6 APPENDIX: Patterns at a Glance

A distilled, one-line-per-pattern index of all fifty patterns in this book, grouped by family. For the full write-up of any pattern — the motivating scenario, Apply, Benefit, Guardrail, Related — read the full entry in its tier section; this is the index, not the text.

Selection is a judgment call. This is a recognition aid, not an auto-apply library — match the pattern to the actual problem in front of you; don’t reach for one just because it’s here. Read it the way you’d consult a table of contents, not the way a script executes a lookup. When unsure, read the full pattern in its section before applying it.

6.1 ONR — On-Ramps (Beginner)

[ONR-01] The Interview Primitive — reach for this when you’re about to ask for something important and haven’t given the model enough of your actual situation to work with.
[ONR-02] The Adversarial Framing Shift — reach for this when you want your draft critiqued, not approved, and suspect a direct “does this look okay?” will just get flattery back.
[ONR-03] The Blind-Spot Inquiry — reach for this before a high-stakes meeting or process, to surface the question you didn’t know to ask.
[ONR-04] Direct Grounding: Don’t Explain — Show — reach for this whenever you’re about to describe a document from memory instead of just showing it the original.
[ONR-05] Multi-Modal Equity: Talk to the Tool — reach for this when typing itself, not the technology, is the real barrier for someone.

6.2 PRAC — Practices (Intermediate–Advanced)

[PRAC-01] Both of You Can Be Wrong — reach for this to set the working relationship up right from day one: both you and the model can be wrong, so both of you check.
[PRAC-02] Active Interrogation & Steering — reach for this the moment an answer feels off mid-conversation, instead of waiting to catch it later.
[PRAC-03] Canonical-File Discipline — reach for this whenever more than one copy of a working file could plausibly exist.
[PRAC-04] Structured Elicitation Before Action — reach for this before a real, consequential piece of work begins, not just a casual first question.
[PRAC-05] Explicit Source-Grounding — reach for this whenever a number or fact matters and shouldn’t be recalled from memory alone.
[PRAC-06] The Decoupled Reference Audit — reach for this after any structural change, to check the result against a copy nobody touched.
[PRAC-07] Pin Down Its Role So It Stops Drifting — reach for this when you need a role to stay stable across sessions instead of quietly drifting.
[PRAC-08] The Disconfirmation Gate — reach for this before committing to a real decision, to make the model argue it’s wrong first.
[PRAC-09] The Unrun Check — reach for this when citing a figure whose verification route was closed — to distinguish a computed result from an inherited number that was never checked.
[PRAC-10] The Negative Must Be Earned — reach for this when a model reports that nothing is there, before treating absence as a clean result.
[PRAC-11] Reader vs. Machine: Choosing the Output Format — reach for this whenever output is headed to another tool or thread, not just a human reader.

6.3 ARC — Architecture (Intermediate–Advanced)

[ARC-01] The Swap Disk Protocol: Give the AI a Permanent Notebook ★ — reach for this the moment a project needs to survive past one conversation.
[ARC-02] Self-Installing Guards — reach for this to stop the same class of mistake from recurring, instead of just fixing today’s instance.
[ARC-03] Have a Different AI Attack the First One’s Work — reach for this when a draft needs a genuinely different set of blind spots looking at it.
[ARC-04] The Single-Writer Reconciliation Pattern — reach for this when more than one thread (or person) might edit the same working file.
[ARC-05] Archived Versions + Embedded Changelog — reach for this so a bad edit is always recoverable, not just the current state.
[ARC-06] Structured Handoff Briefs — reach for this when a fresh thread or a human reviewer needs to get productive fast, without re-deriving context from a raw transcript.
[ARC-07] Recheck the Record Against the Original — reach for this periodically, to catch a wrong-but-plausible value that internal checks alone would never see.
[ARC-08] A Copy You Edit Isn’t a Backup — reach for this when deciding whether a duplicate file is a backup or a liability.
[ARC-09] Resolve Standing Instructions by Rule, Not by Name — reach for this instead of hard-coding a specific filename into a standing instruction.

