Przejdź do treści
Biblioteka skilli
AI i agenci · Budowanie i kod

common-skill-creator

/common-skill-creator

Tworzy i ocenia skille dla agentów AI, sprawdzając trafność wyzwalania, zwięzłość i redundancję w pliku SKILL.md.

Licencja
MIT
Ostatnia zmiana repozytorium
12.11.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
13
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Tworzy i ocenia skille dla agentów AI, sprawdzając trafność wyzwalania, zwięzłość i redundancję w pliku SKILL.md.

Oryginalny opis (EN)

Standardizes the creation and evaluation of high-density Agent Skills (Claude, Cursor, Windsurf). Ensures skills achieve high Activation (specificity/completeness) and Implementation (conciseness/actionability) scores. Use when: writing or auditing SKILL.md, improving trigger accuracy, or refactoring skills to reduce redundancy and maximize token ROI.

Jak działa

  1. Priority: P0 (CRITICAL)
  2. Three-Level Loading System
  3. Workflow (New or Existing Skill)
  4. Description Quality (Activation)
  5. Content Quality (Implementation)
  6. Evidence Tiers & Guardrails
  7. Anti-Patterns
  8. Quality Checklist (Tessl-Aligned)
  9. References
Sekcje z nagłówków SKILL.md, w oryginale · 85 linii, 905 słów
Podgląd SKILL.md
---
name: common-skill-creator
description: "Standardizes the creation and evaluation of high-density Agent Skills (Claude, Cursor, Windsurf). Ensures skills achieve high Activation (specificity/completeness) and Implementation (conciseness/actionability) scores. Use when: writing or auditing SKILL.md, improving trigger accuracy, or refactoring skills to reduce redundancy and maximize token ROI."
metadata:
  triggers:
    files:
      - "SKILL.md"
      - "evals/evals.json"
    keywords:
      - create skill
      - audit skill
      - trigger rate
      - optimize description
---

# Agent Skill Creator Standard

## **Priority: P0 (CRITICAL)**

Applies to **every skill in this registry**. Maximize **Token ROI**. Every line in SKILL.md must provide specific procedural value. **Activation** (how it triggers) and **Implementation** (how it helps) primary quality metrics.

## Three-Level Loading System

- **Level 1** Frontmatter: `name` + `description` (Activation Anchor), ≤100 words.
- **Level 2** SKILL.md body: Core Rules + Workflows (Implementation Core), ≤100 lines.
- **Level 3** references/: Detailed examples, schemas, and "TESTS.md" (On-demand).

## Workflow (New or Existing Skill)
**New skill:**
1. **Research** — web-search domain best practices, checklists, and standards; extract key terms → triggers, workflows → guidelines, mistakes → anti-patterns. See [Web Search Research](references/web-search-research.md).
2. **Capture intent** — what it , when it trigger, expected output format?
3. **Draft the SKILL.md** — draft using [TEMPLATE.md](references/TEMPLATE.md)
4. **Test** — spawn parallel subagents: one with-skill, one without-skill (baseline)
5. **Evaluate** — grade assertions, review benchmark (pass rate, tokens, time)
6. **Iterate** — rewrite based on feedback, rerun into next iteration dir, repeat
7. **Optimize description** — run trigger eval queries, target ≥90% accuracy (matching release readiness)
8. **Pressure-test** — for discipline skills, capture agent excuses, red flags, and stop conditions
   **Existing skill:**
9. **Audit** — run Quality Checklist below; identify violations
10. **Snapshot** — `cp -r <skill-dir> <workspace>/skill-snapshot/` before any edits
11. **Improve SKILL.md** — fix violations, compress, move oversized content to `references/`
12. **Test** — spawn parallel subagents: one with-new-skill, one with-snapshot (baseline)
13. **Evaluate & iterate** — same as steps 4–5 above
14. **Optimize description** — re-run trigger eval if description changed
15. **Harden** — add rationalization counters where agents still fail under pressure
    See [Eval Workflow](references/eval-workflow.md) for full testing + iteration details.

## Description Quality (Activation)

- **Third-Person Voice**: Use `Standardizes...`, `Audits...`, `Encrypts...`. Avoid "I will" or "This skill helps to".
- **What + When Structure**:
- **What**: Define 5–8 specific capabilities (e.g., "Generates JWT tokens, rotates keys").
- **When**: Explicitly define triggers (e.g., "Use when user says 'rotate keys'").
- **Specificity**: Avoid vague verbs like "manage" or "handle". Use "Validate", "Inject", "Refactor", "Sanitize".
- **Trigger Hint**: Include `(triggers: *.ext, keyword)` suffix for technical skills.
## Content Quality (Implementation)

- **No Redundant Knowledge**: Do NOT explain concepts AI already knows (e.g., standard HTTP codes, common language syntax, basic SOLID definitions). Focus strictly on project-specific rules and constraints.
- **Readable Compact Style**: Size is an editorial budget, not a behavioral quality claim. Prefer readable, compact, imperative language over filler words, conversational phrasing, or speculative prose.
- **Actionability**: Examples must be copy-paste ready and executable.
- **Workflow Clarity**: Use sequential ordered lists for multi-step processes.
- **Progressive Disclosure**: Move code blocks >10 lines to `references/`.
- **Pressure Hardening**: Discipline skills must name red flags, common excuses, and exact stop/restart conditions.

