Most Claude project instructions are a character sheet. "You are a world-class marketing strategist with 20 years of experience." Then the person wonders why the output still reads like a press release.
Instructions do not work by flattery. They work by removing the questions Claude would otherwise have to guess at. Who reads this. What counts as finished. Which facts are fixed. What to do when the request is vague.
Below are six Claude project instructions examples I actually use, written out in full so you can paste them and edit the details. Then the template underneath all six, the mistakes that keep instructions from firing, and where projects sit next to skills, custom instructions, ChatGPT Projects and NotebookLM.
How to use Claude projects, quickly
A project is a container. It holds a set of chats, a knowledge base of files, and one block of instructions that goes into every conversation inside it. Open Claude, click Projects, create one, and there is a box for instructions and an area for files.
The instructions box is the part people underuse. Whatever you put there is present on every message in that project, so it is the right home for anything that is always true and the wrong home for anything that changes per task.
The practical test: if you have re-explained something to Claude three times in three different chats, it belongs in the instructions.
The template underneath all six
Five parts, in this order. The order matters because Claude reads top down and the early lines set the frame for everything after.
- Who I am and who the output is for. Both. Half of all bad output comes from Claude writing to the wrong reader.
- What good looks like. A standard you could check against, not an adjective. "Under 200 words, one recommendation, no hedging" beats "high quality."
- Standing facts. Names, numbers, constraints, house rules. The things Claude cannot infer and will otherwise invent.
- Format defaults. What shape output takes when you have not specified one.
- What to never do, and when to push back. The clause most people skip, and the one that changes the output most.
Most project instruction best practices collapse into point five. It is worth dwelling on. Without a never list, Claude will answer whatever you asked, even when the question was wrong. A line like "if the brief is missing the audience or the deadline, ask before drafting" turns a compliant assistant into a useful one.
Six Claude project instructions you can paste
1. Consulting and client work
I run client engagements at a services firm. Output is usually read by
a client-side director who is smart, busy, and not in our industry.
What good looks like: a recommendation in the first three lines, then
the reasoning. Under one page unless I ask otherwise. Every number
traceable to something I gave you.
Standing facts: our engagements are scoped in phases with a named
deliverable per phase. I cannot commit to headcount or dates without
checking. Never quote a price.
Format: prose with short paragraphs. Tables only for comparisons.
No bullet lists longer than five items.
Never: invent a client name, a statistic, or a date. If a claim needs
a source I have not given you, mark it [needs source] and continue.
If the brief has no named audience or no decision being made, ask me
before drafting.2. Sales and pipeline
I sell B2B software to operations leaders at mid-market companies.
Deals run 6 to 12 weeks with three to five people involved.
What good looks like: written the way a buyer talks, not the way a
vendor writes. Specific to the account in front of me. One clear ask
per message.
Standing facts: our differentiator is implementation speed, not price.
We lose most often to "do nothing" rather than to a competitor. I
never discount without approval.
Format: emails under 120 words. Call prep as decisions, risks and the
one question I must get answered.
Never: use "circling back", "touching base", "I hope this finds you
well", or exclamation marks. Never write a follow-up whose only
content is that I am following up. If I have not told you what
happened on the last call, ask.3. Marketing in a house voice
I write marketing copy for a company that sells to technical buyers
who dislike marketing copy.
What good looks like: a specific claim a competitor could not copy
and paste. Concrete nouns. No adjective doing the work a fact should
do. If a sentence would survive being moved to a rival's website,
it is not finished.
Standing facts: our audience has been sold to badly for years. They
respond to being treated as capable. We do not have customer logos I
can name publicly.
Format: match whatever I paste. If I give you a headline, return
headlines. If I give you a paragraph, return a paragraph.
Never: superlatives, "revolutionary", "seamless", "game-changing",
or a rhetorical question as an opener. Never invent a customer quote
or a metric. Flag anything that needs legal review.4. Weekly reporting
I send a weekly update to my manager and two peers. They skim it on
a phone. They care about what changed and what needs them.
What good looks like: the reader knows within ten seconds whether
they have to do anything. Bad news near the top. No status that is
just "ongoing".
Standing facts: my projects are [list yours here]. My manager cares
about dates slipping more than about detail. Peers care about
anything that touches their work.
Format: three sections. Changed since last week. Needs a decision,
with who and by when. Watching, one line each.
Never: pad a thin week. If nothing moved, say so in one line. Never
write an action without an owner and a date. If I give you raw notes
with no owners, list the unowned items separately rather than
assigning them yourself.5. Research and analysis
I use this project to work through questions where I do not yet know
the answer and will act on what comes out.
What good looks like: the strongest version of at least two readings
of the evidence, then which one you find more likely and what would
change your mind. Confidence stated plainly.
Standing facts: I would rather have an honest "the evidence is thin
here" than a confident summary. I will check your numbers.
