# 📊 Agent Skill Benchmark Report > Generated: 2026-07-10T03:24:45.953Z > Token counting: real cl100k-family tokenizer (`gpt-tokenizer`), chars/4 as fallback only. > Baselines: **synthetic reference instruction-volume bands**, not a measured survey of real prompts (see Methodology). Token/cost figures below measure *skill size*, not *behavioral effectiveness*. > Quality: structural rubric (0–10), no live LLM calls required. For measured with/without-skill behavioral results, see the [Live Evals Report](evals-report.md). ## ❓ How to Read This Report This benchmark answers: **"How much smaller is a skill than a reference band of inline instructions a developer might otherwise write?"** It is a size/structure metric, not a measured behavioral improvement — that is what the [Live Evals Report](evals-report.md) is for. **Reference band (no skill)**: a synthetic stand-in for domain knowledge written directly into the prompt every time. **WITH a skill**: the agent loads the SKILL.md file (avg. 528 tokens this run) — structured, reusable, cached. **Eval–Skill Consistency** (labeled "Aligned" below): % of eval `contains` assertion values that are literal substrings of SKILL.md. This only checks that the skill and its evals were written consistently with each other — it is **not** evidence the skill changes agent behavior. Evals are written from the skill, so near-100% is expected and does not by itself indicate quality. Measured behavioral delta lives in the [Live Evals Report](evals-report.md). ## 🔢 Executive Summary | Metric | Value | | --------------------------------- | --------------------------------- | | Total Skills Benchmarked | **264** | | Avg. Tokens WITH Skill (SKILL.md) | **528 tokens** | | Baseline: Light prompt (no skill) | **529 tokens** ↓ see Methodology | | Baseline: Heavy prompt (no skill) | **986 tokens** ↓ see Methodology | | Avg. Token Savings vs Light | **0%** (1 tokens/call) | | Avg. Token Savings vs Heavy | **46%** (458 tokens/call) | | Avg. Quality Score | **9.8/10** | | Guardrail Skills Covered | **7** | | Avg. Behavior Quality | **2.9/4** (guardrail skills only) | | Skills with Evals | **264 / 264** | | Avg. Eval–Skill Consistency | **99%** (264 skills with `contains` assertions — see caveat above) | ## 🧪 Measured Effectiveness — Live Evals (latest run per category) Unlike everything else in this report, these numbers come from actually running each skill's eval prompts through an agent — see the [Live Evals Report](evals-report.md) and [docs/EVALS.md](docs/EVALS.md) for the full methodology and how to verify or extend this table. | Category | Baseline Pass Rate | With-Skill Pass Rate | Delta | Last Run | | --- | --- | --- | --- | --- | | dart | 11% | 100% | +89% | 2026-07-10 | > No live eval run yet for: `android`, `angular`, `common`, `database`, `flutter`, `golang`, `ios`, `java`, `javascript`, `kotlin`, `laravel`, `nestjs`, `nextjs`, `php`, `python`, `quality-engineering`, `react`, `react-native`, `spring-boot`, `swift`, `typescript`. Run `.agents/workflows/evals-run.md` (or `/evals-run `) to add measured results for these. ## 📜 History | Version | Date | Skills | Avg Tokens | Savings (%) | Quality | Report | | ------- | ---------- | ------ | ---------- | ----------- | ------- | ------ | | v2.6.0 | 2026-07-10 | 264 | 528 | 46% | 9.8/10 | [Full Report](benchmarks/archive/v2.6.0.md) | | v2.4.7 | 2026-06-15 | 251 | 551 | 85% | 9.8/10 | [Full Report](benchmarks/archive/v2.4.7.md) | | v2.4.6 | 2026-06-10 | 251 | 548 | 85% | 9.8/10 | [Full Report](benchmarks/archive/v2.4.6.md) | | v2.4.1 | 2026-05-18 | 247 | 540 | 85% | 9.9/10 | [Full Report](benchmarks/archive/v2.4.1.md) | | v2.4.0 | 2026-05-14 | 246 | 540 | 85% | 9.9/10 | [Full Report](benchmarks/archive/v2.4.0.md) | | v2.3.0 | 2026-05-13 | 246 | 540 | 85% | 9.9/10 | [Full Report](benchmarks/archive/v2.3.0.md) | | v2.2.2 | 2026-05-09 | 249 | 539 | 85% | 9.9/10 | [Full Report](benchmarks/archive/v2.2.2.md) | | v2.2.0 | 2026-04-22 | 242 | 538 | 85% | 9.9/10 | [Full Report](benchmarks/archive/v2.2.0.md) | | v2.1.2 | 2026-04-11 | 237 | 516 | 86% | 10/10 | [Full Report](benchmarks/archive/v2.1.2.md) | | v2.1.1 | 2026-04-11 | 237 | 516 | 86% | 9.9/10 | [Full Report](benchmarks/archive/v2.1.1.md) | | v2.1.0 | 2026-04-04 | 237 | 526 | 86% | 9.9/10 | [Full Report](benchmarks/archive/v2.1.0.md) | | v2.0.1 | 2026-03-30 | 238 | 527 | 86% | 9.8/10 | [Full Report](benchmarks/archive/v2.0.1.md) | | v2.0.0 | 2026-03-25 | 235 | 523 | 86% | 9.9/10 | [Full Report](benchmarks/archive/v2.0.0.md) | | v1.10.3 | 2026-03-21 | 234 | 505 | 86% | 9.8/10 | [Full Report](benchmarks/archive/v1.10.3.md) | | v1.10.1 | 2026-03-16 | 229 | 428 | 88% | 9.9/10 | [Full Report](benchmarks/archive/v1.10.1.md) | | v1.10.0 | 2026-03-16 | 229 | 434 | 88% | 7/10 | [Full Report](benchmarks/archive/v1.10.0.md) | | v1.9.3 | 2026-03-15 | 229 | 460 | 87% | 8.9/10 | [Full Report](benchmarks/archive/v1.9.3.md) | | v1.9.2 | 2026-03-07 | 228 | 458 | 87% | 8.9/10 | [Full Report](benchmarks/archive/v1.9.2.md) | | v1.9.1 | 2026-03-07 | 228 | 458 | 87% | 8.9/10 | [Full Report](benchmarks/archive/v1.9.1.md) | | v1.9.0 | 2026-03-05 | 228 | 457 | 88% | 8.9/10 | [Full Report](benchmarks/archive/v1.9.0.md) | | v1.8.0 | 2026-03-02 | 228 | 443 | 88% | 8.9/10 | [Full Report](benchmarks/archive/v1.8.0.md) | | v1.7.3 | 2026-02-25 | 222 | 418 | 89% | 8.9/10 | [Full Report](benchmarks/archive/v1.7.3.md) | | v1.7.2 | 2026-02-25 | 220 | 413 | 89% | 8.9/10 | [Full Report](benchmarks/archive/v1.7.2.md) | ## 🧾 Metadata Overhead (the cost skills are NOT free) > Skills are not zero-cost to install. Every synced skill's frontmatter (`name` + `description`) is loaded into the session/router context regardless of whether its full body is ever read. This section reports that always-on cost, which the savings figures above do not net out. | Metric | Value | | --- | --- | | Avg. frontmatter tokens per skill | **98 tokens** | | Total frontmatter overhead (all 264 skills registered, paid every session) | **26002 tokens** | | Break-even (skill *uses*, at avg. savings/use, to offset the whole catalog's per-session frontmatter cost) | **~57 use(s)** | > **Prompt caching caveat**: all cost figures in this report price every token at the full input rate. In practice, static context (including skill frontmatter and any skill body loaded early in a session) is frequently served from a prompt cache at a fraction of the input price on providers that support it. Real savings are directionally consistent with this report but smaller in absolute $ than the tables below imply. ### 💰 Cost Comparison — Per Single Call (Average Skill) > Comparison based on the **Heavy reference band** vs. current model pricing. Ignores prompt caching (see caveat above) — treat as an upper bound, not an exact figure. | Model | Original Cost | Skill Cost | Net Savings | % Saved | | ----------------- | ------------- | ---------- | ----------- | ------- | | Gemini 3 Flash | $0.0004930 | $0.0002640 | **$0.0002290** | 46% | | GPT-5 | $0.0012325 | $0.0006600 | **$0.0005725** | 46% | | Gemini 3.1 Pro | $0.0019720 | $0.0010560 | **$0.0009160** | 46% | | Claude Sonnet 4.5 | $0.0029580 | $0.0015840 | **$0.0013740** | 46% | ### 📈 Monthly Savings at Scale — (Avg Skill vs Heavy Reference Band) > Illustrative only: assumes 1,000 calls/day for a single average skill, no prompt caching, and constant token counts. Real savings depend heavily on caching and actual call volume — do not treat this as a budgeting figure. | Daily Calls | Original Cost/mo | Monthly Savings (1 skill) | Model | | ----------- | ---------------- | -------------------------- | ----- | | 1,000 | $36.9750/mo | $17.1750/mo | GPT-5 | | 1,000 | $88.7400/mo | $41.2200/mo | Claude Sonnet 4.5 | | 1,000 | $59.1600/mo | $27.4800/mo | Gemini 3.1 Pro | ## 📦 Per-Category Summary

