Referência
Skills
Técnicas reutilizáveis e detectáveis em arquivos SKILL.md, com sinais de detecção, procedimentos e hooks de aplicação para suítes de planejamento e auditorias do ecossistema.
As skills residem como pastas sob src/apothem/skills/, cada uma com um ponto
de entrada SKILL.md. As skills capturam técnicas reutilizáveis e detectáveis —
elas têm um sinal de detecção, um procedimento e um hook de aplicação.
O catálogo completo de skills é gerado mais abaixo, no Inventário de skills gerado.
Contrato de frontmatter
SKILL.md carrega o frontmatter canônico de oito chaves. O contrato legível por
máquina é skill.schema.json
(additionalProperties: false); este bloco o espelha:
---
name: <skill name>
version: <semver>
updated: <YYYY-MM-DD>
description: <one-line role>
archetype: <class-template, e.g. workflow-template>
userInvocable: <true | false>
disable-model-invocation: <true | false>
allowed-tools: <comma-separated read-only tools and scoped rules, e.g. "Read, Glob, Grep" or "Read, Bash(git log *)">
---Emissão por harness
As skills de origem usam as chaves acima. Alguns harnesses leem chaves diferentes, por isso a instalação as traduz somente para esse harness:
| Harness | O que a instalação altera |
|---|---|
| Claude Code (instalação e plugin) | Renomeia userInvocable para user-invocable, a chave que o Claude Code lê. O Claude Code ignora a chave em camelCase, então, sem a renomeação, uma skill marcada com userInvocable: false ainda apareceria no menu /. |
Pastas de skills compartilhadas no nível do usuário: ~/.agents/skills/ (gravada por Codex, Cursor, GitHub Copilot, Windsurf e Zed) e ~/.codebuddy/skills/, ~/.kiro/skills/ e ~/.trae/skills/ | Uma única forma para toda instalação que grava uma dessas pastas, de modo que as instalações nunca reescrevem os arquivos umas das outras. Ela renomeia userInvocable para user-invocable, move version, updated, archetype e portability para um mapa metadata: final de strings entre aspas e grava agents/openai.yaml com policy.allow_implicit_invocation: false ao lado de cada skill cuja origem define disable-model-invocation: true, porque o Codex ignora essa chave e lê ~/.agents/skills/ quer um adaptador do Codex a tenha gravado ou não. Um comando encapsulado como skill mantém seu corpo inteiro e começa com uma nota fixa que nomeia as pastas de regras e de templates, ~/.config/apothem/rules/ e ~/.config/apothem/templates/. As chaves de nível superior que restam são name, description, user-invocable, disable-model-invocation, allowed-tools, argument-hint e metadata. |
| Qwen Code | Reescreve a string allowed-tools separada por vírgulas como uma lista allowedTools, uma entrada YAML por linha, porque o Qwen Code lê somente a chave em camelCase e exige uma lista. Cada entrada mantém a grafia da origem, incluindo padrões como Bash(git status *). Todas as outras chaves e o corpo da skill são instalados sem alteração. |
| Outros harnesses | Nenhuma alteração. |
skill.schema.json aceita qualquer uma das grafias da flag de invocabilidade
pelo usuário e exige exatamente uma delas.
Disciplina de autoria
- Uma pasta por skill. Cada skill é sua própria pasta, com o ponto de
entrada em
SKILL.md. - Sinal de detecção. Os corpos das skills abrem com o sinal de detecção que dispara a técnica.
- Procedimento + aplicação. As skills carregam tanto o procedimento a aplicar quanto o sinal de aplicação que fecha o ciclo.
- Comprimento da descrição.
descriptiontem no máximo 1.024 caracteres, o limite do Agent Skills. Ferramentas mais estritas rejeitam ou truncam uma descrição mais longa, e o schema impõe o limite.
