01
Chat
Use for drafting, explaining, analyzing supplied material, brainstorming, or one-off plans. You remain in the loop every turn.
Free AI Task Router
For AI nerds: describe a task before you overbuild it.
Describe a task and let this free tool tell you whether it can work things out via a chat, one agent, a team, or none of the above—and what to do next.
How the router thinks
Architecture follows the work.
A chat is best for bounded thinking. One agent is useful when a repeatable goal requires tools or files. A team is justified only when independent specialist work can run in parallel and a coordinator can reconcile it. High-stakes judgment remains human-led.
02
One agent
Use for a repeatable goal that needs several steps, tool access, memory, or a schedule. Add permissions and an approval gate.
03
Agent team
Use when parallel specialist roles materially improve speed or quality. Include a coordinator, evidence standards, and a reviewer.
None of the above is a valid answer.
AI may assist with preparation, but final medical, legal, employment, safety, credit, and irreversible financial decisions need an accountable human. Physical work and undefined goals also cannot be solved by adding agents.
AI may assist with preparation, but final medical, legal, employment, safety, credit, and irreversible financial decisions need an accountable human. Physical work and undefined goals also cannot be solved by adding agents.
Token Guide
OpenAI, Claude, and Gemini: what tokens change.
A token is a chunk of text. Context windows limit how much input and prior conversation a model can consider; output limits cap how much it can generate. Bigger is not automatically better: long context costs more, can slow responses, and may bury the important instructions.
ChatGPT / OpenAI
Check the exact model
OpenAI offers models with different context and output limits. ChatGPT product limits and API model limits are not always the same.
Practical rule: For ordinary drafting, do not stuff an entire archive into one prompt. Retrieve only the relevant material, label sources, and reserve room for the answer.
Claude / Anthropic
Long context, model-dependent
Claude model families differ in context, output, price, and extended-thinking behavior. Some token usage may be consumed by reasoning rather than visible prose.
Practical rule: Ask Claude to map a long document first, then analyze named sections. Use prompt caching or retrieval when repeatedly sending the same large reference set.
Gemini / Google
Large multimodal contexts
Gemini models can accept text, images, audio, video, and code, with limits varying by model and access tier.
Practical rule: Large context is useful for video, codebases, and document collections, but still structure the prompt, identify the evidence to use, and request citations or timestamps.
Estimate before you build.
Token usage includes instructions, conversation history, retrieved files, tool results, reasoning where applicable, and the model's response. Agents multiply usage because every step and handoff can create another model call. A five-agent workflow is not one prompt—it may be dozens.
Token usage includes instructions, conversation history, retrieved files, tool results, reasoning where applicable, and the model's response. Agents multiply usage because every step and handoff can create another model call. A five-agent workflow is not one prompt—it may be dozens.
| Situation | Token-conscious approach |
|---|---|
| One short document | Use a normal chat and ask for a specific output. |
| Many recurring documents | Index or retrieve relevant chunks instead of resending everything. |
| Long conversation | Summarize decisions into a compact brief and start a clean thread. |
| Agent workflow | Limit steps, tool outputs, retries, and handoff text; set stop conditions. |
| Agent team | Give each specialist only the context needed for its role; let the coordinator receive concise findings and sources. |
Learn the foundations
Not sure what a token, sandbox, or agent really is?
Take the free fluency test. Beginners receive a direct invitation to the planned self-paced course waitlist.
Take the AI test