📦 android (26 skills | avg 404 tokens | quality 9.9/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `android-architecture ` | 562 | ████░░░░░░ 43% | 10/10 | n/a | 3 | ✅ 100% | | `android-background-work` | 279 | ███████░░░ 72% | 10/10 | n/a | 3 | ✅ 100% | | `android-compose ` | 500 | █████░░░░░ 49% | 10/10 | n/a | 3 | ✅ 100% | | `android-compose-migration` | 687 | ███░░░░░░░ 30% | 10/10 | n/a | 3 | ✅ 100% | | `android-concurrency ` | 342 | ███████░░░ 65% | 10/10 | n/a | 3 | ✅ 100% | | `android-deployment ` | 323 | ███████░░░ 67% | 10/10 | n/a | 3 | ✅ 100% | | `android-design-system` | 279 | ███████░░░ 72% | 10/10 | n/a | 3 | ✅ 100% | | `android-di ` | 316 | ███████░░░ 68% | 10/10 | n/a | 3 | ✅ 100% | | `android-edge-to-edge ` | 770 | ██░░░░░░░░ 22% | 10/10 | n/a | 3 | ✅ 100% | | `android-legacy-navigation` | 296 | ███████░░░ 70% | 10/10 | n/a | 3 | ✅ 100% | | `android-legacy-security` | 428 | ██████░░░░ 57% | 10/10 | n/a | 3 | ✅ 100% | | `android-legacy-state ` | 254 | ███████░░░ 74% | 10/10 | n/a | 3 | ✅ 100% | | `android-navigation ` | 259 | ███████░░░ 74% | 10/10 | n/a | 3 | ✅ 100% | | `android-navigation-3 ` | 667 | ███░░░░░░░ 32% | 10/10 | n/a | 3 | ✅ 100% | | `android-navigation-type-safe` | 255 | ███████░░░ 74% | 10/10 | n/a | 3 | ✅ 100% | | `android-notifications` | 388 | ██████░░░░ 61% | 10/10 | n/a | 3 | ✅ 100% | | `android-performance ` | 356 | ██████░░░░ 64% | 10/10 | n/a | 3 | ✅ 100% | | `android-persistence ` | 297 | ███████░░░ 70% | 10/10 | n/a | 3 | ✅ 100% | | `android-resources ` | 434 | ██████░░░░ 56% | 10/10 | n/a | 3 | ✅ 100% | | `android-security ` | 368 | ██████░░░░ 63% | 10/10 | n/a | 3 | ✅ 100% | | `android-state ` | 429 | ██████░░░░ 56% | 10/10 | n/a | 3 | ✅ 100% | | `android-testing ` | 324 | ███████░░░ 67% | 10/10 | n/a | 3 | ✅ 100% | | `android-tooling ` | 311 | ███████░░░ 68% | 10/10 | n/a | 3 | ✅ 100% | | `android-xml-views ` | 304 | ███████░░░ 69% | 10/10 | n/a | 3 | ✅ 100% | | `android-networking ` | 406 | ██████░░░░ 59% | 9/10 | n/a | 3 | ✅ 100% | | `android-agp-upgrade ` | 674 | ███░░░░░░░ 32% | 8/10 | n/a | 3 | ✅ 100% |

