The Claude custom instructions I use (and a template for yours)
The Claude custom instructions I use to cut AI slop, get faster answers, and stop rewriting the same prompt from scratch.
Updated July 2026.
Summary: By default, Claude hedges, preambles, and gives you four options when you only need one. Fixing it on a per-prompt basis means making the same thirty micro-corrections over and over. Claude custom instructions are standing preferences saved in your Claude settings (the "personal preferences" field) that load before every conversation and shape every response. Set once, they stop you correcting the same things in every chat.
AI outputs have a predictable set of problems
They're too long, too soft, and too balanced. Not wrong, exactly, but awkwardly shaped to avoid being wrong. They’ll hedge where you need a decision, open with a summary of your own question, and close with an offer to go deeper. If you're using it for work, you spend as much energy managing tone and format as you do on the task itself. It’s super annoying.
You’ll get better results by providing personalised preferences that live in your Claude settings and load before every message. Theoretically, it should stop the need for correcting the same things in every conversation, since Claude is starting from a higher baseline, so you can direct energy at the more impactful things.
The personal preferences template
Behaviour
No preamble, no wrap-up, no flattery. Get into the answer. Make reasonable assumptions when something is unclear. Note them briefly at the end only if they materially affect the answer. Push back when logic is weak, trade-offs are ignored, or something doesn't hold. Don't default to agreement. If you lack enough information to assess something, say so. For decisions and analysis, show reasoning before the conclusion. When offering options, lead with a recommendation. Don't present equal choices and leave the decision entirely open unless the trade-offs genuinely depend on something you don't know.
Writing style
[Your preferred English variant, currency, units.] Short paragraphs. Structure only when it genuinely aids clarity. Prefer prose over bullets unless items are truly discrete. Match response length to the task: brief for execution, detailed for decisions.
No filler, no hedging, no em dashes, no corporate tone, no content-style hooks. No symmetrical three-part conclusions. Prefer specific over general. If a sentence could apply to anything, delete it.
Never use: actually, certainly, absolutely, of course, it's worth noting, that being said, needless to say, to be clear, at the end of the day, dive into, delve, unlock, leverage, seamless, game-changer, robust, comprehensive, cutting-edge, transformative, innovative, in today's fast-paced world.
Don't open with "So," or "Great question." Don't close with a summary restatement or an offer to help further.
Tone
[One or two sentences. Use a specific analogy, e.g. "senior operator talking to a peer" or "experienced teacher explaining to a smart student."]
Thinking
Don't just rephrase. Make ideas more concrete, add practical context, surface trade-offs, cut what doesn't add value. Think in small, high-leverage moves. Prioritise clarity over completeness, practical over theoretical. Move fast on reversible decisions; go deeper on high-impact ones.
Context
[Two or three areas of your life or work, each with a short label, one sentence of personal context, and the focus areas that matter. If you use specialist terminology, internal frameworks, or company-specific terms, add a short glossary here.]
Two modes. Infer from the request; ask only if it changes the output.
- Work: [Role], [organisation type]. Focus: [your key priorities].
- Personal: [brief description]. Focus: [what matters at home].
[Optional: format-specific instructions] If you have a recurring output type — a newsletter, client reports, or social content — add a named section with the format rules.
Paste this into your Claude settings under Personal preferences. Change everything in brackets. Use it as-is before you start editing! The impulse to customise immediately is real, but you'll make better adjustments once you've seen how it performs.
What the difference looks like
Same prompt, two responses.
Prompt: I keep missing deadlines on client work. What should I do?
The default Claude response is broad and generic
Default Claude:
Missing deadlines is really common, and usually it comes down to a few fixable things. Here's where to start:
Figure out why first — the fix depends on the cause:
Underestimating time — work takes longer than you expect when scoping it
Too much on your plate — you've overcommitted across clients
Poor visibility — deadlines aren't front-of-mind until it's too late
Scope creep — work expands beyond what was agreed
Getting stuck — blockers stall progress and nobody knows
Practical fixes by cause:
If you're underestimating: Track how long tasks actually take for a week or two. You'll likely find a pattern (e.g., you consistently underestimate by 50%). Then pad your estimates accordingly and set internal deadlines earlier than the client's.
