Attribute the cost before changing config: tsc -p . --noEmit --singleThreaded --extendedDiagnostics, then --generateTrace DIR with @typescript/analyze-trace. Changing flags by habit fixes the wrong phase; only a dominant Check time points at types, and Parse time points at program size (--listFilesOnly, --explainFiles).
Compare counters only with --singleThreaded (or the same --checkers N) and the same tsc version, and delete the *.tsbuildinfo files (see tsBuildInfoFile and outDir) first. tsc 7 changes Types and Memory used with the worker count, and a warm .tsbuildinfo hides the real check.
Check tsc --version and tsc --showConfig before using flags. tsc 7 (the Go port) removes target: es5, downlevelIteration, baseUrl, and moduleResolution: node10 (TS5108, TS5102) and defaults types to []; advice from 5.x docs breaks there. See TypeScript 7.
Do not add skipLibCheck by reflex. It also skips your own hand-written .d.ts files and can hide conflicts, so measure with and without it, diff the errors once, and keep one CI check without it when the repo has hand-written .d.ts files (git ls-files '*.d.ts' outside build output). Use it in both runs when measuring a type-level change.
isolatedDeclarations alone does not speed up tsc. It needs an explicit annotation on every export (TS90xx otherwise), and the gain comes from emitting declarations without type-checking (--noCheck or another emitter) next to a separate tsc --noEmit that still reports errors.
Do not weaken public types (any, object, a base interface) to save checking time without the API owner. Prove an edit changed nothing with the project's type tests or a mutual-assignability probe (Eq<Before, After>).
Split repeated full re-checks with composite project references and tsc -b, or incremental with a persisted .tsbuildinfo. A .tsbuildinfo not restored between CI runs only adds a write. References must form a DAG; split along the existing package graph (typically 5 to 20 projects).
Edit a type only when a trace names it: large unions in hot signatures, A & B & C compared often, conditional return types, TS2589 recursion. Some wiki advice has no measured gain on tsc 7 (named conditional types); keep an edit only if the counter that justified it dropped. See Build settings and type costs.
Narrow include and types only after --listFilesOnly shows surplus files, and keep tests type-checked in their own project. A types list that omits node or a test framework fails with "Cannot find name".
Node and Bun strip types without checking. Keep tsc --noEmit or tsc -b in CI wherever code runs without a tsc emit.
Timing claims need at least 5 runs on a quiet machine; report the median and spread. Report a change with no measured gain as null instead of keeping it.

Workflow

Record tsc --version, the effective options, and the baseline counters with the command above.
Attribute with the trace or the file list, apply one change, and rerun the same command with the same flags.
Run the normal build and tests, and keep the change only if the counter or time dropped and diagnostics are unchanged.
Report the tsc version, the counters or median times before and after with the same flags, the diagnostics diff, and each check not run, by name.

Leave new type errors to the fix-diagnostics-without-suppressing skill: fixing them is not a speed change.

References

Read Measurement when attributing cost: counters, traces, tsc 7 profiles, and program-size checks.
Read Build settings and type costs when choosing skipLibCheck, incremental builds, project references, isolatedDeclarations, --noCheck, or editing expensive types.
Read TypeScript 7 when the compiler is tsc 7 or tsgo, for removed options, new defaults, --checkers, --builders, and tsc6.

optimize-tsc-builds

Optimize tsc Builds

Rules