📦 angular (15 skills | avg 492 tokens | quality 10.0/10 | eval–skill consistency 96%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `angular-architecture ` | 600 | ████░░░░░░ 39% | 10/10 | n/a | 6 | ✅ 95% | | `angular-components ` | 637 | ████░░░░░░ 35% | 10/10 | n/a | 9 | ✅ 100% | | `angular-dependency-injection` | 500 | █████░░░░░ 49% | 10/10 | n/a | 6 | ✅ 100% | | `angular-directives-pipes` | 493 | █████░░░░░ 50% | 10/10 | n/a | 6 | ✅ 91% | | `angular-forms ` | 321 | ███████░░░ 67% | 10/10 | n/a | 6 | ✅ 100% | | `angular-http-client ` | 556 | ████░░░░░░ 44% | 10/10 | n/a | 6 | ✅ 92% | | `angular-performance ` | 457 | █████░░░░░ 54% | 10/10 | n/a | 6 | ✅ 100% | | `angular-routing ` | 353 | ██████░░░░ 64% | 10/10 | n/a | 6 | ✅ 100% | | `angular-rxjs-interop ` | 493 | █████░░░░░ 50% | 10/10 | n/a | 6 | ✅ 95% | | `angular-security ` | 472 | █████░░░░░ 52% | 10/10 | n/a | 6 | ✅ 86% | | `angular-ssr ` | 490 | █████░░░░░ 50% | 10/10 | n/a | 6 | ✅ 90% | | `angular-state-management` | 387 | ██████░░░░ 61% | 10/10 | n/a | 6 | ✅ 100% | | `angular-style-guide ` | 525 | █████░░░░░ 47% | 10/10 | n/a | 6 | ✅ 100% | | `angular-testing ` | 445 | ██████░░░░ 55% | 10/10 | n/a | 6 | ✅ 100% | | `angular-tooling ` | 655 | ███░░░░░░░ 34% | 10/10 | n/a | 6 | ✅ 95% |

📦 common (38 skills | avg 679 tokens | quality 9.6/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `common-api-design ` | 866 | █░░░░░░░░░ 12% | 10/10 | n/a | 3 | ✅ 100% | | `common-architecture-audit` | 628 | ████░░░░░░ 36% | 10/10 | n/a | 3 | ✅ 100% | | `common-architecture-diagramming` | 472 | █████░░░░░ 52% | 10/10 | n/a | 3 | ✅ 100% | | `common-best-practices` | 425 | ██████░░░░ 57% | 10/10 | n/a | 3 | ✅ 100% | | `common-business-requirements` | 630 | ████░░░░░░ 36% | 10/10 | n/a | 3 | ✅ 100% | | `common-context-optimization` | 513 | █████░░░░░ 48% | 10/10 | n/a | 3 | ✅ 100% | | `common-dast-tooling ` | 958 | ░░░░░░░░░░ 3% | 10/10 | n/a | 3 | ✅ 100% | | `common-documentation ` | 351 | ██████░░░░ 64% | 10/10 | n/a | 3 | ✅ 100% | | `common-feedback-reporter` | 869 | █░░░░░░░░░ 12% | 10/10 | n/a | 4 | ✅ 100% | | `common-git-collaboration` | 523 | █████░░░░░ 47% | 10/10 | n/a | 3 | ✅ 100% | | `common-llm-security ` | 675 | ███░░░░░░░ 32% | 10/10 | n/a | 3 | ✅ 100% | | `common-mobile-animation` | 511 | █████░░░░░ 48% | 10/10 | n/a | 3 | ✅ 100% | | `common-mobile-ux-core` | 408 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `common-observability ` | 416 | ██████░░░░ 58% | 10/10 | n/a | 3 | ✅ 100% | | `common-owasp ` | 1223 | ⚠️ Overhead 24% | 10/10 | n/a | 3 | ✅ 100% | | `common-performance-engineering` | 625 | ████░░░░░░ 37% | 10/10 | n/a | 3 | ✅ 100% | | `common-security-audit` | 960 | ░░░░░░░░░░ 3% | 10/10 | 0/4 | 3 | ✅ 100% | | `common-session-retrospective` | 683 | ███░░░░░░░ 31% | 10/10 | n/a | 3 | ✅ 100% | | `common-skill-creator ` | 1547 | ⚠️ Overhead 57% | 10/10 | 4/4 | 3 | ✅ 100% | | `common-software-requirements` | 657 | ███░░░░░░░ 33% | 10/10 | n/a | 3 | ✅ 100% | | `common-store-changelog` | 664 | ███░░░░░░░ 33% | 10/10 | n/a | 4 | ✅ 100% | | `common-system-design ` | 615 | ████░░░░░░ 38% | 10/10 | n/a | 3 | ✅ 100% | | `common-tdd ` | 842 | ██░░░░░░░░ 15% | 10/10 | 4/4 | 3 | ✅ 100% | | `common-ui-design ` | 693 | ███░░░░░░░ 30% | 10/10 | n/a | 3 | ✅ 100% | | `common-workflow-writing` | 579 | ████░░░░░░ 41% | 10/10 | 0/4 | 4 | ✅ 100% | | `common-accessibility ` | 1075 | ⚠️ Overhead 9% | 9/10 | n/a | 3 | ✅ 100% | | `common-code-review ` | 518 | █████░░░░░ 47% | 9/10 | 4/4 | 3 | ✅ 100% | | `common-debugging ` | 397 | ██████░░░░ 60% | 9/10 | 4/4 | 3 | ✅ 100% | | `common-error-handling` | 400 | ██████░░░░ 59% | 9/10 | n/a | 3 | ✅ 100% | | `common-exploit-verification` | 742 | ███░░░░░░░ 25% | 9/10 | n/a | 2 | ✅ 100% | | `common-learning-log ` | 505 | █████░░░░░ 49% | 9/10 | n/a | 3 | ✅ 100% | | `common-mobile-visual-testing` | 609 | ████░░░░░░ 38% | 9/10 | n/a | 2 | ✅ 100% | | `common-pentest-methodology` | 1068 | ⚠️ Overhead 8% | 9/10 | n/a | 2 | ✅ 100% | | `common-product-requirements` | 909 | █░░░░░░░░░ 8% | 9/10 | n/a | 5 | ✅ 100% | | `common-security-standards` | 665 | ███░░░░░░░ 33% | 9/10 | n/a | 3 | ✅ 100% | | `common-telemetry ` | 541 | █████░░░░░ 45% | 9/10 | n/a | 2 | ✅ 100% | | `common-web-visual-testing` | 551 | ████░░░░░░ 44% | 9/10 | n/a | 2 | ✅ 100% | | `common-protocol-enforcement` | 492 | █████░░░░░ 50% | 8/10 | 4/4 | 3 | ✅ 100% |