6.4 SCF — Foundation & Lifecycle (mostly Intermediate)

[SCF-01] One Thread Per Topic, Like a Standing Specialist — reach for this to keep one thread per domain, rather than letting contexts bleed together.
[SCF-02] Match the Model to the Job, Not Habit — reach for this instead of defaulting to the same model out of habit for every task.
[SCF-03] Don’t Trust What It Says About Itself — reach for this instead of trusting a model’s own claim about how full its context window is.
[SCF-04] Communication Style Adjustment — reach for this to get the register you actually need, instead of the model’s default helpful-generalist tone.
[SCF-05] Asymmetric Resource Substitution — reach for this before spending a scarce image-upload slot on content that could just be pasted as text.
[SCF-06] Flat-File Text Injection — reach for this instead of pasting a large document straight into the chat box.
[SCF-07] Workflow-Anchored Connector Discovery — reach for this before manually doing something a native integration might already do for you.
[SCF-08] Park a Task So You Don’t Lose Your Train of Thought — reach for this to park a mid-conversation task without losing your current train of thought.
[SCF-09] Debugging a Spreadsheet Like Production Code — reach for this when a spreadsheet error needs root-cause treatment, not a one-cell patch.
[SCF-10] Encryption Is Not the Answer — reach for this when the instinct to “encrypt everything” is aimed at the wrong actual threat.
[SCF-11] Email Is the Master Key — reach for this before connecting an AI assistant to email, the one account whose compromise cascades into everything else.
[SCF-12] Prune the Ceremony Your AI Can’t See — reach for this when a standing instruction is firing more often than it needs to, and only you can see it.
[SCF-13] Treat “I Can’t” as a Claim to Test — reach for this whenever a model claims it can’t do something it did easily before.
[SCF-14] Test the Whole Channel, Not Just the Step — reach for this when every step reported success but the end result is wrong — to check whether the pipeline itself delivered.

6.5 GOV — Governance (Elite, Gated)

[GOV-01] The Self-Correcting Feedback Loop — reach for this to turn a caught mistake into a standing rule, not just a one-time fix.
[GOV-02] The Structural Unit Test — reach for this to check structural integrity against a reference nobody involved in the change could have touched.
[GOV-03] See What Else Moves When You Move One Thing — reach for this when changing one variable could ripple into domains you weren’t thinking about.
[GOV-04] Screen Instructions by Where They Came From — reach for this whenever an instruction arrives inside content rather than being typed by you directly.
[GOV-06] Name the Ceiling, Not the Comfort — reach for this when describing what a safeguard does — say what it can’t guarantee in the same breath.
[GOV-07] Only One Channel Publishes — Make Forgery Loud — reach for this when multiple threads share a space and you need a way to tell a real standing rule from a fake one.
[GOV-08] Graceful Session Close — reach for this to make the state record the final action of a session, not merely something written before the session ends.

6.6 SAFE — Safety (cross-tier, non-negotiable)

[SAFE-01] The Data Ingestion Boundary — reach for this before typing or uploading anything — decide what never goes in, in the first place.
[SAFE-02] The Decoupling Charter (The Bright Line) — reach for this in medicine, law, or money: the AI preps the questions, a credentialed professional makes the call.
[SAFE-03] Temporal Drift Verification — reach for this for anything current — give the model today’s date rather than trusting its memory.
[SAFE-04] Time-Grounding: The Clock It Doesn’t Have — reach for this whenever a time-sensitive answer arrives without a dated source behind it.

7 APPENDIX: Platform Compatibility

The cells below are hypotheses from published capability docs ([doc]) or hands-on tests ([verify: DD/MM/YY, tier]). Platform capability is the book’s most perishable content — re-check with [SCF-07] before relying on any cell. AS OF: time of writing.

How to read this: Most patterns are platform-agnostic — pure prompting/process disciplines that work on any capable chat model. They carry no compatibility flag. Only the feature-dependent patterns below need one. Baseline = free tier; paid shows what the paid consumer tier unlocks.

A note on DeepSeek: included here as a reader reference, not a tested scope platform. This book’s verified three-platform set remains ChatGPT, Gemini, and Claude (see Preface). DeepSeek cells are documented claims only; none are hands-on verified.

