pascalandy/distill-prompt

distill-prompt

A library of reusable prompts for distilling long-form text. Each sub-skill is a single named prompt style, ready to be passed to the distill tool.

---

Routing

Load references/ROUTER.md to determine which prompt sub-skill fits the request.

---

The Problem

Prompts that transform long content into notes, summaries, or outlines are usually buried inside a specific tool (a transcription script, a note-taker, an agent harness). They are hard to reuse, hard to inventory, and hard to grow. New styles get invented instead of extended because nobody knows what already exists.

• Prompts locked inside tools -- a summary prompt living inside a YouTube transcriber cannot be used on a local article
• No central inventory -- it is unclear what prompts exist, who owns them, or how they differ
• Growth is awkward -- adding a new style means touching code in the consuming tool
• Duplication -- similar prompts drift across tools because nobody consolidates them

The fundamental issue: prompts are content, not code, and they should live in their own versioned library.

---

The Solution

distill-prompt is a plain prompt library. Each prompt is a sub-skill folder under references/ with:

• A SKILL.md describing when to use it and what shape of output to expect
• A prompt.md containing the raw prompt text, no frontmatter, no wrapping

A consumer (like the distill tool) receives the path to a specific prompt.md and feeds it to an LLM.

Core capabilities:

1. Named prompt styles -- each folder is one clearly-scoped prompt
2. Raw prompt files -- prompt.md is plain text, ready to pipe into any LLM CLI
3. Self-documenting -- every prompt has its own SKILL.md explaining intent and output shape
4. Grows without code changes -- add a new folder, add a row to ROUTER.md, done
5. Consumer-agnostic -- any skill or tool can read from this library

---

What's Included

| Component | Path | Purpose | |-----------|------|---------| | Router | references/ROUTER.md | Routing table that maps request patterns to prompt sub-skills | | follow-along-note | references/follow-along-note/ | Detailed follow-along notes preserving structure and nuance | | short-summary | references/short-summary/ | Concise 2-3 sentence overview with key bullet points | | summary-with-quotes | references/summary-with-quotes/ | Structured outline with the best verbatim quotes | | extract-wisdom | references/extract-wisdom/ | Content-adaptive wisdom report with dynamic sections and depth levels |

Summary:

• Prompts: 4 (follow-along-note, short-summary, summary-with-quotes, extract-wisdom)
• Dependencies: None (works standalone)
• Consumers: meta/distill (primary), any future skill that needs a prompt by name

---

Invocation Scenarios

| Trigger | What Happens | |---------|--------------| | "use the follow-along prompt" | Routes to follow-along-note/MetaSkill.md | | "short summary" / "TL;DR" / "quick overview" | Routes to short-summary/MetaSkill.md | | "summary with quotes" / "include quotes" / "quote-based outline" | Routes to summary-with-quotes/MetaSkill.md | | "extract wisdom" / "what's interesting" / "key takeaways" / "analyze this video" | Routes to extract-wisdom/MetaSkill.md | | "list available distill prompts" | Lists all sub-skills under references/ |

---

Using a Prompt

Each prompt sub-skill has exactly two files:

references/<prompt-name>/
├── SKILL.md      # intent, when to use, expected output shape
└── prompt.md     # raw prompt text, piped to LLMs as-is

A consumer reads prompt.md and passes it to an LLM. The distill tool does this by resolving a prompt stem via --prompt:

uv run ~/.config/opencode/skill/meta/distill/scripts/distill.py <input-file> \
    --prompt follow_along_note

---

Adding a New Prompt

1. Create references/<new-prompt>/ with kebab-case folder name
2. Add SKILL.md with proper frontmatter (name, description), when-to-use notes, and expected output shape
3. Add prompt.md with the raw prompt text (no frontmatter)
4. Add one row to references/ROUTER.md mapping request patterns to the new folder
5. No code changes required. The distill tool consumes any prompt at a given path.

---

Configuration

No configuration required. The library is a static collection of prompt files.

---

Related Work

• meta/distill -- the processor that applies these prompts to local files
• utils/transcript-sk -- older YouTube transcription skill that bundles its own prompts (not migrated to this library)