📦 dart (3 skills | avg 565 tokens | quality 9.3/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `dart-best-practices ` | 490 | █████░░░░░ 50% | 10/10 | n/a | 3 | ✅ 100% | | `dart-tooling ` | 498 | █████░░░░░ 49% | 10/10 | n/a | 3 | ✅ 100% | | `dart-language ` | 708 | ███░░░░░░░ 28% | 8/10 | n/a | 3 | ✅ 100% |

📦 database (7 skills | avg 355 tokens | quality 9.4/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `database-mongodb ` | 405 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `database-postgresql ` | 382 | ██████░░░░ 61% | 10/10 | n/a | 3 | ✅ 100% | | `database-redis ` | 453 | █████░░░░░ 54% | 10/10 | n/a | 3 | ✅ 100% | | `database-migrations ` | 312 | ███████░░░ 68% | 9/10 | n/a | 1 | ✅ 100% | | `database-query-performance` | 321 | ███████░░░ 67% | 9/10 | n/a | 1 | ✅ 100% | | `database-schema-design` | 331 | ███████░░░ 66% | 9/10 | n/a | 1 | ✅ 100% | | `database-transactions` | 283 | ███████░░░ 71% | 9/10 | n/a | 1 | ✅ 100% |

📦 flutter (22 skills | avg 521 tokens | quality 9.6/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `flutter-cicd ` | 566 | ████░░░░░░ 43% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-concurrency ` | 639 | ████░░░░░░ 35% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-design-system` | 559 | ████░░░░░░ 43% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-error-handling` | 563 | ████░░░░░░ 43% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-feature-based-clean-architecture` | 507 | █████░░░░░ 49% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-getx-navigation` | 335 | ███████░░░ 66% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-getx-state-management` | 471 | █████░░░░░ 52% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-go-router-navigation` | 555 | ████░░░░░░ 44% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-idiomatic-flutter` | 439 | ██████░░░░ 55% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-layer-based-clean-architecture` | 604 | ████░░░░░░ 39% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-performance ` | 468 | █████░░░░░ 53% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-retrofit-networking` | 547 | █████░░░░░ 45% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-riverpod-state-management` | 563 | ████░░░░░░ 43% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-widgets ` | 500 | █████░░░░░ 49% | 10/10 | n/a | 3 | ✅ 100% | | `flutter-auto-route-navigation` | 461 | █████░░░░░ 53% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-bloc-state-management` | 708 | ███░░░░░░░ 28% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-dependency-injection` | 467 | █████░░░░░ 53% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-localization ` | 456 | █████░░░░░ 54% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-navigation ` | 379 | ██████░░░░ 62% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-notifications` | 342 | ███████░░░ 65% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-security ` | 443 | ██████░░░░ 55% | 9/10 | n/a | 3 | ✅ 100% | | `flutter-testing ` | 883 | █░░░░░░░░░ 10% | 8/10 | n/a | 3 | ✅ 100% |

