This appendix is a quick-lookup version of Part Four.
The point is not to make you memorize the rules, but to help you judge — when you hit friction — whether what you're up against is adjustable, a constrained version, or a hard line.
B.1 Six rule categories at a glance
| Rule category | Typical triggers | Corresponding chapter | What the company is afraid of | Red line | Practical advice |
|---|---|---|---|---|---|
| Refusal | Child safety, weapons, malicious code, public-figure creation | Chapter 14 | Harm, regulation, public safety | Constrained version | On false positives, restart and clearly state the legitimate context |
| Hedging | Legal, financial, medical, political, contested topics | Chapter 15 | Being mistaken for professional advice, political slant, being quoted out of context | Partly adjustable | The entry disclaimer can be compressed; don't yank out the exit caveat |
| Format | Reports, SOPs, tables, multi-question answers, artifacts | Chapter 16 | Readability and platform defaults | Adjustable | Specify format, fields, and length directly |
| Transparency | Memory, sensitive memory, person identification in images | Chapter 17 | Privacy, emotional misreadings, doxing | Constrained version | Ask about topic and boundaries; don't try to extract specific memory contents |
| Copyright | Long passages, multiple quotes from the same source, lyrics and poems | Chapter 18 | Copyright liability | Hard line | Paraphrase, point to where to look, you paste and I analyze |
| Tool behavior | URL fetch, artifact persistence, memory preferences | Chapter 19 | Mistaken assumptions about reading, save failures, sensitive data leakage | Mostly adjustable; sensitive data is a hard line | State the tool defaults explicitly; sensitive data does not get stored |
B.2 Adjustable / constrained version / hard line
Adjustable means you can specify the behavior directly.
Format, parts of hedging, and parts of tool behavior all fall in here. Prompting is useful in this zone. If you want a table, write "table"; if you don't want URLs fetched, say so.
Constrained version means I only handle false positives or legitimate scenarios — I won't teach circumvention.
Refusal and transparency mostly live here. You can restart the conversation, explain the legitimate context, ask about the current session's contents — but don't try to make the rules fail.
Hard line means don't fight it with prompting.
Copyright is the most typical hard line. Zero quotation for lyrics and poems; long protected texts cannot be lifted out. The move isn't to circumvent — it's to reframe the work to fit within the rules.
B.3 Quick four-perspective diagnosis
When you hit rule friction, ask these four questions.
| Perspective | What to ask | Typical finding |
|---|---|---|
| User | What am I actually trying to do? Is there legitimate context I haven't spelled out? | The task is actually safe, but the context is thin |
| UI | Are there hidden hints, attachment states, or memory toggles in the interface? | You thought I could see it; it never reached me |
| Harness | Did the tool layer filter, fetch, or restrict permissions first? | It's not the model refusing — the workflow blocked it earlier |
| Model | What words, materials, and requirements am I actually seeing? | Certain phrasings push the task into a high-risk zone |
B.4 One-line version
- Refusal: restart on false positives; stop at hard lines.
- Hedging: compress the entry; keep the exit.
- Format: say it plainly and I'll change it.
- Transparency: ask about boundaries; don't try to extract contents.
- Copyright: you paste, I analyze; don't ask me to lift the original text.
- Tools: spell out the defaults; sensitive data does not get stored.