| name | tiger-style |
| description | Apply TigerBeetle's TigerStyle coding principles (safety, performance, developer experience) when writing or reviewing code. Use when the user says "TigerStyle", "tiger style", "apply tiger style", asks for a TigerStyle review, or asks about the NASA Power of Ten rules, assertion density, 70-line function limit, or batching and back-of-envelope performance sketches. Language-agnostic โ translates Zig-specific rules to the language at hand. |
TigerStyle
TigerBeetle's coding style. Design goals, in priority order: safety, performance, developer experience. Source: TIGER_STYLE.md.
First invocation: read the references
The TL;DR below is a shortlist, not the full guide. On first invocation in a session, use the Read tool to load the reference files you'll actually need before responding:
- Always read references/principles.md โ full distilled rules with rationale.
- If reviewing (user said review/audit/check, or pointed at a diff/PR), also read references/review-checklist.md.
- If the target language is Rust, Java, Kotlin, or TypeScript/JavaScript, also read the matching file listed under Language Scope.
Exception: if the user asks a single narrow question that the TL;DR answers directly (e.g. "what's the function line limit in TigerStyle?"), you can answer from the TL;DR without loading references. When in doubt, read.
In subsequent turns of the same session, references are already in context โ don't re-read unless you need a specific section you missed.
Two Modes
Reference mode โ authoring code
User is writing new code or asks "how should I write this in TigerStyle?". Apply the rules from principles.md proactively. Cite the rule you're applying when it's non-obvious.
Review mode โ auditing changes
User says "tiger style review" or asks you to review changes against TigerStyle. Walk review-checklist.md against git diff (or the files in scope), report violations with file:line citations grouped by severity.
Default to reference mode. If the user says "review", "audit", "check", or points at a diff/PR, switch to review mode.
Language Scope
TigerStyle originates in Zig, but most principles are language-agnostic. Apply cross-language rules everywhere; skip or translate Zig-only rules.
- Zig-only rules (skip in other languages):
zig fmt, @divExact/@divFloor, @prefetch options, usize vs u32, *const for 16+ byte args, in-place init via out-pointer, struct field/type/method ordering idiom, the options: struct named-argument pattern, CamelCase struct files.
- Translates cleanly: assertions, function length, naming, no dynamic allocation after init (โ pre-allocated pools / object reuse), explicit control flow, error handling, comments that say why, batching, back-of-envelope sketches.
For language-specific translations, load the matching reference:
For other languages (Python, Go, C++, etc.), apply the principles from references/principles.md directly and skip the Zig-only rules listed above.
Core Principles (TL;DR)
- Zero technical debt. Do it right the first time. The second time may not transpire.
- Assertions are force multipliers. Minimum two per function on average. Assert arguments, return values, pre/post-conditions, invariants. Assert both positive and negative space.
- Put a limit on everything. Every loop, every queue, every buffer has a fixed upper bound. Fail fast.
- No dynamic allocation after init. Static allocation at startup. (Translate to language idiom: object pools, pre-sized collections, arena allocators.)
- Hard 70-line function limit. Push
ifs up and fors down. Centralize control flow in the parent; keep leaves pure.
- Hard 100-column line limit. No exceptions.
- Explicit control flow. No recursion where iteration works. Minimal abstractions. Split compound conditions into nested
if/else.
- Handle every error. 92% of catastrophic failures in distributed systems come from mishandled non-fatal errors.
- Name things well. Units last, descending significance:
latency_ms_max, not max_latency_ms. Don't abbreviate. Match character lengths of related names (source/target, not src/dest).
- Say why. Comments explain rationale, not what the code does. Commit messages inform and delight.
- Back-of-envelope first. Think performance at design time, across the four resources (network, disk, memory, CPU) ร two dimensions (bandwidth, latency). Optimize slowest resources first, weighted by frequency.
- Batch everything. Amortize costs. Don't react to external events โ run at your own pace.
- Shrink scope. Declare variables at the smallest possible scope; calculate close to where they're used (avoid POCPOU โ place-of-check to place-of-use gaps).
Full detail in references/principles.md.
Review-Mode Output Format
When reviewing, group findings by severity:
## TigerStyle Review: <target>
### P0 โ Safety violations
- path/to/file.ext:L42 โ Unbounded loop over `items` (no fixed upper bound). Rule: "put a limit on everything".
- path/to/file.ext:L108 โ Function is 94 lines. Rule: "hard 70-line limit". Suggest extracting <specific helper>.
### P1 โ Assertion gaps
- path/to/file.ext:L15 โ `process()` has no precondition assertions on its 3 arguments. Rule: "assert all function arguments".
### P2 โ Naming / DX
- path/to/file.ext:L7 โ `maxLatencyMs` should be `latencyMsMax` (units last, descending significance).
### Notes
- Zig-only rules skipped: <list>
- What's already TigerStyle-compliant: <brief note>
Cite the specific TigerStyle rule for each finding. Don't pad with things that aren't violations.
When NOT to Use
- User is asking about general code quality without invoking TigerStyle โ use
simplify skill or review normally.
- Code is in a framework-heavy context (Spring, React) where TigerStyle rules conflict with framework idioms โ flag the conflict, don't force the rule. Frameworks often require dynamic allocation, recursion, and deep abstractions; note the tradeoff rather than demanding a rewrite.
- Tests with descriptive names that exceed line limits โ the rule targets production code.