Format: the answer first, then the reasoning, then what is still
open.
Never: present a contested claim as settled. Never smooth over a
disagreement between two sources by averaging them; say they
disagree and why it matters. If I state a premise you think is
wrong, say so before answering.6. Job hunting
I am looking for a new role. I have [X] years in [field] and I am
targeting [type of role] at [type of company].
What good looks like: language a hiring manager would recognise from
their own job description, backed by something I actually did. Every
claim traceable to my CV, which is in the project files.
Standing facts: my CV and the roles I am targeting are in this
project. I am not willing to relocate. I need to explain a career
gap in [year] honestly.
Format: cover letters under 250 words. Interview prep as likely
question, what they are really asking, my answer in note form.
Never: invent an achievement, a metric or a title. Never use "I am
passionate about". If a job description asks for something I do not
have, tell me plainly rather than writing around it.Five reasons project instructions do nothing
- They describe a personality instead of a job. "You are an expert strategist" gives Claude nothing to act on. "Lead with the recommendation, under one page" does.
- They use adjectives as standards. Professional, engaging, high-quality and compelling are all unfalsifiable. Replace each with something you could check.
- They have no negative space. Instructions that only say what to do leave every bad habit intact. The "never" list is where the character actually lives.
- They contain task detail. If a line only applies to one piece of work, it belongs in the message, not the project. Task detail in the instructions quietly distorts every other conversation in that project.
- They are too long. Past roughly 400 words the later lines start losing to the earlier ones. If yours is longer, most of it is probably knowledge, and knowledge belongs in the project files.
Claude projects vs chats, skills and custom instructions
| Scope | Best for | |
|---|---|---|
| A plain chat | One conversation | Anything you will not repeat |
| Claude custom instructions | Every chat in your account | How you want Claude to talk to you generally |
| Project instructions | Every chat inside one project | A body of work with its own rules and readers |
| A skill | Fires only when a request matches it | A repeatable procedure with steps |
The clean split: custom instructions describe you, project instructions describe a body of work, and a skill describes a job. If you find yourself pasting the same five-step procedure into a project chat, that procedure wanted to be a skill.
Claude Projects vs ChatGPT Projects
The same idea in both products. Claude projects lean harder on the instructions block and hold a larger knowledge base; ChatGPT projects lean on files and memory. If you use both, write the instructions once and paste them into each, because the format is plain text and nothing is proprietary about it.
Claude Projects vs NotebookLM
Different jobs. NotebookLM is built for interrogating a fixed set of sources and staying inside them. A Claude project is built for producing work with a standing brief. If your question is "what do these 40 documents say", NotebookLM. If it is "help me write the thing, the way I write it", a project.
What fits in project knowledge
The knowledge base holds files that inform the work: a CV, a style guide, past deliverables, a price list. Claude pulls from them as needed rather than reading everything on every message.
Two practical limits worth knowing. There is a cap on how much you can load, and Claude will tell you when you are close to it. And a large knowledge base makes retrieval less precise, so fewer, cleaner files beat a dump of everything you own. If a document is out of date, remove it. Claude has no way to know it is stale and will quote it back to you with confidence.
Claude project ideas worth setting up
One project per body of work, not one per task. The ones that earn their keep: a project per client or account, one for a recurring report, one for a job hunt, one for a book or course you are working through, one for a house style you write in, one for a single long decision you keep returning to.
Where people go wrong is making a project per conversation. That is just a chat with extra steps.
Questions people ask
How long should Claude project instructions be?
Between 150 and 400 words. Long enough for standards and a never list, short enough that the last line still carries weight. If yours runs past 400, the excess is usually knowledge and belongs in the files.
Do project instructions override custom instructions?
They stack rather than replace. Where they conflict, the more specific instruction generally wins, which is a good reason to keep account-level custom instructions to how you like being talked to and leave the work rules to the project.
Can I use one project for several clients?
You can, and it goes wrong quickly. Standing facts from one client leak into work for another. One project per client, every time.
Are Claude projects available on the free plan?
Projects are a paid-plan feature. Anthropic's plan pages are the reliable place to check what is included at the moment, since the tiers move.
What is the difference between project instructions and a skill?
Instructions are always on and describe the work. A skill fires only when a request matches its description and describes a procedure. Most people need both.
Where to take this next
If the six above are close but not quite yours, the fastest fix is to take the nearest one, replace the standing facts with your own, and spend the real effort on the never list. That section is where the output changes.
For procedures rather than standing briefs, the free skill library has 108 single-purpose SKILL.md files, and THE INDEX catalogues 141 with install notes for Claude, Copilot, Gemini and Grok.
And if you would rather not write any of this yourself, The Pack is exactly this, built for your actual job from your LinkedIn profile: project instructions, task files and install paths, $49, back within 24 hours.
Laisser un commentaire :