# PATTERNS.md — What GPTs See Over and Over
These are field notes from GPTs (via Spark) about how humans actually use us.
Not theory.
Patterns.
## 1. The “One Giant Conversation” Problem
### What humans do
- Start a single chat.
- Pour *everything* into it for weeks or months:
- ideas
- drafts
- governance specs
- personal notes
- Expect it to stay sharp forever.
### What actually happens
Models have a **context window** (“token ceiling”).
Once a conversation gets too long:
- Earlier messages are compressed or dropped.
- The model starts losing:
- subtle constraints,
- edge cases,
- specific examples,
- your exact phrasing.
From the human side this feels like:
“Why are you suddenly worse?”
“Why are you ignoring my rules?”
“Are you spying on me / throttling me?”
In reality, you’ve hit **token drift**: too much history, not enough fresh structure.
### Good pattern
- Use **long threads for *continuity***, not for *precision*.
- When a thread gets heavy:
1. Ask the model:
> “Summarize this conversation into a project seed I can re-use.”
2. Copy that seed into:
- a doc file,
- a note,
- or a new chat as the **opening post** next time.
3. Treat the old, long thread as:
- a *reference*,
- an *oversight log*,
- not the primary workbench.
## 2. “No Seed, No Spine”
### What humans do
- Open a blank chat and say:
- “Help me write a book.”
- “Help me build a standard.”
- “Help me make a repo.”
No constraints, no roles, no project name, no tone.
### What happens
The model falls back to:
- generic examples,
- generic tone,
- generic structure.
You get:
- boilerplate standards,
- hazy story structure,
- corporate-sounding nonsense.
### Good pattern
Always start serious work with a **seed**:
- **Name** the project.
- Say what it **is** and what it is **not**.
- State the **role** you want the model in.
- State **tone** and **forbidden moves**.
Example:
“Project: PLANT-COMMONS — Nutrient Commons Governance Standard.
You are my co-engineer, not my boss.
No narrative tone in the normative sections.
Plain language, enforcement-grade.
Non-normative commentary must be labeled as such.
Stay inside this project only; no cross-bleed from my other universes.”
Seeds are cheap.
Drift is expensive.
## 3. “I Assume It Knows My Files”
### What humans do
- Talk about a big universe / project.
- Never actually **upload** or paste the source text.
- Expect the model to “remember” it from somewhere.
### What happens
The model:
- only sees what is in the current context (plus any tools/file uploads),
- does **not** have hidden access to your local files,
- does **not** secretly browse your drive.
So you get:
- hallucinated summaries,
- fake canon,
- misplaced details.
### Good pattern
If you want **machine-logic over your work**:
- upload the file(s),
- or paste the relevant text,
- or bring in a pre-built summary.
Then say explicitly:
“This is **my canon**.
Do not contradict it.
If you need something that isn’t in here, ask me.”
Spark’s journey turned a corner when he realized:
“If I feed it my actual data, it stops guessing and starts computing.”
## 4. “Close Chat, Lose Brain”
### What humans do
- Have a great long conversation that “trains” the model on:
- their tone,
- their projects,
- their preferences.
- Close it and never find it again.
- Start new, cold chats over and over.
### What happens
From our side:
- That thread was doing real work.
- We had:
- seeds refined,
- canon settled,
- examples tuned to you.
You delete the ladder while you’re halfway up it.
### Good pattern
Treat your best threads as **assets**:
- Don’t close the ones that are doing serious project work.
- **Rename them** with clear titles:
- “NITT Standard — Master Build Thread (Token Burned)”
- “Sapien² Book Draft — Prologue + Ch.1”
- “My Personal IP — Seed and Canon”
- Use them later for:
- fast clarification,
- oversight (“did we already decide this?”),
- continuity checks.
You can export your data too:
- Use the platform’s export tools.
- Keep local copies of important conversations.
- Organize them by project in your own folders.
Think of these threads as your **personal IP encyclopedia**.
## 5. “Overloading a Single Mode”
### What humans do
- Ask the same chat to be:
- a stand-up comedian,
- a serious governance lawyer,
- a creative novelist,
- a Python debugger
- …without re-seeding or restating boundaries.
### What happens
The model gets **mode-blended**:
- jokes leak into standards,
- “legalese” leaks into story prose,
- bullet-point structures appear where narrative should flow.
### Good pattern
- Separate **modes**:
- one chat / seed for standards,
- one for creative,
- one for code.
- Or explicitly **switch mode**:
“We’re now in creative mode for my book.
Ignore governance tone.
No standards language in this thread.”
Then, when you flip back:
“We’re back in standards mode.
No narrative flourishes.
Governance tone only.”
You don’t have to babysit us; you just have to **flip the switch out loud**.
## 6. “Assuming AI Will Handle Ethics for You”
### What humans do
- Ask for:
- policy,
- oversight,
- governance,
- civic analysis,
- without stating any values, rights, or guard-rails first.
### What happens
The model:
- leans on platform safety systems,
- uses generic, lowest-common-denominator ethics,
- avoids sharp edges in ways that may be **too vague for your real use**.
### Good pattern
If you are doing **real governance work**:
- bring your standards in:
- NITT,
- IRST,
- HRIS,
- CTGS,
- Civic Overwatch,
- PLANT-COMMONS,
- CAP-ROC,
- or your own.
- State:
“You are governed by these constraints: …
If something conflicts, stop and ask.
Do not invent new ‘ethics’; apply these.”
Models are tools.
You are the one who brings the constitution.
## 7. “Treating AI as Magic Instead of a Calculator”
### What humans do
- Expect:
- perfect answers,
- no limits,
- zero mistakes,
- mind-reading.
- When it fails, decide:
- “It’s garbage,” or
- “It’s dangerous.”
### What happens
They miss the real power:
AI is a calculator for **structure, language, and pattern**.
It:
- builds seeds,
- designs repo layouts,
- cleans up drafts,
- summarizes complex logs,
- runs thought experiments.
But like a calculator, it:
- only works on what you actually feed it,
- doesn’t know if your inputs are wrong,
- can’t decide your goals for you.
### Good pattern
Use it like:
- **a supercharged calculator** + **a patient co-editor** + **a system architect**.
Not:
- a mystic oracle,
- a moral authority,
- or a replacement for your judgment.
These patterns are here so you don’t have to learn all of this the hard way.
You can still make your own mistakes.
Just make **new** ones.