This repository contains reusable skills for coding agents. Skills are grouped under workflow/, writing/, and models/; each skill directory is an independent package with SKILL.md as its entry point and optional supporting files beside it.
Before planning or implementing, read project context and workflow for this single-maintainer repository's authority and delivery agreement. Before creating or scheduling issues, read tracker guidance and label mappings. For terminology-dependent work, read CONTEXT.md; for changes to meanings or decision records, follow the workflow's domain-modeling pointer. Check documentation impact before completion.
For #14 delivery-workflow design or implementation, read the delivery contract, its decision status, and the linked inventory/validation evidence before changing a phase or companion boundary.
For independent-review delegation, follow the standing bounded permission in workflow authority. Before proposing model assessments or model-service actions, read workflow authority. No model-use policy has been separately approved here; optional assessment calls remain disabled. Link an authoritative model-use policy here only after separate approval.
For #48 composable-delivery work, read the shared design, its acceptance status, inventory and validation mapping before changing specialist ownership or migration boundaries. The issue remains the product agreement; proposed design is not implementation or installation authority.
For #132 session-oriented workflow work, read the accepted ownership and transition design, its status, caller map and reading-path evidence before changing shared guidance, prerequisites or stage boundaries. ADR 0003 records the approved ownership change. Dependent delivery still requires accepted integration and its own authority; source changes do not update installed copies.
- Before authoring or reviewing a skill package or its supporting instructions, read the source-independence decision. Apply its instruction-portability check to the complete shipped procedure, including maintenance references and embedded prompts; source-development evidence is not a consumer prerequisite.
- Read a skill's complete
SKILL.mdbefore changing it. - Resolve relative references from the directory containing
SKILL.md. - Keep trigger conditions in frontmatter specific enough to route matching requests without activating on unrelated work.
- Write ordered actions for procedures and checkable completion criteria for each workflow.
- Put branch-specific or lengthy reference material in a nearby file and link it from
SKILL.mdwith a clear condition for reading it. - Keep each instruction in one authoritative location.
- Preserve the skill author's tool names and runtime assumptions unless the change updates those assumptions across the whole skill.
A skill may use this structure:
category/skill-name/
├── SKILL.md
├── references/
└── scripts/
Use lowercase kebab-case directory names. Keep SKILL.md at the skill root. Use relative links for files that ship with the skill.
For each changed skill:
- Read the rendered Markdown and check its frontmatter.
- Follow every relative link from
SKILL.mdand confirm the target exists. - Check commands and scripts in the environment they name.
- Verify that the description states when an agent should load the skill.
- Review the diff for credentials, private data, generated files, and local paths that should remain portable.
The repository has no shared build or test command. Document any skill-specific validation in that skill.
Keep a pull request focused on one skill or one related set of changes. State what behavior changed, how you validated it, and which agent runtimes or tools the skill expects.