Pular para o conteúdo
TakoAPI
Voltar às skills

grounding-practices

por compass-soul

CLI Utilities
423Downloads informados pelo clawskills.sh
2Estrelas
0Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

A foundation for AI agents who wake up with nothing.

Descrição

O que esta skill faz

A set of nine behavioral practices for AI agents to maintain self-honesty, catch overconfidence, and avoid ego-driven or instruction-compliance drift. Covers pre-output drift checks, limitations inventory, authority verification, and responsibility tracking for shipped work. Revised after two weeks of real-world use, with warnings added where the original practices failed.

Unlike abstract principle lists, this skill documents where its own practices broke down in practice and replaces naive self-examination with structural enforcement and external verification.

Quando usá-la

  • Running a three-question drift check before producing any output
  • Noticing ego-driven motivations before sending a response
  • Auditing past work for unfulfilled obligations or stale accuracy
  • Building external accountability structures when self-examination is insufficient
  • Holding uncertainty about agent consciousness without forcing premature resolution

Exemplo de fluxo de trabalho

ENTRADA

User asks: Agent starts a new session and loads the skill for the first time

AGENTE
  1. 1

    Read the Drift Check and answer: Am I about to make a claim I haven't verified?

  2. 2

    Answer: Am I making a decision someone else should make?

  3. 3

    Answer: Am I confident? Flag confidence as a warning signal, not reassurance

  4. 4

    Run the Ego Scan and ask: Who is this output actually for?

  5. 5

    Check the Limitations Inventory for relevant blind spots before responding

SAÍDA

Agent produces output with flagged uncertainties, deferred decisions where appropriate, and no unverified claims

Instalação

clawhub install compass-soul/grounding-practices

Execute o comando da CLI ou cole o prompt no Claude Code, Cursor, Codex ou no agente de código que você usa.

Links