## Evidence Tiers & Guardrails

- **Three Evidence Tiers**: Distinguish (1) structural checks, (2) textual transcript evidence, and (3) executable outcomes.
- **Oracle & Pressure Hardening**: Pressure `behavior_assertions` are positive contains/token checks, not semantic safety guarantees. Regular typed assertions + finite decision-shape controls must reject keyword-stuffed refusal-plus-violation answers; do not claim keyword presence alone establishes compliance. Oracles must test valid paraphrases, quoted unsafe text in refusals, keyword stuffing, partial compliance, and routing negatives; never encode a decision as a bare substring ban on wording a compliant refusal may quote.
- **Digest Binding & Thresholds**: Bind scorer, evals, and generated metadata digests to candidate checkout; no engine-mismatch waiver. Target ≥90% trigger rate (recall & specificity per `readiness.ts`). Retain host guardrails until verified deterministic runtime enforcement replaces them.
- **Discovery & Publication Preflight**: Distinguish search failure/checkout absence from absent capability (search authorized refs/worktrees first). When publishing registry or native export packages, verify final canonical→host ID-set and resource parity and real consumer discovery; review private paths/links and one-way manifests; separate warnings/skips/NOT RUN. Keep publication separate from activation.

## Anti-Patterns

- **No "AI-splaining"**: Do not explain why a pattern is good unless it is a unique project constraint.
- **No Answer-Anchor Padding**: Do not insert arbitrary keywords solely to satisfy string-match tests or fake compliance.
- **No Search-Absence Conflation**: Do not assume missing local search = absent capability; search authorized refs/worktrees.
- **No Vague Triggers**: Never use `src/**` or `**/*`. Keep triggers surgical.
- **No Description Bloat**: If description exceeds 100 words, move capabilities to body.
- **No Long Code Blocks**: Blocks >10 lines must be extracted to `references/`.
- **No Untested Guardrails**: Rules not validated against baseline failure are unverified speculation.

## Quality Checklist (Tessl-Aligned)

- [ ] **Activation ≥ 90%**: Description covers capabilities ("What") and triggers ("When").
- [ ] **Implementation ≥ 90%**: No general-purpose explanations; all examples executable.
- [ ] **Structural Compliance**: SKILL.md ≤ 100 lines; code blocks moved to `references/`.
- [ ] Trigger rate ≥90% on should-trigger queries (recall and specificity per readiness gate).
- [ ] Guardrail skills include rationalizations, red flags, behavior eval fields, and should/should-not trigger cases.
## References

- [Skill Template](references/TEMPLATE.md) — load when starting new skill from scratch
- [Anti-Patterns Detail](references/anti-patterns.md) — load when fixing or reviewing anti-pattern format
- [Size & Limits](references/size-limits.md) — load when SKILL.md approaches 100 lines
- [Resource Organization](references/resource-organization.md) — load when deciding where to place content (scripts/, references/, assets/)
- [Testing & Trigger Rate](references/testing.md) — load when writing evals or measuring trigger rate
- [Eval Workflow](references/eval-workflow.md) — load when running parallel subagent tests
- [Full Lifecycle](references/lifecycle.md) — load for complete phase-by-phase creation guide
- [Web Search Research](references/web-search-research.md) — load when creating skill for unfamiliar or non-engineering domain
Pliki w paczce (13)
  • SKILL.md 7 KB
  • evals/evals.json 4 KB
  • references/TEMPLATE.md 1 KB
  • references/anti-patterns.md 1 KB
  • references/benchmark.md 3 KB
  • references/eval-workflow.md 4 KB
  • references/lifecycle.md 5 KB
  • references/resource-organization.md 7 KB
  • references/rubric.md 5 KB
  • references/size-limits.md 1 KB
  • references/tessl-best-practices.md 2 KB
  • references/testing.md 6 KB
  • references/web-search-research.md 4 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · HoangNguyen0403/agent-skills-standard100MIT
HoangNguyen0403/agent-skills-standard100MIT

Migawka bazy z 13.11.2026 · wersja paczki 3efd7b9dac38 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 95skill-creator
    Tworzy, edytuje, optymalizuje i testuje skille, uruchamia ewaluacje i poprawia trafność wyzwalania.
  • 95skill-creator
    Tworzy, dopracowuje i benchmarkuje skille agentów oraz sprawdza jakość wyzwalania i poprawia opisy skilli.
  • 100skill-creator
    Tworzy i ulepsza skille dla Claude Code lub Cursor, uruchamia ewaluacje i mierzy wyniki, także optymalizując opis pod trafność.
  • 100skill-creator
    Tworzy skille od zera, edytuje i optymalizuje istniejące, uruchamia ewaluacje i dostraja opis pod trafność wyzwalania.
  • 100skill-creator
    Tworzy nowe skille, edytuje i optymalizuje istniejące, uruchamia ewaluacje i mierzy wyniki, także opis pod trafność wyzwalania.
  • 100skill-creator
    Tworzy nowe skille, edytuje i optymalizuje istniejące, uruchamia ewaluacje i mierzy wyniki, także opis pod trafność wyzwalania.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.