How to decide which model to use, how much reasoning effort to spend, when to plan, and when to start a new conversation — with the reasoning separated from the facts that expire.
The research was carried out on 2026-09-07 and is not being updated.
Model names, effort levels, quota ratios and prices change every few months, and when they do, this repository will be wrong silently — no build will fail, no test will go red. That is exactly why it is published as a snapshot with the date in the filename instead of as a living guide that quietly rots.
Everything that expires is confined to one file. Everything durable is in another. If you only read one thing, read
docs/principles.md.
Most people using AI coding tools leave the model and effort settings wherever they happen to be. That is expensive in two directions at once: a frontier model on a rename wastes money, and a weak model on a subtle concurrency bug wastes an afternoon.
The advice that does exist is mostly blog posts copying each other, presenting one person's habit as a universal rule, with no dates and no sources.
This repository is an attempt to do the opposite:
- Label the confidence. Every claim is marked as consensus (5+ independent sources), disagreement (with the dissenting view named), or one person's opinion.
- Preserve the contradictions. Where sources disagree — and on context-window thresholds they disagree wildly — both sides are recorded, because the disagreement is the finding.
- Separate what lasts from what doesn't, so the durable part stays useful after the volatile part expires.
- Say what couldn't be verified. Reddit blocked the crawler; several sources returned 403. That is listed, not hidden.
The design idea worth stealing even if you never read the content:
| Layer | File | Lifespan | Trust it? |
|---|---|---|---|
| Principles | docs/principles.md |
Years | Yes. Written with no model names or numbers so it survives product churn. |
| Findings | docs/findings-2026-09-07.md |
Weeks to months | Only as a record of what was true on 2026-09-07. |
| Your setup | snippet/claude-md-rules.md |
Yours | Fill it in yourself. The placeholders are not recommendations. |
Mixing these three together is the mistake that makes almost every AI-tooling guide on the internet misleading within a year. Keeping them apart means the middle layer can expire without taking the rest down with it.
docs/
principles.md The durable layer — decision logic, no product names
findings-2026-09-07.md The snapshot — everything specific and expiring
sources.md Every source, including the ones that failed to load
snippet/
claude-md-rules.md A block to paste into your CLAUDE.md, with a config
section you must fill in first
Each file exists in English, 繁體中文 and 简体中文.
If you want a rule of thumb right now:
If you can describe the resulting diff in one sentence, skip planning and use a cheap configuration.
That single heuristic is the highest-consensus finding in the whole survey and prevents most overspending.
If something is going wrong, the diagnosis matters more than the escalation:
- It missed a file, skipped the tests, didn't double-check → it didn't think enough. Raise effort.
- It had everything it needed, visibly tried, still failed → it isn't smart enough. Change model.
- You've corrected the same thing more than twice → the context is poisoned. Start a new session. Don't escalate.
The third case is the one people skip, and it's the one that matters most. The full reasoning is in docs/principles.md.
- Not official. No affiliation with Anthropic. For any current number, go to Anthropic's documentation.
- Not maintained. No issues will be triaged, no updates are planned. Fork it freely.
- Not benchmarks. Nothing here was measured by the author. It is a survey of what other people published, with the sourcing made visible so you can judge it yourself.
- Not universal. Quota rules differ by plan. Someone else's numbers are not your numbers.
The underlying survey was conducted by AI research agents (roughly 45 searches across ~70 pages) and then reviewed and organized by a human. This is disclosed because it should change how you read it: the sourcing is unusually complete for this kind of guide, but no claim here was independently verified by hand, and second-hand relays are labelled as such throughout.
CC BY 4.0. Copy it, translate it, adapt it, quote it — just credit the source.
If you republish an adapted version, please keep the snapshot date visible. The date is the most important piece of metadata in this repository.