Source
Bring the real document, data, example, or project.
The practical companion: context patterns, reusable briefs, current models, tools, sources, and a human-review loop.
Choose a task small enough to inspect and useful enough to matter. The model can do the draft work; you still own the outcome.
Bring the real document, data, example, or project.
Name the role, task, context, and output format.
Ask for an artifact you can examine—not just an answer.
Verify facts, fit, risk, and what changed before using it.
Move up only when the task benefits from more durable, inspectable context. “Memory” is never a substitute for checking what the system actually has.
Fast, temporary, and ideal for bounded questions. You supply the relevant context in the conversation.
A curated context boundary with files and instructions for a sustained topic or body of work.
Source, tools, tests, and durable records enable multi-step execution you can inspect and verify.
Aim for inspectable, user-controlled context—not “perfect memory.” Remove information the task does not need, especially sensitive material.
It is a starting structure, not a magic incantation. Clear context and an inspectable output matter more than ornamental prompt language.
Who should do the thinking?
What concrete outcome is needed?
What source material and constraints apply?
What should arrive, and how will it be checked?
Replace generic nouns with your actual material. Never paste confidential information into a system without confirming that its access and data policy fit the work.
Role: You are a senior front-end game designer. Task: Build a polished, playable Snake game in one HTML file. Context: It will be shown live to a nontechnical audience, must work offline, and must be easy to inspect. Format: Return one complete HTML file with keyboard and touch controls, score, restart, accessible contrast, and a short verification checklist.
Research this decision using current primary sources. Separate confirmed facts from inference, cite every material claim, identify unresolved questions, and end with a one-page decision brief.
Inspect the existing project first. Write acceptance tests for the requested behavior, implement the smallest complete change, run the tests, inspect the result in a browser, and leave a concise handoff with changed files and remaining risks.
Explain this topic at five levels: a child, a high-school student, a working professional, a domain specialist, and an executive deciding what to do next. Keep the core idea consistent while changing vocabulary and depth.
Assess this damaged or incomplete source. List what is observable, what is uncertain, and what must not be invented. Then propose a reversible restoration plan and create a first pass that preserves provenance.
Study the supplied brand references before creating anything. Extract the typography, color, spacing, voice, and composition rules; then produce the requested asset and include a short brand-alignment checklist.
Models change quickly. Use these official links to verify availability, terms, and model cards before making a consequential choice.
| Provider / model | Useful when | Access | Openness / license | Official source |
|---|---|---|---|---|
| OpenAI · GPT-5.6 | General knowledge work, coding, research, and agents | ChatGPT, Codex, API | Proprietary · service terms | OpenAI |
| Anthropic · Claude Fable 5 | Long-running complex knowledge and coding work | Claude, platform, cloud marketplaces | Proprietary · service terms | Anthropic Fable |
| Anthropic · Claude Opus 5 | High-capability analysis and professional work | Claude and platform | Proprietary · service terms | Anthropic Opus |
| Anthropic · Claude Sonnet 5 | Balanced everyday writing, analysis, and coding | Claude and platform | Proprietary · service terms | Anthropic Sonnet |
| Google · Gemini 3.7 Flash | Fast multimodal, long-context, coding, and agents | Gemini, AI Studio, Vertex AI, API | Proprietary · service terms | Google AI |
| xAI · Grok 4.6 | Long-running agents, research, coding, and visuals | Grok, partners, API | Proprietary · service terms | xAI |
| DeepSeek · DeepSeek V4 Pro / Flash | Reasoning-effort-controlled agent workflows | App, web, API | Hosted release · service terms | DeepSeek |
| Alibaba · Qwen 3.8 | Hosted production work with long context and tools | Qwen Cloud and products | Hosted; related weights below | Qwen |
| Moonshot · Kimi K3 | Long-context multimodal knowledge and coding work | Kimi app, web, API | Proprietary · service terms | Moonshot AI |
| Mistral · Mistral Medium 3.5 | Multimodal agentic and coding work | API and downloadable weights | Open-weight · Modified MIT v26.04 | Mistral docs |
| Family | Useful when | Access | Openness / license | Official source |
|---|---|---|---|---|
| Qwen 3.8 Flash-Next | Local or self-hosted multimodal experimentation | Weights and common runtimes | Open-weight · Qwen Community License 1.0; conditions apply | Qwen model card |
| Gemma 4 | On-device and self-hosted agentic apps | Weights, AI Edge, Colab, Vertex AI | Open model · Apache 2.0 | |
| Mistral open models | General, coding, vision, audio, and safety workloads | Weights and Mistral hosting | Open-weight · Apache 2.0 or Modified MIT by model | Mistral licensing |
| NVIDIA Nemotron | Customizable agent models and enterprise stacks | Components, NIM, partner runtimes | Open models · NVIDIA Nemotron Open Model License; check card | NVIDIA |
| OpenAI gpt-oss | Local and self-hosted text reasoning with tools | 120B / 20B weights and third-party runtimes | Open-weight · Apache 2.0 plus usage policy | OpenAI open models |
| Meta Llama / Muse | Hosted Meta assistant plus downloadable local options | Meta AI/API; Llama and Muse Glimmer weights | Mixed · proprietary, Llama Community, or Apache 2.0 by release | Meta Muse Glimmer |
“Open-source,” “open-weight,” and “proprietary” are not interchangeable. Verify the specific model card and license for the exact version you intend to use.
Start in chat, move to a project when context recurs, and use an agentic harness when the work must inspect files, use tools, and verify an artifact.
Use this as a companion explanation of the move from a single chat to curated projects and tool-enabled agentic work.
These are attendee-safe sources used in the presentation. The extended model directory above links directly to provider announcements, documentation, or model cards.
A short fluency and safety checklist for every consequential task.
Outcome, audience, evidence, boundary.
The work the model can actually perform.
Truth, taste, risk, downstream effects.
Sources, decisions, changes, verification.
Source it, brief it, build it, check it.