Inventário de skills gerado
| Name | Description | Source |
|---|---|---|
decision-matrix | Choose between two or more options with a weighted, evidence-cited matrix, a runner-up and a first step. Use when the user asks to 'compare these options', 'help me choose between', 'weigh the trade-offs' or 'which one should we pick'. Not for single-answer lookups or building a chosen option. | src/apothem/skills/decision-matrix/SKILL.md |
dependency-upgrade | Upgrade a project's third-party dependencies with changelog review, pinned versions and the host's lint, test and type-check gates. Runs only when the user invokes it (for example to bump deps or refresh the lockfile). Breaking and major bumps stop for a question; --dry-run classifies without writing; --package NAME scopes to one dependency. | src/apothem/skills/dependency-upgrade/SKILL.md |
dev-toolkit | Run a disciplined engineering loop: reproduce-first debugging, test-first red-green-refactor, vertical slices, or a requirement split into issues. Use for 'diagnose a bug', 'do TDD', 'break this into vertical slices' or 'turn this into issues'. Not for one-line edits or multi-agent orchestration. | src/apothem/skills/dev-toolkit/SKILL.md |
diagram-authoring | Author, validate and render diagrams (flowchart, sequence, state, ER, architecture) and data visualizations to SVG, PNG or PDF. Use when the user asks to 'draw a diagram', 'make a flowchart', 'render this Mermaid' or 'export a diagram to SVG'. Source is syntax-validated before render. | src/apothem/skills/diagram-authoring/SKILL.md |
document-authoring | Write a long, cited document (thesis, report, LaTeX) from an approved outline, with checked citations and a PDF compile. Use when the user asks to 'write a thesis', 'author a LaTeX document', 'compile this to PDF' or 'manage the bibliography'. Not for short edits; pipeline papers: research-paper. | src/apothem/skills/document-authoring/SKILL.md |
ecosystem-audit | Audit the apothem configuration tree as a whole for drift, orphans, dangling references and invalid frontmatter. Use when the user asks to 'audit the ecosystem', 'validate the apothem tree' or 'sweep for drift'. Report-only unless --fix is passed. Changed artifacts only: convention-auditor. | src/apothem/skills/ecosystem-audit/SKILL.md |
eval-harness | Build a reproducible evaluation harness for a prompt or model: versioned labeled dataset, scorer, runner and a report with pass rate, interval and regressions. Use when the user asks to 'build an eval' or 'compare models on a labeled task'. One campaign: eval; rubric scoring: prompt-evaluator. | src/apothem/skills/eval-harness/SKILL.md |
incident-runbook | Write or run an incident runbook (Trigger, Diagnosis, Action, Verification, Rollback) with a rollback per action and confirmation before live changes. Use when the user asks to 'write a runbook', 'define an on-call procedure' or 'document recovery steps'. Not for post-mortems. | src/apothem/skills/incident-runbook/SKILL.md |
issue-lifecycle | Takes a reported defect from issue to fix, one stage per run. Use for 'create an issue', 'size this issue', 'reproduce this bug', 'fix issue 123', 'ask the reporter for details' or 'open a draft PR'. Manual only; posts nothing unasked. Debug loops go to dev-toolkit, advisories to vuln-triage. | src/apothem/skills/issue-lifecycle/SKILL.md |
knowledge-wiki | Adds sources to and audits an index-first Markdown knowledge base, when the user asks to 'add this paper to the wiki', 'ingest these sources', 'lint the wiki' or 'what changed in the knowledge base'. Modes: ingest, lint, drift, hubs. Not source discovery or a one-off summary. | src/apothem/skills/knowledge-wiki/SKILL.md |
multi-source-research | Produces a cited, cross-verified research report with confidence levels, when the user asks to 'research X deeply', 'find sources on' or 'gather evidence about'. Finds new sources and writes the report; given documents: source-synthesis, one claim: fact-checker, a bare source list: research-scout. | src/apothem/skills/multi-source-research/SKILL.md |
paper-toolkit | Papers with no research suite: 'write a reading note', 'compare these papers', 'draft related work', 'what would reviewers ask', 'dedupe my references', 'turn DOIs into BibTeX entries'. Suite or referee review: research-* commands. Reference checks: citation-auditor. Loose files: source-synthesis. | src/apothem/skills/paper-toolkit/SKILL.md |
perspectives | Answers a contested question by arguing it from several separate viewpoints, then reconciling them, when the user asks to 'argue both sides', 'what would a skeptic say' or 'we are split on'. Weighted choices go to decision-matrix, claim checks to fact-checker, builds and reviews to workflow. | src/apothem/skills/perspectives/SKILL.md |