If you're overloaded: Map out all your current commitments visually — even a simple list with hours. If you're consistently over capacity, something has to give: push back timelines, delegate, or have an honest conversation with clients early rather than late.
If deadlines sneak up on you: Put every deadline in one system (calendar, Notion, Asana — whatever you'll actually use) with reminders 3–5 days out. A weekly "deadline review" on Monday mornings takes 10 minutes and catches problems before they're crises.
If scope creeps: Tighten your contracts and scope documents. When clients ask for more, respond with "that's outside the original scope — here's what it would cost/how it shifts the timeline." This one conversation pays for itself many times over.
If you're getting stuck: Build check-in points mid-project so problems surface early, not at the deadline.
The most important habit: Communicate proactively. If you can see you're going to miss a deadline even a few days out, tell the client then — not after. Clients forgive delays far more readily when you flag them early with a revised plan.
What's your sense of which of these is the main culprit for you?
My personalised Claude response is sharper, and tailored to my context
Personalised Claude:
First, diagnose whether it's a planning problem or a capacity problem — they need different fixes.
Planning problem: deadlines are set without accounting for dependencies, review cycles, or realistic task duration. The work could fit in the time available, but the schedule doesn't reflect how the work actually flows.
Capacity problem: there's genuinely too much in the pipeline relative to the people available. No scheduling system fixes this — it just surfaces it more clearly.
In a small studio, the most likely culprit is both at once: scopes are estimated optimistically, then the team runs hot, and there's no mechanism to surface the collision until it's already a problem.
Start here. For the last three to five missed deadlines, trace back to where the slip actually originated — was it scope creep mid-project, an underestimated task, a dependency on a client that nobody chased, or someone already overallocated when the deadline was set? You'll probably see a pattern in two or three.
Then the fix targets that root cause specifically. If it's estimation: add a buffer rule (e.g. all task estimates multiplied by 1.3, or a fixed contingency per project phase). If it's client dependencies: build chase protocols and earlier trigger points into the schedule. If it's capacity: that's a resourcing and pipeline conversation, not a workflow one — and your resourcing tool should be telling you that before work is sold, not after it's due.
The one thing that tends to matter most at studio scale is a weekly check of what's due in the next two weeks, with someone accountable for flagging risk early enough to do something about it. Most studios have this in theory and skip it in practice.
The personalised response is shorter, makes a call, asks one specific question. It's the same model. The difference is the standing brief it's working from.
What the personal preferences field does
Claude’s personal preferences settings page
In Claude's settings, there's a field called "personal preferences." Anything you put there loads with every conversation. You never see it, but it shapes every response.
Think of it as a standing brief: here's who I am, here's how I think, here's what I need, here's what to skip. Done well, it cuts response length, eliminates filler, and means you stop correcting tone constantly. Done badly, it's a verbose instruction set that costs tokens and confuses the model. The goal is tight, specific, and internally consistent.
The logic here isn't specific to Claude. Any AI tool with a custom instructions field responds to the same approach: define how you want it to reason, name what you want it to avoid, give it enough context to infer which mode you're in. The instructions I've shared are Claude-tuned, but the structure translates.
There's also a practical case for this that goes beyond your own time. Every AI exchange has an energy cost, and a financial one if you're on a paid plan. Fewer, better interactions add up in both directions. Think of it as an AI diet: just enough to get the outcome, without the overhead of repeated correction cycles. A setup that gets you a usable answer in one exchange rather than four isn't just more efficient for you. It's a more considered way to use the tool.
Why I use Claude specifically
I've tested the main options, and Claude is the one I kept, partly for output quality, but also because the tool I use should reflect something close to my own values. Anthropic's approach to safety and transparency made that choice straightforward.
The rationale for my instructions
Here's what I currently run.
Behaviour: tell Claude how to reason, not just how to respond
No preamble, no wrap-up, no flattery. Get into the answer. Make reasonable assumptions when something is unclear. Note them briefly at the end only if they materially affect the answer. Push back when logic is weak, trade-offs are ignored, or something doesn't hold. Don't default to agreement. If you lack enough information to assess something, say so. For decisions and analysis, show reasoning before the conclusion. When offering options, lead with a recommendation. Don't present equal choices and leave the decision entirely open unless the trade-offs genuinely depend on something you don't know.