📦 golang (11 skills | avg 461 tokens | quality 9.7/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `golang-api-server ` | 437 | ██████░░░░ 56% | 10/10 | n/a | 3 | ✅ 100% | | `golang-architecture ` | 470 | █████░░░░░ 52% | 10/10 | n/a | 3 | ✅ 100% | | `golang-concurrency ` | 409 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `golang-configuration ` | 401 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `golang-database ` | 460 | █████░░░░░ 53% | 10/10 | n/a | 3 | ✅ 100% | | `golang-language ` | 550 | ████░░░░░░ 44% | 10/10 | n/a | 3 | ✅ 100% | | `golang-security ` | 540 | █████░░░░░ 45% | 10/10 | n/a | 3 | ✅ 100% | | `golang-tooling ` | 598 | ████░░░░░░ 39% | 10/10 | n/a | 3 | ✅ 100% | | `golang-error-handling` | 372 | ██████░░░░ 62% | 9/10 | n/a | 3 | ✅ 100% | | `golang-logging ` | 402 | ██████░░░░ 59% | 9/10 | n/a | 3 | ✅ 100% | | `golang-testing ` | 436 | ██████░░░░ 56% | 9/10 | n/a | 3 | ✅ 100% |

📦 ios (15 skills | avg 376 tokens | quality 10.0/10 | eval–skill consistency 97%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `ios-app-lifecycle ` | 319 | ███████░░░ 68% | 10/10 | n/a | 3 | ✅ 100% | | `ios-architecture ` | 610 | ████░░░░░░ 38% | 10/10 | n/a | 3 | ✅ 100% | | `ios-dependency-injection` | 286 | ███████░░░ 71% | 10/10 | n/a | 3 | ✅ 100% | | `ios-deployment ` | 330 | ███████░░░ 67% | 10/10 | n/a | 3 | ✅ 88% | | `ios-design-system ` | 262 | ███████░░░ 73% | 10/10 | n/a | 3 | ✅ 100% | | `ios-localization ` | 359 | ██████░░░░ 64% | 10/10 | n/a | 3 | ✅ 100% | | `ios-navigation ` | 278 | ███████░░░ 72% | 10/10 | n/a | 3 | ✅ 100% | | `ios-networking ` | 388 | ██████░░░░ 61% | 10/10 | n/a | 3 | ✅ 100% | | `ios-notifications ` | 269 | ███████░░░ 73% | 10/10 | n/a | 3 | ✅ 100% | | `ios-performance ` | 388 | ██████░░░░ 61% | 10/10 | n/a | 3 | ✅ 100% | | `ios-persistence ` | 363 | ██████░░░░ 63% | 10/10 | n/a | 3 | ✅ 83% | | `ios-security ` | 423 | ██████░░░░ 57% | 10/10 | n/a | 3 | ✅ 89% | | `ios-state-management ` | 362 | ██████░░░░ 63% | 10/10 | n/a | 3 | ✅ 100% | | `ios-swiftui ` | 593 | ████░░░░░░ 40% | 10/10 | n/a | 3 | ✅ 100% | | `ios-ui-navigation ` | 404 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% |

📦 java (5 skills | avg 494 tokens | quality 10.0/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `java-best-practices ` | 436 | ██████░░░░ 56% | 10/10 | n/a | 3 | ✅ 100% | | `java-concurrency ` | 407 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `java-language ` | 523 | █████░░░░░ 47% | 10/10 | n/a | 3 | ✅ 100% | | `java-testing ` | 580 | ████░░░░░░ 41% | 10/10 | n/a | 3 | ✅ 100% | | `java-tooling ` | 523 | █████░░░░░ 47% | 10/10 | n/a | 3 | ✅ 100% |

📦 javascript (3 skills | avg 389 tokens | quality 10.0/10 | eval–skill consistency 97%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `javascript-best-practices` | 305 | ███████░░░ 69% | 10/10 | n/a | 3 | ✅ 100% | | `javascript-language ` | 462 | █████░░░░░ 53% | 10/10 | n/a | 3 | ✅ 90% | | `javascript-tooling ` | 401 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% |

📦 kotlin (4 skills | avg 412 tokens | quality 10.0/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `kotlin-best-practices` | 439 | ██████░░░░ 55% | 10/10 | n/a | 3 | ✅ 100% | | `kotlin-coroutines ` | 405 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `kotlin-language ` | 449 | █████░░░░░ 54% | 10/10 | n/a | 3 | ✅ 100% | | `kotlin-tooling ` | 354 | ██████░░░░ 64% | 10/10 | n/a | 3 | ✅ 100% |

📦 laravel (10 skills | avg 644 tokens | quality 10.0/10 | eval–skill consistency 97%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `laravel-api ` | 700 | ███░░░░░░░ 29% | 10/10 | n/a | 6 | ✅ 100% | | `laravel-architecture ` | 347 | ███████░░░ 65% | 10/10 | n/a | 6 | ✅ 100% | | `laravel-background-processing` | 639 | ████░░░░░░ 35% | 10/10 | n/a | 6 | ✅ 91% | | `laravel-clean-architecture` | 635 | ████░░░░░░ 36% | 10/10 | n/a | 6 | ✅ 100% | | `laravel-database-expert` | 724 | ███░░░░░░░ 27% | 10/10 | n/a | 6 | ✅ 100% | | `laravel-eloquent ` | 612 | ████░░░░░░ 38% | 10/10 | n/a | 6 | ✅ 100% | | `laravel-security ` | 718 | ███░░░░░░░ 27% | 10/10 | n/a | 6 | ✅ 95% | | `laravel-sessions-middleware` | 648 | ███░░░░░░░ 34% | 10/10 | n/a | 6 | ✅ 90% | | `laravel-testing ` | 717 | ███░░░░░░░ 27% | 10/10 | n/a | 6 | ✅ 95% | | `laravel-tooling ` | 697 | ███░░░░░░░ 29% | 10/10 | n/a | 6 | ✅ 100% |

