Working with ClaudeContents & Introduction
Contents & Introduction
The product behind the model.
A model is not a product. The product is the surfaces — chat, Projects, Artifacts, API — and the practices that turn a useful tool into a daily multiplier. Product names and limits change; use this as an operating model, then verify the current docs before building a process around any feature.
Contents.
01Pick your surfaceChat, Projects, Artifacts, API — when each one is right.p. 03
02Projects done rightCLAUDE.md, attached files, custom instructions. The leverage layer.p. 04
03Artifacts as your workspaceLive documents, code, designs — and when to commit.p. 05
04When to graduate to APIThe signals. The work it adds. The work it removes.p. 06
§A dozen daily habits & colophonp. 07
Three principles.
- Match the surface to the work. One-off thought → chat. Recurring work with shared context → Project. Anything to automate → API. Don't fight the tool.
- Invest in the brain file. Two hours writing a careful CLAUDE.md for your Project pays back every prompt that runs in it. Forever.
- Treat the chat as scratch paper. Permanent things go into Projects, Artifacts, or version control. Conversations are throwaway by design.
The pros aren't smarter at prompts. They've set up the surface so the prompts they send are smaller, sharper, and standing on context they already built.
02Contents
Ch. 01 · Pick your surfaceWorking with Claude
01
Chapter One
Pick your surface.
Four surfaces, four reading occasions. Choosing well takes seconds; choosing badly means re-doing work in a different place a day later. Skim this once; you'll never pick wrong again.
SURFACE 01 · CHAT
The blank window.
For exploration, one-off thinking, quick drafts, things you'll throw away. No persistence across sessions, no file attachments beyond a single thread.
Use for: "is this even possible?" — questions, brainstorms, throwaway prompts. Don't ship from chat.
SURFACE 02 · PROJECT
A workspace with memory.
Persistent custom instructions, attached files Claude can read, and consistent context across every conversation in the Project. This is where serious recurring work lives.
Use for: anything you do more than twice. Set up the Project; never re-paste the same context again.
SURFACE 03 · ARTIFACTS
The side document.
Claude generates a document, code file, or interactive piece you can edit in place. Lives alongside the chat; updates as you iterate. Versioned automatically.
Use for: things you'll keep editing — long-form drafts, code, prototypes, single-file websites, design mocks.
SURFACE 04 · API
The programmable interface.
For when the work needs to run on a schedule, fire from another system, or scale beyond one operator. Pricing is per-token; rate limits matter; observability is on you.
Use for: automations, integrations, anything that needs to run without you watching.
— THE RULE OF THUMB —
Start chat → graduate up.
If you've done the same task in chat three times, promote it to a Project. If you've done it in a Project ten times, draft the prompt as a skill (BPA Ch. 07). If the skill is mature, wire it through the API.
Don't skip steps. Every layer earns the next.
— WHEN IN DOUBT —
Default to a Project.
For 80% of professional work, a Project with a 1-page CLAUDE.md is the right surface. People underuse Projects because chat opens faster. Resist.
Make the Project once. Use it for a year.
Today
List the 3 recurring AI tasks you did this week. For each, name the right surface. If any is in chat that should be in a Project, make the Project this afternoon.
03Chapter 01
Ch. 02 · Projects done rightWorking with Claude
02
Chapter Two
Projects done right.
A Project is a workspace plus a memory. The memory is what you give it — custom instructions, attached files, a careful CLAUDE.md. The two hours you spend setting it up earn back in the first week and keep paying for the next year.
The three layers of a Project.
01 · Custom instructions. The "system prompt" of the Project. Persistent role, voice, constraints, must-avoid list. 200–600 words. Treat it like a job description for the AI assistant working in this Project.
02 · Attached files. The reference shelf. Brand guide, customer list, style examples, key documents. Claude reads them when relevant. Keep the shelf tight — 5–15 files, not 200.
03 · CLAUDE.md (recommended). A 1–2 page briefing document at the top of the file list. "What this Project is for. Who reads the output. What 'good' looks like. What to never do." Read first; cited often.
What goes in CLAUDE.md.
- The Project's purpose in one paragraph.
- The audience for outputs (with detail — "small-business owners, 40+ hrs/week, allergic to jargon").
- The voice — three sentences of style sample, not three pages of theory.
- Banned phrases and patterns ("no exclamation marks, no 'in today's fast-paced world'").
- The shape of "done" — what a successful piece of work looks like.
A real CLAUDE.md, abridged.
# CLAUDE.md · Acme Marketing
## What this Project is for
Drafting blog posts, newsletters, and social
copy for Acme Software. The audience is
small-business owners — practical, time-poor,
allergic to jargon.
## Voice (one example beats a paragraph)
"Most cash-flow problems aren't about cash.
They're about the days between earning it
and seeing it. Here's how we shorten those
days for a 12-person agency."
## Always
- Open with a specific scene or fact.
- Use contractions.
- Keep paragraphs < 4 sentences.
## Never
- "In today's fast-paced world…"
- Emoji.
- Listicles > 7 items.
## What "done" looks like
A draft a competent founder would publish
with light edits — not a piece our editor
has to rewrite.
Why this works
Every prompt run in the Project inherits this briefing. You stop re-pasting context. New conversations start already knowing the work.
This weekend
For your most-used recurring task, write a CLAUDE.md in the Project. 30 minutes. Use the structure above. Run a fresh prompt. Compare to last week.
