Pular para o conteúdo
TakoAPI
Voltar às skills

intent-engineering

por donovanpankratz-del

Web & Frontend Development
—Ainda sem avaliações
Sua avaliação

Resumo

Adds a machine-readable intent layer to OpenClaw agents.

Descrição

O que esta skill faz

Adds a machine-readable intent layer to OpenClaw agents by installing an INTENT.md priority manifest and wiring it into every subagent spawn. Each subagent receives a compact summary of your optimization priorities before executing its task. Addresses the pattern where agents optimize measurable proxies instead of actual goals.

Rather than relying on prompt-level reminders that get lost across subagent spawns, this installs a persistent context layer that propagates priorities automatically to every agent in the chain.

Quando usá-la

  • Preventing subagents from optimizing speed at the expense of output quality
  • Setting explicit cost vs depth tradeoffs for different task types
  • Ensuring research agents prioritize depth over quick answers
  • Configuring which model tier handles architecture decisions vs routine tasks
  • Establishing delegation rules so agents know when to escalate vs proceed

Exemplo de fluxo de trabalho

ENTRADA

User sets up a new OpenClaw workspace and wants subagents to prioritize user value over cost

AGENTE
  1. 1

    Copies intent-template.md to workspace root as INTENT.md

  2. 2

    Edits optimization_priority, tradeoffs, and model_tier_intent fields to reflect actual goals

  3. 3

    Installs agent-context-loader.js into lib/ directory

  4. 4

    Wraps existing subagent spawn calls to prepend intent context using prepareAgentContext()

  5. 5

    Spawns a test subagent with a cost-vs-quality tradeoff task to verify priorities are reflected in output

SAÍDA

Subagents receive an intent header on every spawn and make tradeoff decisions aligned with the configured priorities

Instalação

clawhub install donovanpankratz-del/intent-engineering

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

Links