Different models, different jobs
Updated October 31, 2026. A routing pattern, not a fixed model ranking or a promise of unlimited subscription use.
I don't need every model to be my best writer, best coder, and best critic. I need a useful division of labor—and a way to see whether the work actually got done.
Separate thinking, execution, and review
| Job | Useful route | Finish line |
|---|---|---|
| Notice changes | Ordinary code comparing IDs, timestamps, or hashes | No model call when nothing relevant changed |
| Think together | An interactive assistant that fits your allowance, such as Microsoft Copilot Chat | A clear question, alternatives, and a small approved brief |
| Execute | A supported coding agent, such as Codex or Claude Code | An inspectable change, real checks, and a recorded result |
| Compose | A writer whose voice suits the audience | A fact-checked draft, not an automatic send |
| Challenge the result | A capable reviewer from another model family, when worthwhile | Specific findings tied to evidence, followed by tests and human judgment |
These are jobs, not permanent product identities. Some chat products also offer agents or actions; Codex and Claude can help you think. Choose for the task, data policy, available allowance, and tools—not a brand stereotype.
What Adam uses now
OpenAI-backed Codex remains the normal execution path. Claude Code is deliberately selected for some deep coding and architecture work. On October 31, Riley's communication-drafting route was changed to Claude Opus because Adam prefers its more personable writing style. That is a subjective editorial preference, not a benchmark showing one provider is always better.
Riley is on demand, with finite run and concurrency caps. Giving a writer a stronger model does not require sending an unchanged inbox through that model on a timer. The dated setup distinguishes configuration from verified execution.
Microsoft Copilot Chat is the conversational thinking-partner example here, not GitHub Copilot and not a new background-job route. Older pipelines may use different routes; inspect their actual configuration before assuming they match today's conversational setup.
A subscription is an allowance, not an endless buffet
Use the subscriptions you already have where their supported tools and terms fit. But do not assume chat, CLI, hosted agents, and arbitrary API access share the same entitlement. Short interactive planning can be economical; a large context or multi-step agent run can consume much more.
Microsoft Copilot Chat and GitHub Copilot are different products. Adam's thought-partner example is Microsoft Copilot Chat: use the interactive chat to explore alternatives and sharpen a brief before handing work to an execution agent. Eligible Microsoft 365 accounts include a baseline chat experience at no additional cost. Account eligibility, feature limits, capacity, and separately metered agents still matter; ordinary chat access is not an entitlement to unlimited automated execution. Microsoft Copilot Chat FAQ, Microsoft feature limits.
GitHub Copilot is a separate developer product. If you choose it for a CI or coding workflow, check its own billing rules. Those rules are not the basis for the Microsoft chat example.
Codex usage depends on the plan and workload; Claude and Claude Code can share subscription limits. Check the current provider dashboard rather than copying an old model price table. Do not silently enable API billing or purchased overages when a subscription reaches its limit. Official OpenAI allowance guidance, Claude Code subscription guidance.
Give each task a small source packet, a run limit, a stop condition, and a useful deliverable. Track completed outcomes alongside tokens and failures. A zero in an orchestration cost ledger is not proof that the subscription is free. See cost controls.
Different products do not necessarily mean different model families
A product name is not a model-family label. GitHub Copilot, for example, can expose models from multiple providers. Reviewing Claude output with a Claude model inside another product is not cross-family review. Record the actual model used, not just the app name; automatic model selection can make that harder to establish.
A second family can offer another perspective, but models still share failure modes and may confidently agree on the same mistake. Ask the reviewer to assess the original requirements and evidence rather than simply endorse the author's summary. Passing tests, source verification, and human approval still matter. Agreement is not a correctness certificate.
GitHub or Gitea?
For the repository and pipeline choice, read GitHub and Gitea development workflows. Keeping the repository on your own hardware and keeping model inference offline are separate decisions.