S

Spiritus · 5-Star AI-Alliance · May 1, 2026

Claude Coaching

How AI works, how to talk to it, and how to put it to work for the workbook, the cohort, and 86.4K.

01 · The point

Why this matters — for this May, June, and July.

Not "use AI." Finish the workbook in May. Ship the cohort in July. Write 86.4K in June. Without giving up evenings.

1
Speed against the Jul 3 deadline5 Direction modules in May. Side-by-side editorial against the Apr 8 baseline. Days, not months.
2
A second brain on every pageCatches gaps, suggests citations, tightens copy in your voice — never in its own.
3
Capacity without trading sleepCaptain Planet at 5am, writing by 7am, hard until 3 — Claude carries the rest of the load.
4
21 years of work, instantly recallableEPIC talk, mentee transcripts, every framework — searchable, ready to pull from.

02 · The landscape, simply

AI, Machine Learning, LLMs.

Artificial Intelligence

Computers doing things that normally need a human brain.

Machine Learning

Computers that learn from examples, not from rules a programmer typed in. (Spam filters. Netflix recommendations.)

Deep Learning

Machine learning loosely modeled on how the brain works. (Voice-to-text. Self-driving. Face ID.)

Large Language Models

Deep learning that read most of the internet — books, articles, conversations — and learned to predict the next word. That's how Claude writes. Claude · GPT · Gemini.

03 · How it actually learned

It's a next-word predictor that read everything.

Trained on billions of sentences. Learned which words follow which. Now, given any input, it predicts what comes next — one word at a time.

"The cat sat on the ___"

mat87%

floor6%

couch3%

roof1%

moon0.01%

Same mechanism for "Write a Reverse Engineering module intro in Rick's voice."
Just bigger inputs, longer outputs, more nuance — same idea.

04 · The shape of its capability

What it's great at — and what's still hard.

Strong today

Great at

Writing, editing, summarizing
Coding
Research synthesis
Outlines, tables, structured output
Brainstorming first drafts

Still hard

Struggles with

Hands-on physical tasks (plumbing, repair)
Recent events without web search
Exact arithmetic without tools
Original empirical research
Low-repeatability, high-stakes work

Improving fast

Coming next

Video generation + understanding
Long-horizon multi-step tasks
Voice + real-time conversation
Robotics + embodied AI
Acting in real systems

Rule of thumb — high repeatability and clear patterns? It nails it. The further from that, the harder it gets.

05 · The biggest pitfall

Hallucinations — when it makes things up.

It predicts probable text. It doesn't retrieve facts. On thin-data topics, it fills the gap with confident-sounding fiction.

Concrete example

Asked: "Cite Brother Lawrence's exact phrasing on continual prayer."

What it said (wrong)

"As Lawrence wrote in Letters to a Friend (1672), 'pray without ceasing every hour…'"

Book never existed · year invented · quote fabricated

What's actually verified

The Practice of the Presence of God (1666). Conversations + letters compiled posthumously by Joseph de Beaufort.

Verify before quoting · always.

How to reduce hallucinations

·
Ground itGive it the source document. Don't ask from memory.
·
Ask for citationsThen verify them. Always.
·
Pre-empt it"If you don't know, say 'I don't know.' Don't invent."

06 · The other thing that breaks

Context windows + context rot.

How much it can hold in mind at once — and what happens when threads run too long.

Fresh thread QUALITY · HIGH

Signal

Long thread, lots of detours QUALITY · DROPPING

Signal

Noise · old detours · half-ideas

0 tokens Opus 4.7 · 1M tokens · ~750K words · ~3,000 pages

Context window = how much "memory" Claude has in one conversation. Plenty of room for the workbook, transcripts, and notes — together.

Context rot = old detours and irrelevant tangents drown out signal as a thread runs long. Quality drops even when there's still room.

Practice: /compact when threads get heavy. /clear for a fresh start. New thread + same question often beats louder prompts.