Pattern	Needs	ChatGPT (free / Plus)	Gemini (free / AI Plus+)	Claude (free / paid)	DeepSeek (free)
[ONR-04] Direct Grounding	file/image upload	`[doc]` Free: 3 files/24h · Plus: 80/3h rolling, 20/message	`[doc]` Free: 10 files/session (no code files) · AI Plus+: Drive attach, code files	`[doc]` consistent across tiers (message limits vary, not file count) / Projects: persistent	`[doc]` app: yes — session-based text extraction; not retained between chats
[ONR-05] Multi-Modal Equity	voice + image input	`[doc]` Free: limited voice · Plus: full voice + vision	`[doc]` Free: Gemini Live (mobile only) + image · AI Plus+: full	`[doc]` app: voice (iOS/Android) / yes	`[doc]` No voice in official app; image input not supported in consumer chat
[ARC-01] Swap Disk	file re-hydration; persistent project	`[doc]` Free: Projects (5 files/project, since Sep 2025) · Plus: Projects (25 files/project)	`[doc]` Free: Gemini Notebook (50 sources/notebook, syncs in-app) + Gems for persona · AI Plus: 100 sources/notebook	`[doc]` manual upload / Projects (paid): persistent workspace	`[doc]` No Projects or notebook equivalent; each session starts fresh
[ARC-03] Different AI Attacks the Work	access to 2nd model lineage	`[doc]` use any 2 platforms (even 2 free tiers)	—	—	—
[ARC-04] Single-Writer Reconciliation	store write/copy	`[doc]` read-only connector (paid); no write	`[doc]` Gemini Notebook + Drive: read/write with human-review step (AI Plus+, Workspace)	`[verify: 02/07/26, this env]` create/write(text)+copy verified; no move/delete; binary human-placed	`[doc]` no file store connector; no write capability
[ARC-05] Archived Versions	file store + archive	`[doc]` manual versioning; no AI copy	`[doc]` manual + AI can create in Gemini Notebook	`[doc]` manual + AI `copy`	`[doc]` manual only; no AI copy; session-based
[ARC-07] Recheck Against Original	read linked sources	`[doc]` manual re-upload (free) / read connector (paid)	`[doc]` upload / Gemini Notebook maintains source links	`[doc]` manual / read connector	`[doc]` manual re-upload only; session-based
[SCF-01] One Thread Per Topic	persistent threads/persona	`[doc]` Free: Projects (basic, 5 files) · Plus: full Projects	`[doc]` Free: Gems (persona, all tiers) + Gemini Notebook (source workspace) · AI Plus: more notebook sources	`[doc]` manual threads / Projects (paid)	`[doc]` no persistent threads; no Projects/Gems equivalent
[SCF-02] Model Routing	tier/model choice	`[doc]` Free: GPT-5 (~10 msg/5h then mini) · Plus: GPT-5.5 + model picker	`[doc]` Free: Gemini 3.6 Flash + 30 Pro prompts/day · AI Plus: 128K ctx · AI Pro: 1M ctx	`[doc]` — / model choice (Sonnet, Opus etc.)	`[doc]` DeepSeek V4 + V4-Flash available; no consumer model picker
[SCF-05] Resource Substitution	(manages upload quotas)	`[doc]` Free: tight (3 files/24h) · Plus: 80/3h rolling	`[doc]` Free: generous (10/session); no code files free	`[doc]` paid-leaning (message limits vary, file count consistent)	`[doc]` session-based; quota not displayed
[SCF-06] Flat-File Injection	file attach	`[doc]` Free: 3/24h · Plus: 80/3h, 20/message	`[doc]` Free: 10/session · AI Plus+: Drive attach	`[doc]` paid (limited free)	`[doc]` app: yes (session-only, text extraction)
[SCF-07] Connector Discovery	connectors	`[doc]` paid only — Agent mode: GitHub, Google services	`[doc]` Free: Google Search · AI Plus+: Gmail sidebar · AI Pro+: full Workspace	`[doc]` paid only	`[doc]` no connectors in consumer app

The cross-platform finding worth stating in prose (at time of writing): none of the four platforms grants an AI unrestricted autonomous write/move/delete on your file store as a standard consumer feature — Claude reads/copies and (in this project’s configuration) writes raw text at a path; ChatGPT’s connector is read-only; Gemini creates/edits and proposes moves behind a human-review dashboard; DeepSeek has no file store connector at all. This is why [ARC-04]’s split-responsibility model (AI writes content, human controls placement) is the correct pattern everywhere, not a workaround for one platform.

Gemini nomenclature note (updated at time of writing): two distinct tools that are often confused. Gems = reusable custom AI personas with full Gemini capabilities (analogous to ChatGPT’s custom GPTs); available on all tiers including free. Gemini Notebook (renamed from NotebookLM on 16 Jul 2026) = source-grounded research workspace where the AI answers only from uploaded sources; now integrated in the Gemini app sidebar; available free (50 sources/notebook) with higher limits on paid tiers. The two can be combined. Gemini tier names as of time of writing: Free · AI Plus ($4.99/mo) · AI Pro ($19.99/mo) · AI Ultra ($99.99+/mo); compute-based usage limits since I/O 2026.

