skill-to-prompt-porter

name	skill-to-prompt-porter
description	Port Claude Code skills into standalone prompts for external LLMs (Grok, ChatGPT, Gemini, NotebookLM) — methodology extraction, infrastructure stripping, modality adaptation, source-grounded adaptation
license	MIT
metadata	{"version":"1.1.0","hermes":{"tags":["Skills","LLM","Portability"],"related_skills":["skill-pattern-alignment","skill-decomposition-methodology"]}}

Porting Skills to Standalone LLM Prompts

Core Principles

1. Methodology is portable, infrastructure is not: Decision trees and quality standards transfer. Git, MCP, wiki-links do not.
2. Preserve the quality floor: Non-negotiable standards are the skill's real value — never drop them to simplify.
3. Adapt modality, not methodology: If original reads files, ported prompt processes screenshots or pasted text. Analysis stays identical.
4. Self-contained means self-contained: Works with zero context about your notes system or specific tools.

Three-Layer Separation

Layer	Action	Contents
Methodology (KEEP)	Port entirely	Decision trees, quality standards, forbidden patterns, templates, grouping heuristics
Infrastructure (STRIP)	Remove completely	Git ops, MCP calls, wiki-link verification, README updates, archival, file paths, batch processing
Modality (ADAPT)	Transform for platform	Input/output interfaces

Modality Adaptation Table

Original (Claude Code)	Adapted (External LLM)
Read tool on .md files	Screenshot, pasted text, URL
Write tool to create files	Raw markdown in chat
Grep/Glob for existing notes	User handles deduplication
Wiki-links [[Note Name]]	Plain text or omit
Archival workflow	Omit entirely

Source-Grounded Platforms (NotebookLM)

The Modality Adaptation Table above targets chat LLMs (Grok, ChatGPT, Gemini) that generate from model knowledge — port quality standards to constrain generation. Source-grounded tools (NotebookLM) answer only from uploaded sources — adapt differently:

Original (Claude Code)	Adapted (NotebookLM)
Resource suggestions per item	"Best source in this notebook" per item
Analysis from model knowledge	Require a source citation on every output node
Quality floor vs hallucination	Instruct: mark uncovered items "GAP — not in sources", never invent to fill
Interactive clarification steps	Inline placeholders the user fills ([TOPIC], [LEVEL])

The GAP rule has a side benefit: the output doubles as a coverage audit of the notebook's source bundle.

Added 2026-06-07 from porting skill-tree-generator to NotebookLM.

7-Step Porting Workflow

1. Read source skill — classify every section as Layer 1/2/3. Pull in any files the SKILL.md points to (references/, scripts/, assets/); a port that omits them drops content the skill declares required. Scrub before porting: source content going into a third-party LLM must carry no secrets, internal hostnames, or private-infra names, even inside Layer-1 methodology kept verbatim.
2. Extract decision tree — core analysis logic transfers verbatim
3. Extract quality standards — consolidate all REQUIRED/CRITICAL/FORBIDDEN into one section
4. Extract forbidden patterns — port ones that apply without infrastructure
5. Build output template — strip tool-specific fields, replace with platform equivalents
6. Add input handling — screenshot (transcribe first, describe visuals, flag incomplete), pasted text (accept as-is, ask for source), URL (fetch and extract metadata)
7. Write prompt document — single markdown file.
8. Verify the floor survived — before declaring done, list every REQUIRED/CRITICAL/FORBIDDEN item from the source and confirm each appears in the port. If the target platform has a length ceiling (Custom GPT ~8k chars, Gem limits) and the source overflows it, cut illustrative examples and prose FIRST; never drop a quality-floor rule to fit — if the floor itself won't fit, say so rather than shipping a silently weakened port.

# [Skill Name] Prompt

> **Usage**: [One-line instruction]

---

## System Prompt

[Instructions: input handling + methodology + output template + quality standards + forbidden patterns]

---

*Methodology adapted from [skill name].*

Key Rules

• Never keep wiki-link syntax on platforms where it's meaningless
• Collapse multi-phase workflows into single-pass: analyze input, produce output