TTT-b/stremio-ai-translator

Model routing and token saving

Which model

• Sonnet (default, ~90% of work): edits, bug fixes, tests, refactors of known code, git, docs, README/TROUBLESHOOTING updates, running npm run check.
• Opus (only when it pays off): a new feature spanning several modules, architecture or data-flow changes, a bug whose cause is unknown after one Sonnet attempt, security review, large migrations. Use it to produce a short written plan (max ~30 lines), then execute the plan on Sonnet.
• Haiku: pure lookups, grep-and-report, summarizing logs, listing files. Delegate with a subagent and model: haiku.
• Do not escalate to Opus for style, naming, or one-file changes. If a Sonnet attempt fails twice, escalate once, then return to Sonnet.

Saving tokens without losing quality

1. Read PLAN.md first. Then read only the files the task touches. Use Grep/Glob before opening files; read line ranges, not whole files.
2. Never read node_modules, dist-electron, release, logs, data, package-lock.json, or .env.
3. Run npm run check once at the end, not after every edit. Run a single test file (npx vitest run tests/<file>) while iterating.
4. Pipe long command output through Select-Object -First 40 / -Last 40; do not dump full logs.
5. Do not re-read a file just edited. Do not restate facts already established in the conversation.
6. Keep replies short: what changed, what was verified, what the user must do next.
7. Offload broad searches or log reading to a subagent so only its conclusion enters the main context.
8. One task per session. Start a new session when the topic changes instead of carrying a long history.
9. Never skip the check, tests, or the reading of files that affect correctness to save tokens.