The AI Compass
You're in the free realm. Enjoying it? Unlock the other six realms + the Memory Vault →
C
Realm 1 of 7 · The COMPASS

Conversations

Master the chat itself — how to ask, iterate, and get answers worth acting on.

Mnemonic: Great conversations spell PROMPTS — seven topics, seven letters
0 / 7
explored

Everything in Claude.AI begins in a conversation. This realm teaches you the anatomy of a chat and the craft of prompting — the single highest-leverage skill on the platform. Its seven topics spell PROMPTS, because that is literally what they build.

Choose any topic — constructivist learning means you pick the path. Each opens into three depths: Learn the idea, Try it now in your real business, and Go deeper for mastery.

Claude performs best when you write to it the way you'd brief a brilliant new hire: state what you want, why you want it, and what "good" looks like. Vague prompt: write me a sales email. Precise prompt: Write a 120-word sales email to gym owners introducing my $497 onboarding course. Warm but direct tone. End with one clear call to action.

Three ingredients of precision: the task (what to produce), the context (audience, situation, constraints), and the success criteria (length, tone, format, must-include points).

Memory aid: think of every prompt as a work order, not a wish. A wish says "make it good." A work order says exactly what "done" means.
Constructivist mission — build the knowledge by doing
  1. Open a new chat in Claude.AI.
  2. Take a task you did this week and write it as a one-line vague prompt. Send it.
  3. Now rewrite it as a work order: task + context + success criteria. Send that.
  4. Compare the two answers side by side.

Reflect: Which specific detail in your second prompt changed the output the most?

Pro-level precision techniques:

  • Assign a role: "Act as a direct-response copywriter reviewing this landing page" primes better instincts than a bare request.
  • State the negative space: what to avoid ("no buzzwords, no exclamation marks") is as powerful as what to include.
  • Give the downstream use: "this will be pasted into a cold email tool" changes formatting decisions Claude makes silently.
  • One prompt, one job: if a request has three deliverables, either number them explicitly or split into three messages.

Entrepreneurs lose more time to under-specified prompts than to any limitation of the model. Precision is free leverage.

Claude doesn't know your business unless you tell it (or teach it via memory and Projects — covered in later realms). At the conversation level, two sentences of context transform everything: who you are, and who Claude should act as.

Example opener: I run a one-person publishing company producing non-fiction books and online courses. Act as my developmental editor with 20 years in trade publishing. Every answer after that line is filtered through the right lens.

Memory aid: the two-badge rule — pin a name badge on yourself and one on Claude before the meeting starts.
Constructivist mission — build the knowledge by doing
  1. Write a two-sentence 'badge pair' for your most common task: one sentence about you, one assigning Claude a role.
  2. Save it in a note on your device — this becomes a reusable opener.
  3. Start a chat with it and ask a question you asked Claude last week, without the badges. Compare.

Reflect: What expertise do you keep wishing you could hire? That's the role to assign.

Advanced role work:

  • Panel prompting: "Answer as three advisors — a CFO, a brand strategist, and a skeptical customer — each in turn." You get built-in debate instead of one voice.
  • Calibrated seniority: "junior assistant" produces cautious, option-heavy answers; "seasoned operator" produces committed recommendations. Choose deliberately.
  • Standing context: once you find badge pairs that work, promote them into a Project's custom instructions (Organization realm) or your user preferences (Memory realm) so you never type them again.

Claude will match almost any output shape you name: a table, a numbered checklist, a 60-second script, a JSON object, an email with a subject line, plain prose with no bullet points. If you don't specify, Claude guesses — and you spend time reshaping.

Useful levers: length ("under 150 words", "one page"), structure ("a table with columns X, Y, Z", "exactly five options"), register ("plain language a 12-year-old could follow", "board-ready"), and medium ("formatted to paste into LinkedIn").

Memory aid: L·S·R·M — Length, Structure, Register, Medium. Four dials on the output machine.
Constructivist mission — build the knowledge by doing
  1. Ask Claude for advice on any business question, with no format instructions.
  2. Now re-ask, adding all four dials: a length, a structure, a register, and a medium.
  3. Try one exotic shape: 'as a two-column pros/cons table' or 'as a 30-second voicemail script'.

Reflect: Which dial do you leave unset most often — and what does that cost you per week?

Power-user shaping:

  • Show, don't describe: paste one example of a past output you loved and say "match this shape exactly." An example beats three paragraphs of description.
  • Tagged sections: for complex outputs, ask Claude to wrap parts in labels — "put the headline options under HEADLINES and the body under BODY" — so you can grab pieces cleanly.
  • Reusable format specs: keep a note of your three most-used format blocks (e.g., your proposal skeleton) and paste them in. Later, bake them into a Project or a Style.

The model picker (top of every chat) lets you choose which Claude model answers. The trade-off is simple: faster, lighter models for everyday drafting and quick questions; the most capable models for complex analysis, long documents, strategy, and code. You can switch models mid-conversation.

On paid plans you can also enable extended thinking, which lets Claude reason step-by-step before answering — dramatically better for math, planning, legal-style reasoning, and anything with many moving parts.

Memory aid: bicycle vs. freight train. The bicycle (fast model) for errands; the freight train (top model + thinking) when you're hauling something heavy.
Constructivist mission — build the knowledge by doing
  1. Find the model picker in a new chat and note which models your plan offers.
  2. Ask a heavyweight question (e.g., 'stress-test my pricing strategy') on the most capable model with extended thinking on if available.
  3. Ask a lightweight question on a faster model. Feel the difference in speed and depth.

