First contact: the chat, the project, the artifact
Get fluent with the Claude UI: chats, Projects (persistent context), and Artifacts (live in-place editing of docs and components).
01 · Read
Explanation
Claude is a chat-first assistant from Anthropic. Unlike a search box, every reply is a function of three inputs: the system prompt (who Claude should be), the conversation history, and your latest message. If the answer is bad, one of those three is wrong — usually the first.
A Project is a persistent container. You attach style guides, prior work, PDFs, or docs once and they apply to every chat in that project. This is the single biggest unlock for non-trivial work: the model stops "forgetting" who you are between sessions.
An Artifact is a doc, table, image, or React component that Claude renders next to the chat and edits in place. Asking for output "as an artifact" is the difference between regenerating a 500-word draft six times and editing a single living document.
Key concepts
- System prompt
- A persistent instruction that shapes every reply — Claude's "role".
- Project
- A persistent context container: docs, files, and instructions that survive across chats.
- Artifact
- A live, editable output (doc, code, component) Claude updates in place.
- Stop condition
- A clause like "Stop and ask if you are unsure" that reduces hallucinations.
02 · Watch
Case study
A solo founder writes a launch announcement
Setting
Maya runs a 1-person SaaS. She has a product changelog, three blog posts in her voice, and a deadline of Friday for a launch email.
Challenge
Her last attempt at AI-written copy sounded like a marketing intern. She does not have time to rewrite line-by-line.
Walkthrough
- Maya creates a Project called "Launches" and uploads three blog posts as voice samples.
- In the system prompt: "You write in the voice of the attached samples. Default to short sentences. Never use the words leverage, unlock, or empower."
- She asks Claude to draft the email "as an artifact, 180 words, single CTA at the end."
- She uses the artifact panel to highlight one paragraph and ask "rewrite this less corporate" — Claude edits in place instead of rewriting the whole email.
- Forty-five minutes later she ships the email. Three iterations, one artifact, no copy-paste rituals.
Outcome
A launch email that reads like Maya, not like AI. The Project becomes her permanent "voice memory" — every future launch is faster than the last.
03 · Do
Workshop
Build your own Voice Project
Set up a Claude Project that produces drafts that sound like you, on demand.
Prerequisites
- A free or Pro Claude account at claude.ai
- Two or three pieces of writing you have published (blog posts, emails, anything 300+ words)
- 1
Create a new Project
Name it after a recurring writing job you do — "Weekly update", "Sales emails", "Blog drafts". Specificity matters; a Project for "writing in general" is too broad to be useful.
- 2
Upload your voice samples
Drop in 2–3 of your own pieces. Avoid mixing voices: if you co-author a newsletter, only upload sections you wrote.
- 3
Write a 5-line system prompt
Lines: 1) what to write, 2) tone, 3) sentence-length default, 4) banned words/phrases, 5) a stop condition like "If the brief is missing a CTA, ask before drafting."
- 4
Test with a real brief
Paste a real brief from your inbox. Ask for the answer "as an artifact." Iterate by editing the artifact, not by sending new chats.
- 5
Lock the prompt
Once a draft passes the "does this sound like me?" test, save the system prompt. Re-use the Project for every future piece of that type.
04 · Ship
Exercise — graded by Claude
Reverse-engineer a system prompt
You are handed three Claude replies that all sound terrible — too corporate, full of buzzwords, generic. Write the *system prompt* you would give Claude to fix them. Your prompt must address: (1) tone, (2) sentence-length defaults, (3) banned words, and (4) a stop condition for ambiguous briefs. Reply with just the prompt itself, no preamble.
Expected length: 5 to 10 lines. Pass score: ≥ 70.
Rubric (the grader uses this)
- 25%Specifies a clear role or persona. Open with "You are X" or equivalent.
- 25%Constrains tone and sentence length. Numbers help: "average 12 words per sentence" beats "be concise".
- 20%Bans at least three buzzwords explicitly. Examples: leverage, synergy, empower, unlock.
- 20%Includes a stop condition for ambiguity. Something like "Ask before drafting if the CTA is missing."
- 10%Reads like a prompt, not an essay. Imperative voice, no apologies, no preamble.