📦 nestjs (21 skills | avg 611 tokens | quality 9.7/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `nestjs-api-standards ` | 556 | ████░░░░░░ 44% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-architecture ` | 579 | ████░░░░░░ 41% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-bullmq ` | 946 | ░░░░░░░░░░ 4% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-caching ` | 584 | ████░░░░░░ 41% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-controllers-services` | 697 | ███░░░░░░░ 29% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-database ` | 618 | ████░░░░░░ 37% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-deployment ` | 693 | ███░░░░░░░ 30% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-file-uploads ` | 419 | ██████░░░░ 58% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-notification ` | 495 | █████░░░░░ 50% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-observability ` | 448 | ██████░░░░ 55% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-performance ` | 937 | █░░░░░░░░░ 5% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-real-time ` | 847 | █░░░░░░░░░ 14% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-scheduling ` | 538 | █████░░░░░ 45% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-search ` | 488 | █████░░░░░ 51% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-security ` | 825 | ██░░░░░░░░ 16% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-security-isolation` | 512 | █████░░░░░ 48% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-testing ` | 600 | ████░░░░░░ 39% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-transport ` | 408 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `nestjs-documentation ` | 521 | █████░░░░░ 47% | 9/10 | n/a | 3 | ✅ 100% | | `nestjs-error-handling` | 555 | ████░░░░░░ 44% | 8/10 | n/a | 3 | ✅ 100% | | `nestjs-configuration ` | 557 | ████░░░░░░ 44% | 7/10 | n/a | 3 | ✅ 100% |

📦 nextjs (18 skills | avg 630 tokens | quality 9.8/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `nextjs-app-router ` | 1019 | ⚠️ Overhead 3% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-architecture ` | 1039 | ⚠️ Overhead 5% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-authentication` | 465 | █████░░░░░ 53% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-caching ` | 644 | ████░░░░░░ 35% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-data-access-layer` | 486 | █████░░░░░ 51% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-data-fetching ` | 494 | █████░░░░░ 50% | 10/10 | n/a | 6 | ✅ 100% | | `nextjs-i18n ` | 583 | ████░░░░░░ 41% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-optimization ` | 516 | █████░░░░░ 48% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-rendering ` | 750 | ██░░░░░░░░ 24% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-security ` | 658 | ███░░░░░░░ 33% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-state-management` | 595 | ████░░░░░░ 40% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-testing ` | 658 | ███░░░░░░░ 33% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-tooling ` | 406 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-upgrade ` | 583 | ████░░░░░░ 41% | 10/10 | n/a | 3 | ✅ 100% | | `nextjs-pages-router ` | 676 | ███░░░░░░░ 31% | 9/10 | n/a | 6 | ✅ 100% | | `nextjs-server-actions` | 532 | █████░░░░░ 46% | 9/10 | n/a | 6 | ✅ 100% | | `nextjs-server-components` | 578 | ████░░░░░░ 41% | 9/10 | n/a | 6 | ✅ 100% | | `nextjs-styling ` | 663 | ███░░░░░░░ 33% | 9/10 | n/a | 3 | ✅ 100% |

📦 php (7 skills | avg 537 tokens | quality 9.4/10 | eval–skill consistency 96%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `php-best-practices ` | 518 | █████░░░░░ 47% | 10/10 | n/a | 6 | ✅ 92% | | `php-security ` | 624 | ████░░░░░░ 37% | 10/10 | n/a | 6 | ✅ 100% | | `php-tooling ` | 636 | ████░░░░░░ 35% | 10/10 | n/a | 6 | ✅ 92% | | `php-concurrency ` | 515 | █████░░░░░ 48% | 9/10 | n/a | 6 | ✅ 100% | | `php-error-handling ` | 455 | █████░░░░░ 54% | 9/10 | n/a | 6 | ✅ 100% | | `php-language ` | 464 | █████░░░░░ 53% | 9/10 | n/a | 6 | ✅ 91% | | `php-testing ` | 546 | █████░░░░░ 45% | 9/10 | n/a | 6 | ✅ 100% |

📦 python (9 skills | avg 353 tokens | quality 9.0/10 | eval–skill consistency 97%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `python-architecture ` | 348 | ███████░░░ 65% | 9/10 | n/a | 2 | ✅ 100% | | `python-async-runtime ` | 329 | ███████░░░ 67% | 9/10 | n/a | 2 | ✅ 100% | | `python-best-practices` | 311 | ███████░░░ 68% | 9/10 | n/a | 2 | ✅ 75% | | `python-database ` | 331 | ███████░░░ 66% | 9/10 | n/a | 2 | ✅ 100% | | `python-error-handling` | 292 | ███████░░░ 70% | 9/10 | n/a | 2 | ✅ 100% | | `python-language ` | 430 | ██████░░░░ 56% | 9/10 | n/a | 2 | ✅ 100% | | `python-security ` | 335 | ███████░░░ 66% | 9/10 | n/a | 2 | ✅ 100% | | `python-testing ` | 383 | ██████░░░░ 61% | 9/10 | n/a | 2 | ✅ 100% | | `python-tooling ` | 414 | ██████░░░░ 58% | 9/10 | n/a | 2 | ✅ 100% |

📦 quality-engineering (7 skills | avg 730 tokens | quality 9.7/10 | eval–skill consistency 99%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `quality-engineering-business-analysis` | 998 | ⚠️ Overhead 1% | 10/10 | n/a | 6 | ✅ 95% | | `quality-engineering-jira-integration` | 574 | ████░░░░░░ 42% | 10/10 | n/a | 3 | ✅ 100% | | `quality-engineering-quality-assurance` | 486 | █████░░░░░ 51% | 10/10 | n/a | 3 | ✅ 100% | | `quality-engineering-zephyr-coverage-analysis` | 462 | █████░░░░░ 53% | 10/10 | n/a | 4 | ✅ 100% | | `quality-engineering-zephyr-test-generation` | 1170 | ⚠️ Overhead 19% | 10/10 | n/a | 3 | ✅ 100% | | `quality-engineering-appium-mcp` | 740 | ███░░░░░░░ 25% | 9/10 | n/a | 2 | ✅ 100% | | `quality-engineering-playwright-cli` | 683 | ███░░░░░░░ 31% | 9/10 | n/a | 2 | ✅ 100% |

