consult-compendium

name

consult-compendium

description

Consult the guild's craft knowledge compendium for reference material relevant to the current task. Use when starting a code review, writing a spec, writing a plan, making design decisions, conducting research, beginning implementation from a plan, writing a commission prompt, working with TypeScript patterns, generating images, designing visual assets, conducting strategic analysis, setting up a new project's architecture, designing a CSS color system, building with the Claude Agent SDK, or entering any domain the compendium covers. Triggers include "check the compendium", "consult reference", "craft guidance", "reference entries".

Consult Compendium

Load relevant craft knowledge from the guild's reference library before starting work in a covered domain. The compendium contains curated entries on topics like spec writing, planning, design decisions, research, code review practices, TypeScript patterns, implementation workflow, commission prompt craft, and other domains added over time.

This skill is passive guidance. It does not change your posture, identity, or tool access. It points you at reference material and lets you proceed with that knowledge loaded.

How to Use

1. Check What's Available

Use Glob to list the entries in the reference/ directory relative to this skill:

Glob: packages/guild-compendium/plugin/skills/consult-compendium/reference/**/*.md

If the directory is empty or contains no .md files, proceed with your task without reference material. The compendium may not have entries for every domain yet.

Reference files are organized into subdirectories:

• workflow/ — Task-type guidance (how to do code review, planning, research, etc.)
• craft/ — Domain knowledge and reusable patterns (TypeScript, architecture, color systems, Claude Agent SDK)
• commissions/ — Commission-specific craft (prompts, chaining)

2. Read Relevant Entries

Based on your current task, read the entry or entries that match the domain you're working in. You don't need to read every entry. Pick the ones that are relevant:

• Starting a code review? Read workflow/code-review.md.
• Writing a spec or requirements? Read workflow/spec-writing.md.
• Writing an implementation plan? Read workflow/planning.md.
• Making technical design decisions? Read workflow/design-decisions.md.
• Conducting research? Read workflow/research.md.
• Beginning implementation from a plan? Read workflow/implementation.md.
• Writing a commission prompt? Read commissions/commission-prompts.md.
• Chaining commissions? Read commissions/commission-chaining.md.
• Working with TypeScript? Read craft/typescript-practices.md.
• Setting up a new project or defining its architecture? Read craft/daemon-first-architecture.md.
• Designing a color system, adding CSS color tokens, or auditing existing colors? Read craft/color-system-principles.md.
• Building with the Claude Agent SDK, defining custom tools, or configuring MCP servers? Read craft/claude-agent-sdk.md.
• Working in another domain? Browse the subdirectories for a match.

Use the Read tool on each relevant file.

3. Absorb and Proceed

Each entry is 500 to 1000 words of distilled, actionable guidance. Read the key points and carry them as context into your work. You do not need to cite the entry or follow it rigidly. It orients you in the domain; your posture and the task itself shape how you apply what you read.

Entry Format

Reference entries are markdown files with YAML frontmatter containing four fields:

---
title: What Makes a Good Spec
domain: software-development
last_updated: 2026-03-23
source: "research commission (Verity, 2026-03-23)"
---

• title: Descriptive name for the entry.
• domain: Categorization keyword in kebab-case (e.g., software-development, visual-craft, project-management).
• last_updated: ISO date of last revision.
• source: Where the entry came from (research commission, direct write, or retro promotion).

Entries are self-contained. You can understand the topic from a single entry without reading others. They are also worker-agnostic: a code-review.md entry is useful whether you are conducting a review or preparing code for one.

What This Skill Does NOT Do

• Does not inject content automatically. You pull what you need via file reads. No context window cost when the material isn't relevant.
• Does not modify your behavior. The reference material informs; your posture directs.
• Does not require all entries be read. Read what's relevant. Skip what isn't.