plan-suite | Master Plan Suite template, read by path by the /plan pipeline stages (spec, generate, review, design, audit, execute, status, amend). Holds master-template.md, the canonical plan-suite specification. A template, not invoked directly and not listed; use the /plan commands instead. | src/apothem/skills/plan-suite/SKILL.md |
projectify | Builds the Description, Instruction and knowledge files for a chat-app Project or custom assistant within a measured context budget, when the user asks to 'projectify this', 'write Project instructions' or 'set up a custom assistant'. Not for repository rules, skills or commands. | src/apothem/skills/projectify/SKILL.md |
prompt-engineering | Designs a structured prompt and validates it against happy, edge and adversarial eval cases, when the user asks to 'design a prompt', 'write a system prompt for' or 'improve this prompt'. Delivers the prompt plus its eval set; not a running agent or a model benchmark (eval-harness). | src/apothem/skills/prompt-engineering/SKILL.md |
refactor-extract | Extracts a symbol or file without changing behavior, interactively: specifies what it does, re-derives it clean-room, then checks nothing regressed, when the user asks to 'extract this function' or 'pull this out into its own module'. Delegated refactors: refactor-surgeon. | src/apothem/skills/refactor-extract/SKILL.md |
research-suite | Research Suite template, read by path by the /research pipeline stages (ideate through disseminate). Holds the canonical research-suite surface: the ten rigor mandates and the thirteen-stage lifecycle. A template, not invoked directly and not listed; use the /research commands instead. | src/apothem/skills/research-suite/SKILL.md |
secret-rotation | Use when a credential has leaked or been committed ('rotate a secret', 'a key leaked') and must be retired safely: revokes the exposed secret at its issuer first, re-issues a least-privilege replacement, rewires config through env vars or a secret manager, then checks no plaintext remains. Runs only when the user invokes it. | src/apothem/skills/secret-rotation/SKILL.md |
source-synthesis | Reconciles given documents into one cited synthesis separating consensus from contested claims, when the user asks to 'synthesize these sources', 'combine these documents' or 'what do these say together'. Needs two or more sources. Finding them: multi-source-research; gap map: research-synthesis. | src/apothem/skills/source-synthesis/SKILL.md |
surgical-guard | Makes a minimal, anchor-bounded edit, then reviews only its diff for swallowed errors, hollow tests and docs drift, when the user asks to 'edit surgically', 'make a minimal diff' or 'guard this change'. Not for working out what the change should be (dev-toolkit). | src/apothem/skills/surgical-guard/SKILL.md |
teaching-pack | Builds a teaching pack from one paper or technical document: goals, curriculum, formula walk-throughs, worked examples, quiz with key, glossary, slides on request. For 'teach me this paper' or 'make a study guide from this spec'. Several sources: source-synthesis; critique: research-review. | src/apothem/skills/teaching-pack/SKILL.md |
test-authoring | Writes new behavior-first tests in Arrange-Act-Assert form, failing assertions first, covering happy path, edge boundaries and failures, when the user asks to 'write tests' or 'test this function'. Authors tests only. Whole-repo suite: test-suite; running tests: test-runner. | src/apothem/skills/test-authoring/SKILL.md |
ui-design | Designs or refines a user interface: classifies the surface, applies design rules and keeps DESIGN.md current, when the user asks to 'design a landing page', 'improve this UI', 'make it look better' or 'pick a palette and type scale'. Not for critique-only review (design-critic) or backend logic. | src/apothem/skills/ui-design/SKILL.md |
vuln-triage | Triages one named CVE or security advisory: severity band, affected surface in the lockfile, reachability, and a patch, upgrade, mitigate or accept route, when the user asks to 'triage this CVE' or 'does this advisory affect us'. Does not apply fixes or scan the whole repository. | src/apothem/skills/vuln-triage/SKILL.md |
workflow | Use for a non-trivial mission that needs independent parallel work plus adversarial verification: decomposes it, dispatches independent subagents under named return contracts, runs a refute-by-default critique on every finding, and discloses any beyond-mission fixes. Runs only when the user invokes it, as in 'fan out and verify'. | src/apothem/skills/workflow/SKILL.md |
Regras
Regras comportamentais sempre ativas e filtradas por caminho que governam a resolução de problemas, o ofício do código, a arquitetura limpa, o planejamento e a qualidade em todo o conjunto de regras comportamentais.
Comandos
Definições de comandos de barra que implementam o pipeline da suíte de planejamento, além de auditorias de código, arquitetura, segurança, dependências e revisão de documentação.