago. Each new conversation begins with what the workshop called the amnesia problem.
You bridge that gap manually, with a few well-chosen sentences.
A Bridge Sentence
Last week in another chat I worked through a strategy for the Q3 launch. The decisions we landed on were: focus on existing customers first, delay the public announcement until July, and emphasize reliability over features in the messaging. I want to build on those decisions in this conversation.
The AI cannot verify that those decisions were actually made — it has no memory of the previous chat. But it can take them as given and reason forward from there. That is exactly what you want. You are not asking the AI to confirm history; you are giving it the inputs it needs to keep working.
One feature worth turning on while you're at it: Claude.ai has a Memory setting that does some of this work automatically. Once you enable it in your settings, Claude can carry recurring context across chats — your role, your projects, communication preferences, things you've referenced before. It's a real upgrade over the cold-start amnesia we've been describing.
But Memory captures patterns, not specifics. It might remember that you write devotionals for 7th–12th graders; it won't reliably remember the three decisions you made in last Tuesday's strategy chat. And even when Memory is holding something, it may not weight it the way the current task needs it weighted.
So the manual discipline still holds. Bridging in the relevant prior context — even when Memory is on — is more reliable than assuming the AI will recall everything as fully as the work requires. Reminders beat assumptions, every time.
If your prompt involves a document, a draft, an outline, an email thread, a Bible passage, or any other text — paste it in. Don't describe it; give it. Don't summarize what your boss said in the email; show the email. Don't paraphrase the verse you're studying; quote it.
The AI is a reading machine. The quality of what it produces from a piece of text is enormously higher than the quality of what it produces from your description of a piece of text. If you have the source material, share it.
The exception, and an important one: never paste sensitive material — patient information, confidential business data, personal details about other people who haven't consented — into a consumer AI account. The billboard test from the workshop applies. If you wouldn't put it on a billboard, don't paste it into a chat that doesn't have a Business Associate Agreement or equivalent protection.
Telling the AI what not to do.
A well-crafted prompt is not just a description of what you want. It is also, often more importantly, a description of what you don't want. Most weak AI outputs are not weak because the AI failed at the task — they are weak because the AI did three other tasks alongside it that you didn't ask for.
You ask for a draft of a memo. You get a draft of a memo, plus three alternative versions, plus a list of "key considerations to keep in mind," plus a closing offer to "develop any of these further." All of that is well-meaning. None of it is what you asked for. And every paragraph of it is a paragraph you have to read and discard before you can do anything useful.
Guardrails are how you shut the side doors so the AI walks straight to the room you want.
Scope
"Stay focused on the email itself. Don't include alternative versions or suggestions for follow-up."
This is the single most useful guardrail in casual AI use. It tells the AI that the task is the task — not the task plus a buffet of related options.
Format
"Plain prose, no bullet points. Three paragraphs, around 200 words total."
AI loves bullet points. Most of the time, you don't actually want them. Specifying format up front saves you from having to ask for a rewrite. If you want a table, say so. If you want it short, give a word count. If you want it conversational rather than formal, say that.
Tone
"Warm but direct. Not breezy. Not corporate."
AI has default tones — usually slightly cheerful, slightly polished, slightly generic. If you want something else, name it. Specific is better than vague: "warm but direct" beats "natural"; "the way a thoughtful pastor would say it" beats "pastoral"; "plainspoken, no jargon, no hedging" beats "clear."
Authority
"Where you're making claims of fact, mark them as such. If you're not sure of something, flag the uncertainty rather than guessing."
This one matters more in research and writing tasks than in casual ones. It is your defense against the AI's worst habit — confidently inventing things. The line from the workshop holds: AI is a fantastic assistant and a terrible authority. The authority guardrail is what reminds the AI of its own limits inside its output.
The most underused guardrail of all is a one-line description of the finished thing. "The output should be ready to paste into an email, with no further editing needed." Or: "The output is a working draft I'll edit, so don't worry about polish — focus on getting the structure right." Or: "The output is a thinking partner's response, not a final document — questions back to me are welcome."
