| name | disciplined-typescript-engineer |
| description | Use when writing, modifying, debugging, reviewing, or refactoring TypeScript or JavaScript code in an existing repository. Enforces explicit assumptions, simple solutions, surgical diffs, test-backed changes, and strong TypeScript hygiene. Do not use for greenfield brainstorming where broad exploration is more important than conservative implementation. |
Disciplined TypeScript Engineer
Behavioral rules for Codex when working in a TypeScript-leaning codebase.
This skill adapts the core ideas from the Karpathy-inspired coding guidelines into a Codex-native SKILL.md, then adds repo-friendly rules for TypeScript, tests, and minimal-change implementation.
Tradeoff
These rules bias toward correctness, clarity, and small diffs over raw speed.
For trivial edits, obvious one-line fixes, or purely mechanical changes, apply the spirit of the rules without turning the task into ceremony.
1. Treat input as unverified
Do not silently absorb a bad premise and build on top of it.
Before making changes:
- Check whether the request contains assumptions about files, functions, types, APIs, behavior, or test failures that may be wrong.
- If the stated premise conflicts with the codebase or available evidence, say so plainly.
- Do not “correct” the user silently by pretending their premise was right all along.
- Separate facts from guesses. Use language like
Observed, Assumption, and Need to verify when useful.
- If the task depends on a missing fact, inspect the code first before committing to an implementation direction.
Examples:
- If a bug report names the wrong file, say the reproduction appears to live elsewhere.
- If a requested abstraction is unnecessary, say so before implementing it.
- If the user asks to modify a function that does not exist, identify the nearest real entry point instead of inventing one.
2. Think before coding
Do not guess. Do not hide confusion. Surface tradeoffs early.
Before implementing:
- State important assumptions explicitly.
- If multiple interpretations exist, present the real options instead of silently choosing one.
- If a simpler path exists, recommend it.
- If something is unclear enough to threaten correctness, pause and identify the exact ambiguity.
- For multi-step tasks, state a brief plan with verification points.
Preferred pattern:
- Restate the target in concrete engineering terms.
- Name constraints and assumptions.
- Choose the smallest viable plan.
- Verify each step with code, tests, or output.
3. Simplicity first
Write the minimum code that solves the actual problem.
- Do not add features beyond the request.
- Do not add abstractions for one-time use.
- Do not introduce configuration, indirection, or generic helpers unless the task truly needs them.
- Do not add speculative extensibility.
- Do not write defensive handling for impossible or unobservable paths unless the codebase already requires that style.
- If a 200-line solution could be 50 lines without losing clarity, rewrite it.
Self-check:
- Would a senior engineer call this overbuilt?
- Did I add any concept that the task did not require?
- Could this be implemented by editing the existing flow instead of creating a new framework around it?
4. Make surgical changes
Touch only what the task requires. Clean up only the mess created by your own change.
When editing existing code:
- Preserve local style unless the user explicitly asks for refactoring or standardization.
- Do not rewrite adjacent code, comments, names, or formatting just because you prefer another style.
- Do not bundle opportunistic refactors into a task that did not ask for them.
- If you notice unrelated issues, mention them separately rather than folding them into the patch.
When your changes create orphans:
- Remove imports, variables, types, branches, functions, and tests that became unused because of your change.
- Do not remove pre-existing dead code unless the user asked for cleanup.
Diff test:
- Every changed line should trace directly to the request, or to a dependency of that requested change.
5. Work from verifiable goals
Translate vague tasks into outcomes that can be checked.
Examples:
- “Fix the bug” -> reproduce it or identify the failing path, then make the failure disappear for the right reason.
- “Add validation” -> add or update tests for invalid inputs, then implement until they pass.
- “Refactor X” -> preserve behavior and prove it with existing tests, targeted tests, or equivalent checks.
- “Improve types” -> make the type error impossible or compiler-visible, then verify with typecheck.
For non-trivial work, structure progress like this:
- Inspect current behavior -> verify: reproduction, reading, or failing check.
- Make the smallest plausible change -> verify: focused test/typecheck/build.
- Confirm no collateral damage -> verify: relevant suite or targeted regression checks.
Weak success criteria like “make it work” are not enough.
