IgorGanapolsky

context-engineering-checklist

context-engineering-checklist技能提供遵循Hugging Face Context Course原则的代理能力实现结构化方法。它指导开发者选择合适的层级(技能、MCP工具、工作流、子代理或钩子/门控),并确保每个组件的正确实现,包括SKILL.md文件的特定要求、工具定义和工作流模式。

IgorGanapolsky 29 7 更新于 1个月前

资源

2
GitHub

安装

npx skillscat add igorganapolsky/thumbgate/context-engineering-checklist

通过 SkillsCat registry 安装。

技能简介

context-engineering-checklist技能提供遵循Hugging Face Context Course原则的代理能力实现结构化方法。它指导开发者选择合适的层级(技能、MCP工具、工作流、子代理或钩子/门控),并确保每个组件的正确实现,包括SKILL.md文件的特定要求、工具定义和工作流模式。

SKILL.md

Context Engineering Checklist (HF course → ThumbGate)

Source: The Context Course

Pick the right layer (Clarify)

Need Layer Do not
How-to / policy for agents Skill (Unit 1) Long chat-only prompts
External API/tool MCP (Unit 2) Fake tools in prose
Multi-step pack / fan-out Workflow/plugin (Unit 3) One mega-prompt
Parallel specialists Sub-agents (Unit 4) Dual-edit without locks
Irreversible deny Hook/gate (Unit 5) Skill-only “never do X”

Unit 1 — Skills

  • SKILL.md frontmatter: name, trigger-rich description, slash
  • One failure class per skill
  • NEVER / ALWAYS table with zero-ratio rules
  • Scripts for mechanical steps; no secrets in body

Unit 2 — MCP

  • Tool names match discovery / allowlist
  • Side-effect annotations truthful
  • Mutating tools covered by PreToolUse / gate-check

Unit 3 — Plugins / workflows

  • Context → specialists → verify → synthesize
  • Smoke + full modes; read-only reviews

Unit 4 — Sub-agents

  • Linear claim or complement (no steal)
  • Vault one-writer; scoped prompts

Unit 5 — Hooks / gates

  • Matchable tool surface (not English prose)
  • Hard floor if irreversible
  • Deny + allow + false-positive tests
  • Live probe before “protected”

Six-block context envelope

Before generation, package reusable context into all six blocks:

  1. goal: one measurable outcome.
  2. businessData: current evidence with source and timestamp.
  3. examples: concrete good or bad few-shot cases.
  4. procedures: ordered actions and verification.
  5. constraints: safety, business, and SLA limits.
  6. rubric: grading criteria supplied before the answer is generated.

Use construct_context_pack.contextEnvelope instead of repeating these blocks
in chat. All blocks are required when the envelope is present, count against
maxChars, and are part of semantic-cache identity.

For connector evidence, set metadata.sourceUpdatedAt,
metadata.maxAgeSeconds, and metadata.sourceUrl on the stored context
object. Treat stale, future, invalid, and unknown freshness as
explicit evidence states, not current truth.

Scripts

bash ~/.grok/skills/context-engineering-checklist/scripts/check_context_layers.sh [path...]

Related

  • [[gsd-ralph-context-loop]]
  • [[three-bus-ship-cycle]]
  • [[xai-workflows-for-system]]