Reflect: Which of this week's tasks were freight-train jobs that you gave to the bicycle?

Economics of model choice: usage limits on every plan are consumed faster by bigger models and longer conversations. A working pattern for solopreneurs:

  • Do exploratory back-and-forth on a fast model.
  • When the shape of the task is clear, switch to the top model for the final, high-stakes pass.
  • Keep conversations focused — a fresh chat with a tight brief often outperforms message #60 of a sprawling one, and costs less of your limit.

Model names change over time; the picker always shows what's current, and Anthropic's release notes explain what each is best at.

Treat Claude's first response as the opening of a negotiation. You can: follow up ("shorter, punchier, cut point 3"), ask for variations ("give me three more takes, one contrarian"), challenge it ("what's the weakest part of this plan?"), or retry the same message for a fresh attempt.

The secret weapon is editing your earlier message: hover over any message you sent, edit it, and the conversation branches from that point — a new timeline with new answers, without losing the old one. Perfect for exploring two directions of the same idea.

Memory aid: sculptor's rule — the first response is the block of marble. Iteration is the chisel.
Constructivist mission — build the knowledge by doing
  1. Ask Claude to draft anything (a bio, an offer, a headline).
  2. Send three chisel strikes in a row: one cut ('remove X'), one push ('bolder'), one challenge ('argue against this').
  3. Then edit your original message to change the brief, and watch the conversation branch.

Reflect: Do you usually accept draft one? What would a three-strike habit change?

Iteration patterns worth memorizing:

  • Critic pass: "Before I use this, review it as a harsh editor and list every weakness." Then: "Now fix everything you found."
  • Rubric pass: give Claude a scoring rubric ("clarity, specificity, proof, call-to-action — score /10 each") and ask it to score and revise its own work.
  • Branch-and-compare: edit the original prompt into two different strategic directions, then paste both results into a new chat and ask Claude to synthesize the best of each.

The paperclip / plus button attaches files directly to a message: PDFs, Word documents, spreadsheets, CSVs, images, screenshots, code files. Claude reads them and works on your actual material — summarizing a contract, critiquing a landing-page screenshot, cleaning a messy spreadsheet, answering questions about a 100-page report.

Images are first-class: Claude can read text in screenshots, interpret charts, describe photos, and review designs. On mobile, you can snap a photo straight into the chat.

Memory aid: show-and-tell beats tell. If it exists as a file, attach it instead of describing it.
Constructivist mission — build the knowledge by doing
  1. Attach a real business document — a proposal, contract, or report — and ask for a one-paragraph summary plus three risks.
  2. Screenshot your own website's homepage and ask for a conversion-focused critique.
  3. Attach a spreadsheet (or CSV export) and ask three questions about the numbers.

Reflect: What document have you been meaning to read carefully for weeks? Attach it today.

File-handling craft:

  • Ask before, not after: put your questions in the same message as the attachment, so Claude reads with purpose.
  • Multiple files: attach several and ask comparative questions — "which of these three proposals has the strongest scope section?"
  • Long documents: for very long material, ask for a structured map first ("outline this with section headings"), then drill into sections.
  • Recurring reference files (brand guide, price list) belong in a Project's knowledge base — covered in the Organization realm — so they're available in every chat automatically.

Claude lives on iOS and Android with the same account, history, Projects, and memory as the web — start a chat at your desk, continue it in line at the bank. The mobile apps support voice input: tap the microphone, talk through your idea, and Claude transcribes and responds.

Voice changes what you use Claude for. Spoken briefs are naturally richer in context than typed ones — rambling is a feature. Dictate the whole messy situation, then ask Claude to structure it.

Memory aid: walking meetings. Claude on mobile is the colleague who joins your walk and takes perfect minutes.
Constructivist mission — build the knowledge by doing
  1. Install the Claude app and sign in to your account.
  2. On a walk or commute, dictate a 2-minute brain-dump about a current business problem.
  3. Ask: 'Structure everything I just said into: situation, options, recommendation, next actions.'

Reflect: Which recurring thinking time (commute, gym, dishes) could become Claude time?

Mobile-first workflows for founders:

  • Photo capture: whiteboards, receipts, book pages, competitor shelf displays — photograph and interrogate on the spot.
  • Meeting debriefs: straight after a client call, dictate everything you remember; ask for a follow-up email draft and a task list.
  • Idea inbox: keep one ongoing chat as your capture stream, then weekly ask Claude to sort it into projects, discards, and this-week actions.
  • Agentic tools like Claude Code and Cowork (Agents realm) can also be monitored from the mobile app — kick off work at your desk, check it from anywhere.
Memory device

The realm in one word: PROMPTS

Close your eyes and spell it. If you can recover all seven keywords from the letters, this realm is yours.

PPrecisionBe clear, be specific, be generous with context
RRoles & contextTell Claude who you are and who it should be
OOutput shapingSpecify format, length, and structure
MModels & thinkingChoose the right model and thinking depth
PPush back & iterateFollow up, edit, retry, and branch
TTransfer filesUpload documents, images, and data into the chat
SSpeakVoice, dictation, and Claude on mobile
← Back to
The Atlas
Stay on course

Liked this realm? There's more where it came from

Opt in for occasional dispatches on learning AI — practical techniques you can use the same day, plus early word on atlas updates.

Sent by The AI Compass · learnaicompass.com. We only email when there's something worth learning — never more than a few times a month, never sold or shared.