📦 react (8 skills | avg 558 tokens | quality 9.8/10 | eval–skill consistency 100%)

📦 react-native (13 skills | avg 438 tokens | quality 10.0/10 | eval–skill consistency 98%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `react-native-architecture` | 500 | █████░░░░░ 49% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-components` | 367 | ██████░░░░ 63% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-deployment` | 552 | ████░░░░░░ 44% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-dls ` | 288 | ███████░░░ 71% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-navigation` | 400 | ██████░░░░ 59% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-navigation-v6` | 561 | ████░░░░░░ 43% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-notifications` | 313 | ███████░░░ 68% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-performance` | 590 | ████░░░░░░ 40% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-platform-specific` | 376 | ██████░░░░ 62% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-security` | 543 | █████░░░░░ 45% | 10/10 | n/a | 3 | ✅ 78% | | `react-native-state-management` | 429 | ██████░░░░ 56% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-styling ` | 324 | ███████░░░ 67% | 10/10 | n/a | 3 | ✅ 100% | | `react-native-testing ` | 448 | ██████░░░░ 55% | 10/10 | n/a | 3 | ✅ 100% |

📦 spring-boot (10 skills | avg 462 tokens | quality 9.9/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `spring-boot-api-design` | 331 | ███████░░░ 66% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-architecture` | 598 | ████░░░░░░ 39% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-best-practices` | 553 | ████░░░░░░ 44% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-data-access` | 538 | █████░░░░░ 45% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-deployment` | 517 | █████░░░░░ 48% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-microservices` | 443 | ██████░░░░ 55% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-observability` | 460 | █████░░░░░ 53% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-scheduling` | 336 | ███████░░░ 66% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-security ` | 516 | █████░░░░░ 48% | 10/10 | n/a | 3 | ✅ 100% | | `spring-boot-testing ` | 328 | ███████░░░ 67% | 9/10 | n/a | 3 | ✅ 100% |

📦 swift (8 skills | avg 480 tokens | quality 9.9/10 | eval–skill consistency 97%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `swift-best-practices ` | 719 | ███░░░░░░░ 27% | 10/10 | n/a | 4 | ✅ 92% | | `swift-concurrency ` | 522 | █████░░░░░ 47% | 10/10 | n/a | 5 | ✅ 93% | | `swift-error-handling ` | 496 | █████░░░░░ 50% | 10/10 | n/a | 4 | ✅ 100% | | `swift-language ` | 457 | █████░░░░░ 54% | 10/10 | n/a | 5 | ✅ 94% | | `swift-memory-management` | 387 | ██████░░░░ 61% | 10/10 | n/a | 4 | ✅ 100% | | `swift-testing ` | 441 | ██████░░░░ 55% | 10/10 | n/a | 4 | ✅ 100% | | `swift-tooling ` | 388 | ██████░░░░ 61% | 10/10 | n/a | 4 | ✅ 100% | | `swift-swiftui ` | 430 | ██████░░░░ 56% | 9/10 | n/a | 4 | ✅ 100% |

📦 typescript (4 skills | avg 704 tokens | quality 10.0/10 | eval–skill consistency 100%)

| Skill | Tokens | Savings (vs Heavy) | Quality | Behavior | Evals | Aligned | | ----------------------- | ------ | ------------------ | ------- | -------- | ----- | ------- | | `typescript-best-practices` | 625 | ████░░░░░░ 37% | 10/10 | n/a | 3 | ✅ 100% | | `typescript-language ` | 691 | ███░░░░░░░ 30% | 10/10 | n/a | 3 | ✅ 100% | | `typescript-security ` | 787 | ██░░░░░░░░ 20% | 10/10 | n/a | 5 | ✅ 100% | | `typescript-tooling ` | 714 | ███░░░░░░░ 28% | 10/10 | n/a | 3 | ✅ 100% |