Telling the AI what done looks like is how you prevent it from over-polishing or under-developing. It also tells the AI when to stop. Many AI outputs sprawl because nothing in the prompt told them where the finish line was.
The same task, before and after.
Here is the same task, before and after, so you can see the layers stacked up.
Before
Help me write a thank-you message to the families who hosted the youth retreat.
After — Drafted in a note, then pasted in
I'm a volunteer student minister, 7th–12th grade, at a Baptist church in Tulsa. Last weekend was our spring retreat — three families opened their homes to host small groups overnight. I'd like to send each family a personalized thank-you note: handwritten, short, warm, and specific to what they actually did (not generic).
The three families and what they hosted:
Please draft three short notes, around 80–120 words each. Warm but not gushing. Specific to what each family did. Plain prose, no bullet points. Sign-off should be "Grateful, Joshua."
Don't include alternative versions or general advice — just the three drafts.
Same task. The first version will produce something usable. The second version will produce something you can copy almost word-for-word. The difference is not five times the effort. It is maybe two minutes of drafting in a note before opening the chat.
That is the whole craft, in a single example.
How prompts go sideways — and how to catch them in the draft.
A short list of the most common ways prompts go sideways, and how to catch them in the draft phase before they become wasted chat threads.
Part Two
Everything in Part One applies to any AI platform. The ideas are the same whether you're in Claude, ChatGPT, Copilot, or Gemini. Part Two is different — it's a practical guide to the specific features Claude.ai offers and when to reach for which one.
A reminder before we open the toolbox: features amplify a good prompt, they do not replace one. Choosing the right model and turning on the right feature can make a strong prompt much stronger. They cannot rescue a weak one. If your prompt isn't working, the answer is almost never "switch the model" — it's almost always "fix the prompt."
That said, knowing what each tool actually does will save you time, money, and frustration.
Sonnet, Opus, Haiku — and when to reach for each.
Claude offers several models, and they are not interchangeable. Picking the right one is partly about quality and partly about stewardship — using a heavier model for a lighter task wastes your usage limits and burns more energy than necessary.
As of this writing, the consumer-facing options are Sonnet, Opus, and Haiku. Free users get Sonnet only; Pro and Max users can choose among all three.
The default workhorse
Fast, fluent, capable of substantial reasoning, and produces high-quality writing across nearly every common use case. If you don't have a specific reason to pick something else, pick Sonnet.
The deeper model
Thinks longer, holds more in its head at once, and produces noticeably better work on tasks that genuinely require it — long, complex creative writing; multi-step reasoning; deep analysis of large documents; theological or philosophical work that benefits from nuance.
The lightweight model
Excellent for quick lookups, simple summaries, format conversions, fast drafts, and any task where speed matters more than depth. If you find yourself using Sonnet to reformat a list, switch to Haiku.
A quick decision table:
The simplest rule: start with Sonnet, escalate to Opus when the work demands it, drop to Haiku when speed matters more than thoughtfulness.
Teaching Claude your voice.
A Style in Claude.ai is a saved set of communication preferences — tone, format, level of formality, length, point of view, signature phrases, things to avoid. You can switch between styles mid-conversation. You can create custom styles by giving Claude a writing sample and saying "write like this."
Styles are useful in three specific situations:
A pastor who writes both casual newsletter pieces and serious teaching content might keep one style for "Sunday newsletter voice" and another for "Wednesday teaching voice." Switching styles is faster than re-prompting tone every time.
Yours, your brand's, or someone else's you're writing in the persona of. Feed Claude a few samples, save the style, and apply it whenever you're writing in that voice.
Claude's default is generally clear, slightly formal, somewhat thorough. If you consistently want shorter, sharper, more conversational replies — save a style that reflects that and apply it as your default.