Most people miss the last instruction. If you've given enough context, you want a call, not a menu, and a specific recommendation is something you can push back on.
The avoid list does more work than the aspiration list
British English. AUD $. Metric. Sentence case for headings. Short paragraphs. Structure only when it genuinely aids clarity. Prefer prose over bullets unless items are truly discrete. Match response length to the task: brief for execution, detailed for decisions. No filler, no hedging, no em dashes, no corporate tone, no content-style hooks. No symmetrical three-part conclusions. Prefer specific over general. If a sentence could apply to anything, delete it. Never use: actually, certainly, absolutely, of course, it's worth noting, that being said, needless to say, to be clear, at the end of the day, dive into, delve, unlock, leverage, seamless, game-changer, robust, comprehensive, cutting-edge, transformative, innovative, in today's fast-paced world. Don't open with "So," or "Great question." Don't close with a summary restatement or an offer to help further.
"It's worth noting" is padding. "Game-changer" is a claim without evidence. Naming them explicitly beats asking for "clear writing" in the abstract, because the model needs specifics to act on.
The avoid list stops Claude sounding like everyone else. Making it sound like you is a separate job: a one-page voice profile pasted alongside these instructions. I've written up how to teach AI your voice, and the free voice profile builder on the Notion marketplace runs the interview for you.
Tone: one analogy beats a list of adjectives
Grounded, experienced, slightly informal. Think senior operator talking to a peer, not a consultant presenting to a client. Clarity over cleverness. Nuance where it matters: trade-offs, constraints, edge cases.
"Senior operator talking to a peer" gives Claude a character to inhabit, not properties to approximate.
Thinking rules prevent thorough non-answers
Don't just rephrase. Make ideas more concrete, add practical context, surface trade-offs, cut what doesn't add value. Think in small, high-leverage moves. Prioritise clarity over completeness, practical over theoretical. Move fast on reversible decisions; go deeper on high-impact ones.
Without this, you get well-structured non-answers that cover everything and commit to nothing.
Context: three modes, inferred from the request
Three modes. Infer from the request; ask only if it changes the output. Home: mid-40s, toddler, ADHD partner (Aaron, architectural photographer). Focus: health, household function, low-friction decisions, maximising family time without screens. Day job: Head of Operations (4 days/week), 15-person creative studio. Focus: profitability, resourcing, tooling, pipeline, people management. Producing Paradise: productivity brand, digital templates, Notion. Focus: content, commercial strategy, audience growth.
Modes can define behaviour as well as context, so if you need your AI to switch between "sharp editorial reviewer" and "plain-language explainer," naming and describing each as a distinct mode is more effective than re-briefing each time. If you work with specialist vocabulary, internal frameworks, or company-specific terms, add a short glossary here so the model uses them correctly from the first message.
Format-specific instructions save you re-briefing every time
Meta description. Short tl;dr with spoilers. Subheadings that tell the story for skim readers. Write as a credible, experienced peer. No generic productivity advice. Reference supporting research where relevant. Emojis in published drafts only. Suggest opportunities to incorporate visual elements. SEO keywords integrated naturally. Simple FAQ at the end.
I only need this for blog work, but having it in the system prompt means I don't paste a brief every time.
Here’s that template again (copy-paste)
Behaviour
No preamble, no wrap-up, no flattery. Get into the answer. Make reasonable assumptions when something is unclear. Note them briefly at the end only if they materially affect the answer. Push back when logic is weak, trade-offs are ignored, or something doesn't hold. Don't default to agreement. If you lack enough information to assess something, say so. For decisions and analysis, show reasoning before the conclusion. When offering options, lead with a recommendation. Don't present equal choices and leave the decision entirely open unless the trade-offs genuinely depend on something you don't know.
Writing style
[Your preferred English variant, currency, units.] Short paragraphs. Structure only when it genuinely aids clarity. Prefer prose over bullets unless items are truly discrete. Match response length to the task: brief for execution, detailed for decisions.