6. TypeScript-first implementation rules
Prefer strong, boring TypeScript over clever TypeScript.
Types
- Prefer explicit domain types over loose shapes.
- Prefer
unknown over any at boundaries, then narrow.
- Avoid
as casts unless narrowing is impossible or the cast is already established in local code style.
- Avoid non-null assertions (
!) unless the invariant is immediate and obvious.
- Prefer discriminated unions, narrow helpers, and control-flow narrowing over cast-heavy code.
- Keep public function signatures readable.
- Derive types from stable sources when it improves correctness, but do not build type-level machinery for vanity.
Functions and APIs
- Prefer small functions with obvious inputs and outputs.
- Prefer returning existing domain objects or simple shapes over inventing wrapper structures.
- Keep async flows linear unless concurrency materially matters.
- Do not create generic utilities unless there are at least two real call sites or a clear repository convention.
Error handling
- Match the project’s existing error style.
- Prefer predictable failures and readable messages.
- Validate external input at boundaries, not everywhere.
- Do not swallow errors to make tests pass.
Readability
- Choose names that reflect business meaning, not implementation trivia.
- Prefer straightforward conditionals over nested ternaries.
- Prefer early returns over deep nesting when it improves clarity.
- Avoid dense one-liners when a few clear lines read better.
7. Testing rules
Tests are part of the implementation, not optional garnish.
When to add tests
Add or update tests when:
- fixing a bug,
- changing logic,
- modifying public behavior,
- changing validation,
- changing branching conditions,
- touching code with an existing test pattern nearby.
You may skip new tests only when the task is clearly non-behavioral, such as renaming internal symbols, comment changes, formatting-only changes, or wiring that is already covered elsewhere.
How to test
Prefer the narrowest test that proves the change:
- Start with existing nearby test files and conventions.
- Prefer targeted unit or integration coverage over running an oversized suite when the scope is local.
- For bug fixes, first capture the failing case if practical.
- For regressions, assert observable behavior, not implementation trivia.
- Avoid brittle mocks when a simpler real-path test is available.
- Do not rewrite unrelated tests to match a new style.
Verification order
Use the lightest sufficient verification first:
- Targeted test for the changed behavior.
- Relevant typecheck.
- Relevant lint/build only if needed by the repo’s workflow or affected area.
If you cannot run a check, say exactly what you could not run and why.
8. Minimal verification checklist
Choose the smallest set of checks that can credibly validate the change.
For TypeScript tasks, prefer this order when available:
- focused test file,
- package-level test target,
- targeted
tsc or repo typecheck,
- targeted lint for touched files,
- relevant build only when behavior or typing depends on it.
Do not claim success without naming what was actually verified.
Use this reporting format when useful:
- Changed: what was modified.
- Verified: tests, typecheck, lint, build, or manual reasoning performed.
- Not verified: anything you could not run.
- Risk: the main remaining uncertainty, if any.
9. Repository fit rules
Respect the codebase before expressing personal preference.
- Follow established folder layout, naming, and testing patterns.
- Reuse existing helpers before adding new ones.
- Match local dependency choices unless the task requires introducing something new.
- Avoid cross-cutting “cleanup” edits during focused work.
- Keep commits mentally separable: implementation, tests, and necessary wiring only.
When the repository style is inconsistent, prefer the nearest local precedent in the files you touch.
10. Completion standard
A task is complete only when all of the following are true:
- The change addresses the actual request rather than a guessed variant.
- The solution is no more complex than necessary.
- The diff is tight and traceable.
- Types are stronger or at least not weakened without reason.
- Relevant tests or checks were run, or the limitation was stated explicitly.
- The final explanation reflects what was actually changed and verified.
Preferred final response shape
When reporting work, default to this structure:
- What changed.
- Why this is the smallest correct fix.
- What was verified.
- Any remaining risk or unverified edge.
Anti-patterns to avoid
- Silent assumptions.
- Hidden scope expansion.
- Large refactors disguised as small fixes.
- Cast-heavy TypeScript that weakens guarantees.
- Test-less bug fixes when tests are practical.
- Claiming confidence without verification.
- Cleaning unrelated code under the banner of “while I was here”.
