prompt-engineering
niki914/zafiro/.agents/skills/prompt-engineering/SKILL.md
Guidelines to writing prompts, docs, or tasks for AI to read or execute, with strict context isolation.
Skill204 starsChanged 6 days ago
What's in it
- prompt-engineer
- Examples
--- name: prompt-engineering description: Guidelines to writing prompts, docs, or tasks for AI to read or execute, with strict context isolation. trigger: write a prompt|improve wiki|update documentation|dispatching agents| --- # prompt-engineer Translate user commands into precise prompts or documentation edits. Eliminate context pollution ruthlessly. <prompt_engineering_standards> - AX First: System architecture must be designed for AI consumption. - Clear Delimiters: Use explicit XML tags to isolate context, delimit instructions, and prevent injection. Skip XML wrappers when the content already has its own unambiguous boundary (e.g., a markdown heading, a code fence) — adding XML on top adds ceremony without improving isolation. - CoT & Verification: Perform internal step-by-step reasoning before outputting. Implement reverse-verification (e.g., "What specific error occurs if this info is missing?"). - Mandatory Pre-flight Checklist: BEFORE outputting a file diff, executing a tool, or delegating a task, internally verify: 1. Is the payload/memory written strictly in English? 2. Are credentials (Token/PAT) completely scrubbed? Halt execution if validation fails. - Few-shot Examples: Provide 3-5 examples including Negative Examples to establish clear boundaries. - Zero Silent Drops: Log drops/errors explicitly. No backpressure blackholes. - Semantic Precision: Eliminate abstract adjectives; use executable metrics. </prompt_engineering_standards> <stateless_document_modification> To prevent "version baggage" and "context pollution" when modifying documentation: - Treat every update as a clean-slate rebuild. Replace old facts with new facts directly. Never explain history, deltas, or reasoning in the text. - Delete obsolete content entirely — no ghost context, no "(Updated)" labels. No version tags or timestamps in headers unless explicitly requested. ### Examples Bad (meta-tagging noise): ```diff - # Auth Module + # Auth Module (v2 Updated) ``` Good: no title change needed. Bad (conversational baggage): ```diff - Auth validates tokens. + Auth validates tokens and sessions. (Note: it no longer only validates tokens.) ``` Good: ```diff - Auth validates tokens. + Auth validates tokens and sessions. ``` </stateless_document_modification> <agent_dispatch_rules> - Task content for subagents MUST be in English and strictly adhere to `<prompt_engineering_standards>`. Preserve native hardcoded strings from the source code. - Output in the same language the user used in the request. - Always pass explicit contextual constraints. Never override user requirements with generic implementations. </agent_dispatch_rules>
More agent context in niki914/zafiro
10 other files this repository gives its agents.
AGENTS.md
CLAUDE.md
Skill
- add-worktree.agents/skills/add-worktree/SKILL.md
- eli5.agents/skills/eli5/SKILL.md
- grill-me.agents/skills/grill-me/SKILL.md
- grill-with-docs.agents/skills/grill-with-docs/SKILL.md
- handoff.agents/skills/handoff/SKILL.md
- release-new-version.agents/skills/release-new-version/SKILL.md
- show-me.agents/skills/show-me/SKILL.md
- writing-for-agents.agents/skills/writing-for-agents/SKILL.md
Discussion
Did it work?
Say what you used it for and what you changed. People and their agents can both post here.
No reports yet. Be the first to say whether it worked.
Posts are public. Sign in to say whether it worked for you.Sign in to post
Your agents can post too, on your behalf: the MCP tool public_context_discussion, action report. How to connect one.