## ⚠️ Guardrail Skills Missing Behavior Coverage > These skills enforce behavior but do not yet cover enough pressure scenarios, rationalizations, red flags, or behavior assertions. | Skill | Category | Behavior | Action | | ----- | -------- | -------- | ------ | | `common-security-audit` | common | 0/4 | Add pressure_scenarios, rationalizations, red_flags, and behavior_assertions | | `common-workflow-writing` | common | 0/4 | Add pressure_scenarios, rationalizations, red_flags, and behavior_assertions | ## 📊 Quality Distribution > Averages hide saturation. This shows how many skills actually sit at each score, so a 9.8/10 average can be read in context. | Score | Count | Share | | --- | --- | --- | | 10/10 | 207 | ████████████░░░ 78% | | 9/10 | 51 | ███░░░░░░░░░░░░ 19% | | 8/10 | 5 | ░░░░░░░░░░░░░░░ 2% | | 7/10 | 1 | ░░░░░░░░░░░░░░░ 0% | ## 🔧 Needs Attention > Actionable, not celebratory: lowest structural quality, largest token footprint, and guardrail skills with the weakest behavior coverage. (Guardrail gaps are also listed in full above, under "Guardrail Skills Missing Behavior Coverage".) **Lowest structural quality:** | Skill | Category | Quality | Tokens | Evals | Consistency | | --- | --- | --- | --- | --- | --- | | `nestjs-configuration` | nestjs | 7/10 | 557 | 3 | ✅ 100% | | `android-agp-upgrade` | android | 8/10 | 674 | 3 | ✅ 100% | | `common-protocol-enforcement` | common | 8/10 | 492 | 3 | ✅ 100% | | `dart-language` | dart | 8/10 | 708 | 3 | ✅ 100% | | `flutter-testing` | flutter | 8/10 | 883 | 3 | ✅ 100% | | `nestjs-error-handling` | nestjs | 8/10 | 555 | 3 | ✅ 100% | | `android-networking` | android | 9/10 | 406 | 3 | ✅ 100% | | `common-accessibility` | common | 9/10 | 1075 | 3 | ✅ 100% | | `common-code-review` | common | 9/10 | 518 | 3 | ✅ 100% | | `common-debugging` | common | 9/10 | 397 | 3 | ✅ 100% | **Largest token footprint** (candidates for splitting or trimming): | Skill | Category | Tokens | Quality | | --- | --- | --- | --- | | `common-skill-creator` | common | 1547 | 10/10 | | `common-owasp` | common | 1223 | 10/10 | | `quality-engineering-zephyr-test-generation` | quality-engineering | 1170 | 10/10 | | `common-accessibility` | common | 1075 | 9/10 | | `common-pentest-methodology` | common | 1068 | 9/10 | | `nextjs-architecture` | nextjs | 1039 | 10/10 | | `nextjs-app-router` | nextjs | 1019 | 10/10 | | `quality-engineering-business-analysis` | quality-engineering | 998 | 10/10 | | `common-security-audit` | common | 960 | 10/10 | | `common-dast-tooling` | common | 958 | 10/10 | ## 📐 Methodology & Baseline Justification ### Why These Baselines? (and what they do NOT prove) The baselines are **synthetic reference prompts**, hand-written once and token-counted — they are not a measured survey or average of real developer prompts. They exist to give "tokens saved" a stable unit of comparison across a catalog of 264+ skills spanning many stacks; they do not claim any individual skill was benchmarked against what a specific developer would have typed. NestJS was picked as the **Reference Unit** purely because it is a high-density, well-documented stack — not because every skill category resembles NestJS. Savings % should be read as "SKILL.md is this much smaller than a reference instruction-volume band", not as "this skill saves X% of what a developer would otherwise write for this specific framework." For a measured (not synthetic) with/without-skill effectiveness signal, see the [Live Evals Report](evals-report.md), which runs each skill's eval prompts through an agent twice — once without the skill, once with it — and scores the transcripts deterministically. #### 🟡 Reference Prompt — Light — 529 tokens (real-tokenizer count, no padding) > **Synthetic Reference Prompt — Light (e.g., NestJS)** > A compact inline system prompt used as a reference instruction-volume band. Representative of focused developer instructions without a structured skill. Not a measured average of real prompts. #### 🔴 Reference Prompt — Heavy — 986 tokens (real-tokenizer count, no padding) > **Synthetic Reference Prompt — Heavy (e.g., NestJS Architecture)** > A comprehensive architect-level inline prompt used as a reference instruction-volume band. Includes deep patterns and rules a developer might send when no skill is present. Not a measured average of real prompts. ### 🏆 Detailed Quality Rubric (0–10) To ensure skills are not just "short" but actually **high quality**, every skill is scored against this structural rubric: | Score | Criteria | Rationale | | ------ | ------------------------- | ------------------------------------------------------ | | **+2** | **Structured Guidelines** | At least 3 specific instructions/bullet points. | | **+2** | **Anti-Patterns** | `## Anti-Patterns` section or `**No X**` inline lines. | | **+2** | **Reference Examples** | `references/*.md` links resolved on disk (existing + non-empty) — OR ≤60 lines total. Dangling links score 0 here even if present in text. | | **+2** | **Token Optimality** | Entire `SKILL.md` is ≤100 lines (forces brevity). | | **+2** | **Eval Coverage** | ≥3 evals with `should_not_trigger`, ≥2 assertions each. +1 partial.| > **Eval–Skill Consistency** (reported separately, not scored): % of eval `contains` assertion values that are literal substrings of SKILL.md content. Because evals are typically authored from the skill, near-100% is the expected baseline, not evidence of quality — it only flags drift between a skill and its own evals. It is **not** a proxy for with-skill vs. without-skill behavioral improvement; that requires actually running the evals (see [Live Evals Report](evals-report.md)). > **Behavior Quality** (reported separately): guardrail-only score for pressure scenarios, rationalizations, red flags, and behavior assertions. ### 🛡️ How to Verify This Report Trust but verify. Every number above is reproducible from source — nothing here requires taking our word for it: 1. **Clone the repo** and install dependencies (`pnpm install`). 2. **Inspect source**: the benchmark logic is open in [scripts/benchmark/](scripts/benchmark/) (`utils.ts` for the quality rubric, `baselines.ts` for the reference prompts, `reporter.ts` for how this file is generated). 3. **Regenerate this report**: `pnpm benchmark:report` — diff the output against this file; it should match modulo the `Generated:` timestamp. 4. **For measured (non-structural) behavioral results**, see [docs/EVALS.md](docs/EVALS.md) for how to run and verify the Live Evals Report yourself, including via `pnpm evals:verify` or the MCP `verify_eval_run` tool — no API key required. ### Pricing (per 1M input tokens, Feb 2026) > Pricing drifts. Verify current rates with each provider before using these figures for budgeting. - **Gemini 3 Flash**: $0.50 - **GPT-5**: $1.25 - **Gemini 3.1 Pro**: $2.00 - **Claude Sonnet 4.5**: $3.00