Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Expert skill for binary and text serialization formats, schema design, and optimization
.claude/skills/a5c-ai-serialization/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-07 | ✓→✗ | ▼ Worse | 49% | 0% |
| case-02 | ✓→✓ | = Same ✓ | -3% | 0% |
| case-03 | ✓→✓ | = Same ✓ | 23% | 0% |
Expert skill for binary and text serialization formats, schema design, and performance optimization.
protoc - Protocol Buffer compilerflatc - FlatBuffers compilermsgpack-tools - MessagePack utilitiescbor-tools - CBOR utilitiescapnp - Cap'n Proto compileravro-tools - Apache Avro utilitiesprotobufsyntax = "proto3"; package network; message Packet { uint32 sequence = 1; bytes payload = 2; int64 timestamp = 3; map<string, string> headers = 4; }
fbsnamespace Network; table Packet { sequence: uint32; payload: [ubyte]; timestamp: int64; headers: [KeyValue]; } table KeyValue { key: string; value: string; } root_type Packet;
bashprotoc --python_out=. --go_out=. packet.proto flatc --python --go packet.fbs
bashhyperfine 'protoc-bench encode message.proto' 'msgpack-bench encode message.json'
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 21,687 | 33,020 | +52% | 1 | 1 | 0% | 3,916 | 4,695 | +20% | 0 | 0 | — |
case-02 | pass→pass | 22,010 | 16,895 | -23% | 1 | 1 | 0% | 3,749 | 3,636 | -3% | 0 | 0 | — |
case-03 | pass→pass | 6,687 | 5,829 | -13% | 1 | 1 | 0% | 1,237 | 1,526 | +23% | 0 | 0 | — |
case-04 | pass→pass | 3,446 | 2,826 | -18% | 1 | 1 | 0% | 615 | 1,000 | +63% | 0 | 0 | — |
case-05 | pass→pass | 11,348 | 12,068 | +6% | 1 | 1 | 0% | 2,046 | 2,794 | +37% | 0 | 0 | — |
case-06 | fail→pass | 18,083 | 20,324 | +12% | 1 | 1 | 0% | 2,888 | 3,729 | +29% | 0 | 0 | — |
case-07 | pass→fail | 15,204 | 22,325 | +47% | 1 | 1 | 0% | 2,477 | 3,693 | +49% | 0 | 0 | — |
case-08 | pass→pass | 10,437 | 10,806 | +4% | 1 | 1 | 0% | 1,896 | 2,418 | +28% | 0 | 0 | — |
case-09 | pass→pass | 9,445 | 12,979 | +37% | 1 | 1 | 0% | 1,688 | 2,531 | +50% | 0 | 0 | — |
case-10 | pass→pass | 5,645 | 5,762 | +2% | 1 | 1 | 0% | 1,042 | 1,522 | +46% | 0 | 0 | — |
case-11 | pass→pass | 20,603 | 23,362 | +13% | 1 | 1 | 0% | 3,475 | 4,088 | +18% | 0 | 0 | — |
case-12 | pass→pass | 22,066 | 23,005 | +4% | 1 | 1 | 0% | 3,932 | 5,077 | +29% | 0 | 0 | — |
case-13 | fail→fail | 19,713 | 21,775 | +10% | 1 | 1 | 0% | 3,030 | 3,876 | +28% | 0 | 0 | — |
case-14 | pass→pass | 18,962 | 17,955 | -5% | 1 | 1 | 0% | 2,845 | 3,750 | +32% | 0 | 0 | — |
case-15 | pass→pass | 4,101 | 4,918 | +20% | 1 | 1 | 0% | 734 | 1,225 | +67% | 0 | 0 | — |
case-16 | pass→pass | 17,279 | 20,088 | +16% | 1 | 1 | 0% | 3,207 | 4,369 | +36% | 0 | 0 | — |
case-17 | pass→pass | 9,700 | 11,291 | +16% | 1 | 1 | 0% | 1,784 | 2,143 | +20% | 0 | 0 | — |
case-18 | pass→pass | 15,774 | 20,705 | +31% | 1 | 1 | 0% | 2,686 | 3,797 | +41% | 0 | 0 | — |
case-19 | pass→pass | 9,826 | 13,187 | +34% | 1 | 1 | 0% | 1,763 | 2,511 | +42% | 0 | 0 | — |
case-20 | pass→pass | 18,179 | 20,250 | +11% | 1 | 1 | 0% | 3,100 | 3,973 | +28% | 0 | 0 | — |
case-21 | pass→pass | 18,792 | 15,490 | -18% | 1 | 1 | 0% | 3,877 | 3,685 | -5% | 0 | 0 | — |
case-22 | pass→pass | 19,004 | 22,890 | +20% | 1 | 1 | 0% | 3,116 | 4,201 | +35% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +5 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.