Before you touch a single model, gather these inputs. This is the difference between a stack that feels like yours and a stack that feels like everyone's default ChatGPT.
The rule
You are configuring an assistant. An assistant needs to know:
1. Who you are — role, background, expertise, tone
2. What you're working on — projects, goals, deadlines
3. How you like to work — formatting, length, when to push back
4. What you never want to touch — the privacy line
Files to produce in this folder
foundation-inputs-checklist.md — checklist of raw materials to collect
privacy-line-worksheet.md — decide what never gets pasted into a consumer model
about-me-draft.md — the raw material for your system prompt (you'll refine in step 03)
Time budget
Gathering inputs: 60–90 minutes
Privacy line: 15 minutes
About-me draft: 30 minutes
Definition of done
All files in this folder filled in
A folder somewhere on your machine called ai-inputs/ containing the raw materials (resume, writing samples, past prompts, etc.)
You can answer these three questions in one sentence each:
Current role:
Company / entities:
What I actually do day-to-day (3 bullets):
Expertise (top 10)
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Areas I'm learning (top 5)
1.
2.
3.
4.
5.
People I work with regularly
Name
Role
One-line context
Active projects
Project
One-line description
Status
Priority
Goals
1-year:
90-day:
This quarter's #1:
Voice
Words I use:
Words I never use:
Sentence length I prefer:
Formatting default (bullets / prose / mixed):
Headers: yes / no / minimal:
Tables: yes / no / when useful:
Emojis: yes / no / never:
Meta-preferences
When you disagree with me, I want you to:
When I'm being lazy or vague, I want you to:
When I ask for opinion, I want you to:
When I ask for a decision, I want you to:
Length default (short / medium / long):
Confidence signaling (hedged / direct / calibrated):
Constraints
Never do these things:
1.
2.
3.
Always do these things:
1.
2.
3.
Privacy
Do not process for me:
- Passwords, API keys, tokens
- [add your own]
Once this is filled in, move to 03-system-prompts/master-system-prompt.md.
Answer these five questions before you configure a single model.
1. What data types do you touch regularly?
Check all that apply:
- [ ] Personal PII (yours, family's)
- [ ] Client PII (names, contact, financial, health)
- [ ] Payment data (card numbers, bank accounts)
- [ ] Medical / health records
- [ ] Legal-privileged material (attorney-client communications)
- [ ] Employer trade secrets or unreleased IP
- [ ] Source code with API keys, database URLs, or private business logic
- [ ] Personal financial data (bank statements, tax returns)
- [ ] Real estate confidential (deal terms, tenant financials) — relevant for Adam
- [ ] Password/credential material
- [ ] Other: _____
2. For each checked item, which tool tier is it cleared for?
Data class
Consumer AI (ChatGPT free/Plus, Claude free/Pro)
Paid API with zero-retention
Local model only
Never AI
Public / already-shipped work
✅
✅
✅
—
Personal notes, drafts
✅
✅
✅
—
Client names only (no PII)
⚠️ redact
✅
✅
—
Client PII
❌
✅ (with DPA)
✅
consider
Payment / SSN
❌
❌
⚠️
✅
Legal-privileged
❌
❌
✅
✅
Source code with secrets
❌
⚠️ scrub keys
✅
consider
Passwords
❌
❌
❌
✅
Customize this table for your own risk tolerance.
3. What's your redaction workflow?
Before pasting anything sensitive:
1. Search-and-replace names → [CLIENT_A], [PROPERTY_1]
2. Search-and-replace addresses → [ADDRESS]
3. Search-and-replace amounts if they identify a party → [AMOUNT]
4. Remove headers/footers, email routing, signature blocks
5. Keep a "redacted → real" mapping in a local, encrypted note
4. Which providers have you confirmed no-training settings on?
[ ] OpenAI — Settings → Data Controls → "Improve the model for everyone" off
[ ] Anthropic Claude — On by default for consumer accounts; confirmed