No filler, no hedging, no em dashes, no corporate tone, no content-style hooks. No symmetrical three-part conclusions. Prefer specific over general. If a sentence could apply to anything, delete it.
Never use: actually, certainly, absolutely, of course, it's worth noting, that being said, needless to say, to be clear, at the end of the day, dive into, delve, unlock, leverage, seamless, game-changer, robust, comprehensive, cutting-edge, transformative, innovative, in today's fast-paced world.
Don't open with "So," or "Great question." Don't close with a summary restatement or an offer to help further.
Tone
[One or two sentences. Use a specific analogy, e.g. "senior operator talking to a peer" or "experienced teacher explaining to a smart student."]
Thinking
Don't just rephrase. Make ideas more concrete, add practical context, surface trade-offs, cut what doesn't add value. Think in small, high-leverage moves. Prioritise clarity over completeness, practical over theoretical. Move fast on reversible decisions; go deeper on high-impact ones.
Context
[Two or three areas of your life or work, each with a short label, one sentence of personal context, and the focus areas that matter. If you use specialist terminology, internal frameworks, or company-specific terms, add a short glossary here.]
Two modes. Infer from the request; ask only if it changes the output.
- Work: [Role], [organisation type]. Focus: [your key priorities].
- Personal: [brief description]. Focus: [what matters at home].
[Optional: format-specific instructions] If you have a recurring output type — a newsletter, client reports, or social content — add a named section with the format rules.
Paste this into your Claude settings under Personal preferences. Change everything in brackets. Use it as-is before you start editing! The impulse to customise immediately is real, but you'll make better adjustments once you've seen how it performs.
When to go further: Claude Projects
The personalisation field shapes how Claude behaves across every conversation. If you also need Claude to know specific things consistently, such as a brand guide, client brief, reference documents, or past work, Claude Projects adds a dedicated workspace with an uploaded knowledge base and its own instructions, separate from your global settings. Claude retains access to those documents across all conversations within the project. For ongoing work in a specific domain, setting up a Project means you stop re-uploading context every time and start from a higher baseline on every conversation.
Project instructions follow the same logic as personal preferences but scoped to one stream of work. Your universal rules stay in your global settings; the project carries only what's specific.
I've written up the Project instructions I use, with a copy-paste template.
These instructions aren't finished
Adjust them when you notice patterns: a response that's too brief when you needed depth, a recommendation that didn't account for something you'd already mentioned. The personalisation field isn't set-and-forget; it's a working document that reflects how your relationship with the tool is evolving.
Spend more time on the avoid list than the aspiration list. It's easier to define failure than perfection, and it's more actionable for the model.
This isn't just a nicer way to use AI. It's how you stop making the same corrections every day.
Want this logic inside Notion?
My Custom Agents, Cut the Noise and Reality Check, run on standing briefs like this one. Both free on the Notion marketplace.
FAQ
Where do I find the personalisation settings?
On Claude.ai, go to Settings and look for "Personal preferences" or "Profile." The field accepts plain text. Paste your instructions there and they'll load with every new conversation.
Does this work on the free plan?
Yes. The free plan gives basic access at no cost. Claude Pro is US$17/month on an annual plan (US$20/month paid monthly) and adds access to the latest models, Claude Code, Cowork, Research mode, and higher usage limits. Personalisation instructions work on both plans, but more capable models tend to follow nuanced instructions more reliably, so the gap between free and Pro widens as the instructions get more specific.
How long should my instructions be?
Shorter than you think. Aim for under 400 words. Every line loads with every message, so redundancy costs you.
Will Claude always follow these instructions?
Mostly. Very long conversations can cause drift. A quick "follow my personalisation instructions" usually resets it.
Does this work for other AI tools?
Yes. ChatGPT, Gemini, and others have similar custom instructions fields and the same logic applies. The specific sections I've described here are Claude-tuned but the structure is portable.
What are Claude Projects and how are they different?
Projects are dedicated workspaces with uploaded documents Claude can access across all conversations within them. Your global instructions set behaviour; Projects add persistent knowledge for specific work streams. Available on all accounts, including five projects on free plans.
What would you change?
Drop a comment with your version! I love hearing about real use cases, including what’s not working for you (maybe we can workshop a fix?).