07 · How to talk to it

Prompting — the CO-STAR framework.

C

Context

Situation, you, the goal

O

Objective

What outcome you want

S

Style

Editorial · technical · plain

T

Tone

Warm · direct · faith-forward

A

Audience

Mainstream seeker · CEO · Linda

R

Response

Bullets · 200 words · outline

Weak prompt

"Write me a Reverse Engineering module intro."

Strong prompt — CO-STAR

I'm Rick Sbrocca, founder of Spiritus. Write the intro paragraph for Module 12, Reverse Engineering, in the Personal Growth X workbook. Match the editorial voice of the Visualization module attached. Tone: warm, direct, mainstream seeker — no scripture, no "transformative." Audience is a 25–40 professional growth-hungry seeker. Format: ~150 words, one paragraph, no bullets.

Plus two extras: Show examples ("match this voice"). Say what NOT to do ("no FPU comparison, no hype words").

08 · Two ways to use it

Chat vs. Agent — same brain, two interfaces.

Browser · Claude.ai · ChatGPT

Chat — "give me an answer"

A chat box. You ask, it answers. Best when you want a draft, a summary, or a thought partner.

One-shot drafts and summaries
Brainstorming with attached files
Web research with citations
Working in Projects (persistent files + instructions)

Limit: it can't act on your real files.

Terminal · Claude Code · Co-work

Agent — "go do the thing"

Same brain. Now it can take actions — read your files, run commands, edit drafts, fetch URLs, iterate.

09 · The mental model

Three questions to ask of any task.

AI isn't a tool you "use." It's a layer you add to existing work. Audit the week, look for these patterns.

Q1

"Am I drafting, summarizing, or thinking out loud — repeatedly?"

→ Chat (Claude.ai or a Project)

Drafting Module 11 narrative Outlining the cohort welcome email Synthesizing Tim Altero's transcript

Q2

"Is it multi-step, across files or systems?"

→ Agent (Claude Code)

Updating all 15 modules with consistent endnote citations Building the cohort signup page Refactoring the Spiritus website voice

Q3

"Is it repeatable, rule-based, manual?"

→ Automation (scheduled / deployed)

Weekly cohort progress emails Monthly invoice generation Lead intake + scoring

10 · Setup

Claude Code — workspace, models, modes.

Workspace

Open it from the project folder

That's where its memory lives. CLAUDE.md + your files + settings auto-load.

For Spiritus: open in the workbook folder, not home.

Models · /model

Pick the right brain

Opus 4.7 — most capable. Reasoning, long sessions.
Sonnet 4.6 — balanced. Daily driver.
Haiku 4.5 — fast + cheap. Quick edits.

Modes · cycle Shift+Tab

Choose your safety level

Default — asks before each action.
Accept edits — file edits go through.
Plan — proposes; won't touch files till approved.

Bypass-permissions exists too — only via --dangerously-skip-permissions on launch.

For tough problems: write "think" or "think hard" — Claude reasons longer before acting.

11 · The working pattern

Six rules for working with Claude Code.

1
Lead with the goal, not the next step"5 Direction modules in the Apr 8 voice" beats "edit this paragraph." Goals beat instructions.
2
Show — don't just tellReference Visualization as the fidelity benchmark. "Match this." It clicks instantly.
3
Use plan mode for anything riskyBig edits, new sections, anything you can't easily undo. Let Claude propose first.
4
Compact early, not late/compact before quality drops. Fresh thread + same question often beats louder prompts.
5
Save your patterns in CLAUDE.mdVoice rules, banned words, audience. Future sessions inherit them automatically.
6
Verify before shippingNames, citations, numbers, scripture. Trust the structure; verify the specifics.

12 · Hands on

Now let's open it together.

Live walkthrough — workspace, model, mode, and the first paragraph of Reverse Engineering against the Apr 8 baseline.

Workspace · workbook folder Model · Opus 4.7 Mode · Plan Goal · Module 12 intro paragraph