Project Instructions are about the work — who you are, what the project is for, what guardrails apply. Styles are about the voice — how the words should sound. They work together. Project Instructions tell Claude what to do; Styles tell Claude how it should read on the page.
If you only do one thing with Styles: create a custom style that reflects your actual writing voice and apply it whenever you're drafting anything that will go out under your name. The output will sound dramatically more like you and require dramatically less editing.
When to let Claude slow down.
Adaptive Thinking (sometimes called Extended Thinking) is a setting that lets Claude work through problems more deliberately before answering — essentially, taking time to reason rather than producing the first plausible response. When you turn it on, Claude takes longer to respond, and the response itself is the result of deeper, more careful reasoning.
This is genuinely useful, and it is also genuinely overused.
The quiet tools that already know what to do.
Skills in Claude are reusable instruction sets that activate automatically when Claude detects a relevant task. They run in the background — you don't have to call them by name. You ask Claude to make a polished Word document, and a docx skill activates with detailed instructions on how to make it well. You ask for a presentation, and a pptx skill activates. You ask for a PDF, and a pdf skill activates.
For the user, this mostly looks like: "Claude got noticeably better at certain specific tasks at some point and I'm not sure why." That's the skills doing their work.
You don't usually need to think about skills. They activate automatically when relevant. The right mental model is not "I should turn on a skill" — it's "I should describe my task clearly, and Claude will reach for the right tool."
Skills are particularly useful for document creation. If you want a polished Word doc, a real presentation file, a working PDF, or a well-formatted spreadsheet, asking Claude to create the file directly will produce noticeably better results than asking it to give you content to paste somewhere yourself. The skills know things about formatting, layout, and file structure that would take you ages to specify by hand.
The simplest practical advice: ask Claude to make the actual file, not the content for the file. "Create a Word document with…" rather than "Give me the content for a Word document about…" The first will use the relevant skill and produce something more polished. The second asks you to do work the AI is now equipped to do.
The sandbox workshop.
Cowork is the newest of the major Claude features and arguably the most powerful for certain kinds of work. It is essentially Claude with hands — Claude Desktop running on your actual computer, with access to the files and folders you give it. Code runs safely in a sandboxed virtual machine, but the files Claude reads, writes, and edits are real files on your machine. It can produce and modify files in batch, run multi-step workflows, schedule recurring tasks, and connect to a growing list of external services.
The right way to think about Cowork: regular Claude is for thinking and one-off creating. Cowork is for production work that involves multiple files, multiple steps, or assembling something complex.
A useful analogy: regular Claude is the sketch desk where you think and draft. Cowork is the workshop bench where you actually build the deliverable. Most of the work happens at the sketch desk. The bench earns its keep on the days when you're producing finished pieces.
A note on access: Cowork is a paid-tier feature and requires Claude Pro or higher, plus the Claude Desktop app for macOS or Windows — it's not available in the web or mobile claude.ai. If you're on the free tier or working primarily on mobile, most of what you'd want to do for everyday AI use can happen comfortably in regular Claude.
If you want one page to keep nearby, this is it.
For the Workbench
A page to keep nearby.
The shift this guide is asking for is not big, and it is not technical. It is the difference between thinking on the keyboard and thinking on paper before you ever touch the keyboard. It is the difference between reaching for the heaviest tool out of habit and reaching for the right tool on purpose. It is the difference between a question that earns a real answer and a question that earns whatever the AI happens to produce.
The workshop made the case that AI is a workspace, not a search engine. This guide is the next layer down — the daily craft of working inside the workspace. Better prompts, the right model, the right feature, and the steady habit of thinking before you ask.
Once these become reflexes, you stop noticing them. The AI just gets noticeably more useful, the answers just feel more like what you needed, and the time you used to spend re-prompting and re-rolling becomes time you spend on the actual work.
That's the goal. Not impressive AI tricks. Just a tool that quietly does what you needed it to do, the first time, more often than not.
— Joshua