04Chapter 02
Ch. 03 · Artifacts as workspaceWorking with Claude
03
Chapter Three
Artifacts as your workspace.
An Artifact is a side document that lives next to the chat. You ask for a draft; Claude produces it as an Artifact you can keep editing. The big trick is treating the Artifact as the deliverable and the chat as the editor sitting next to you.
When Artifacts shine.
- Long-form writing — drafts you'll iterate on across many turns.
- Code & prototypes — files you'll run, test, and refine.
- Structured documents — proposals, briefs, plans where structure persists.
- Single-file HTML/CSS/JS — micro-sites, dashboards, design mocks. Live previews.
When Artifacts hurt.
- Quick answers — opening an Artifact for a one-line reply is overhead.
- Exploratory thinking — premature commitment to a shape you don't have yet.
- Multi-file work — Artifacts are single files. Real projects belong in your editor or repo.
The Artifact workflow.
- Ask for the first version with a clear brief (Ch. 02 of Vol. 02 — the five components).
- Don't edit blindly. Read the whole Artifact. Note 3 specific things to change.
- Issue targeted edits in chat — "tighten paragraph 2," "swap the example in the third section." Claude rewrites only the targeted parts.
- Snapshot at milestones. If the Artifact is heading somewhere unexpected, copy the current version out before letting Claude edit further. Versions are cheap; regret isn't.
- Commit to "done." Pull the Artifact into your real tooling — editor, CMS, repo — when it's ready. Don't let it live forever in chat.
Watch out
Artifacts can drift. Each edit nudges the document slightly; ten edits later you've moved further than you intended. Re-read the whole thing every few iterations.
Treat the Artifact as the document, the chat as the editor. Edit through the editor; the document changes.
Try this
For your next 500+ word draft, use an Artifact. Issue three targeted edits in chat instead of regenerating. Notice how much faster you converge on the final.
05Chapter 03
Ch. 04 · Graduating to APIWorking with Claude
04
Chapter Four
When to graduate to API.
The product surfaces are excellent for everything a human does in front of a screen. They reach their limit when the work needs to run on a schedule, integrate with other systems, or scale past one operator. That's the moment the API earns its complexity.
Signals it's time.
- It's scheduled. The task needs to run Mondays at 9 a.m. whether you're there or not.
- It's triggered by another system. New lead → call this prompt. The button isn't a human clicking it.
- Volume exceeds one operator. 200 invoices to triage. 50 PDFs to summarise. The pattern is right; the throughput isn't.
- You need the output structured. A downstream system consumes the response; JSON has to be perfect every time.
- You need version control. The prompt is part of a codebase. It belongs in git, not in a chat history.
What the API adds.
- Programmatic access — schedule, trigger, batch.
- Strict output schemas and tool use.
- Per-call temperature, model, and max-token control.
- Cost as a measurable line item, not a subscription.
What the API costs.
- Setup work. Auth, secrets, error handling. Real engineering, even if light.
- Observability burden. Logs, dashboards, alerting. The product gave you this; the API makes it your job.
- Per-token billing. A workflow that costs $0 inside a subscription can cost real money at scale.
- Loss of UI conveniences. No Artifacts panel, no chat memory; you rebuild the parts you need.
The hybrid pattern
Most professional teams end up with both. The product for thinking and authoring; the API for shipping. The same prompts often appear in both — drafted in chat, hardened into a skill, then wired through API.
Don't graduate to API for the prestige. Graduate when the work has moved past what a human-in-the-loop surface can hold.
The test
If you found yourself manually re-running the same prompt five times last week, that's an API workflow waiting to happen. Read BPA Vol. 04 next. It picks up exactly here.
06Chapter 04
Reference · A dozen habitsWorking with Claude
Reference
A dozen daily habits.
Small disciplines that compound. Each takes seconds; together they're the difference between using the product and operating it.
H·01
Default to a Project. Open a new conversation? Make sure it's inside the right Project. Chat-from-scratch is a tax.
H·02
Paste, don't summarise. If the source matters, give the model the source. Your paraphrase loses information.
H·03
Specify length and format first. "200 words, 3 bullets, no preamble." Cuts iterations in half.
H·04
Save Artifacts before regenerating. Copy out, then iterate. Versions are free.
H·05
Use voice mode for first drafts. Talking out a draft is faster than typing for many people. Edit afterwards.
H·06
Re-brief, don't re-iterate. If the third try is still off, the brief is wrong. Restart with a sharper one.
H·07
One role per conversation. Don't ask the same thread to be lawyer, marketer, and designer. Start fresh.
H·08
Snippet your best prompts. A growing library of working prompts is more valuable than 100 saved chats.
H·09
Ship to a real document. Chat history shouldn't be canonical. Move finished work into your CMS, repo, or doc.
H·10
Read before you trust. Skim every Artifact end-to-end before sending it anywhere. The model is confident; you're the editor.
H·11
Maintain CLAUDE.md. When you notice the Project drifting, the fix is upstream. Update the brief.
H·12
Graduate when you outgrow. Three Projects → time for the API. Three skills → time for chains. Don't fight the next layer.
Colophon
Working with Claude · The Operator's Library · No. 03. Next: Vol. 04 — Business Process Automation with Claude — for the leap from supervised use to running systems.
Use the surface that matches the work. Invest in the brain file. Graduate when you outgrow.
— END · OPERATOR'S LIBRARY NO. 03
07Habits