Two caveats that outlast any version number. Many readers work in Microsoft environments (OneDrive/SharePoint) rather than Google Drive — the Swap Disk Protocol’s mechanics are cloud-agnostic, but connector tooling differs by provider and should be re-verified per platform. And the exact click-sequence to grant or re-grant a write scope is interface-specific, and has been observed to change between releases: treat the need to re-verify write access per thread as durable, the steps as perishable.

Verification note: every [doc] cell is a documented hypothesis, not a hands-on test. The single [verify] cell is from this project’s own production environment — Claude, at time of writing. Two AI models agreeing on a cell is not verification. This appendix is itself an instance of [PRAC-05]: a documented estimate, flagged as such, awaiting reconciliation to actuals.

8 APPENDIX: Standing Rules of Engagement

Working rules for producing this book across multiple AI threads — and directly reusable by any reader running their own multi-model workflow. Every rule here was earned by a specific failure during this book’s own production.

No self-attestation. Never claim your own output is “verified,” “audited,” “locked,” or “consensus.” State what you changed; leave confirmation to an independent pass. An audit entry is written after the reviewer returns, never pre-filled.
Reference by stable ID, never display number. IDs are frozen; display order is cosmetic; never renumber a reference.
After any structural change, list the affected pointers as unverified for the reviewer — don’t assert they’re fine.
The verifier must be external to the verified. Any check must sit outside the thing checked; no model grading its own work.
Perishable content is principle-first and timestamped (SST). Lead with the durable principle; mark specifics (models, prices, limits) as “as of [date], illustrative.”
One canonical file; the frozen version wins ties. When a pattern exists in two places and they differ, flag it — don’t silently reconcile. Never compress or drop cleared content when porting; if compressing for display, say so.
Flat engineering register. No hype adjectives, in the manuscript or the surrounding notes.
Standard submission format: what changed · affected pointers (unverified) · the draft.
Bright line on high-stakes content stated in-section: AI prepares questions and surfaces what to check; human and professional decide.
Beginner bar on Beginner-tier content: followable by a non-technical reader; no jargon.

(Note: the disciplines in these rules also appear as reader-facing patterns — [PRAC-06] Decoupled Reference Audit, [PRAC-03] Canonical-File Discipline, [SCF-03] Don’t Trust What It Says About Itself, [SAFE-03] Temporal Drift Verification and [SAFE-04] Time-Grounding. The rules are the operational version; the patterns are the taught version.)

The project’s internal work queue, backlog, status notes, and full production changelog are maintained in the separate private editorial file — not in this public manuscript.

9 APPENDIX: Further Reading

This book covers one layer. These cover the ones around it. Listed because a reader who wants the whole picture should have it — not as prerequisites. You can start here and read outward.

The layer below: prompting. DAIR.AI, Prompt Engineering Guide — promptingguide.ai. Free, open-source, community-maintained, kept current. The definitive practical treatment of the turn: zero-shot and few-shot, chain-of-thought, self-consistency, and the rest. If a pattern in this book assumes you can get a decent answer out of a single question, this is where that skill is taught.

White, J. et al., A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT — arXiv:2302.11382 (Vanderbilt, 2023). Sixteen prompt patterns documented in the software-pattern form this book also borrows. Written for software development tasks, so the examples are developer-facing, but the pattern discipline transfers.

The layer sideways: building things on top of models. Lakshmanan, V. and Hapke, H., Generative AI Design Patterns (O’Reilly). Thirty-two patterns for people shipping applications and agents — hallucination, non-determinism, guardrails, retrieval. Assumes you write code.

Gullí, A., Agentic Design Patterns (Springer). Patterns for building autonomous agents, worked through agent frameworks. Also assumes you write code.

Why these are not competitors to this book. Every one of them is written for someone building with a model — engineers, application developers, agent authors. This book is written for someone working alongside one in a chat window. The prompting resources sit underneath it: they make each turn better. This book is about everything that has to survive after the turn ends. Read them together and you have both floors of the same building.

A caution, in this book’s own spirit. Links rot and editions change. Titles and authors are durable; the URL above was checked at time of writing and may not be accurate by the time you read this.