ship and groundwork
Two Claude Code skills that make a coding agent behave predictably across sessions.
- Markdown
- JavaScript
- Claude Code
What it is
An agent is only as reliable as what it is told and what it is allowed to do. These two skills are that layer, written for my own workflow and composed from several existing skills rather than built from nothing.
groundwork interviews a project once and writes a small always-read rules file plus a set of notes read only when relevant, so a later session starts already knowing the project's constraints instead of rediscovering them. ship handles getting work to GitHub with every step confirmed and a collaborator's incoming commits checked before a push.
This is context engineering in the most literal form available: deciding what the model sees, what it may act on, and where it must stop and ask. Most production failures with agents are not the model being weak, they are the wrong context reaching it.
Decisions worth defending
- Separates always-loaded rules from notes read on demand, which keeps the always-on context small
- Distinguishes soft rules from hard ones, and routes hard rules to real enforcement rather than leaving them as text an agent may drift past
- Composes existing skills into a workflow instead of duplicating them
- This portfolio was set up using groundwork, so its own docs directory is the worked example
Not built yet
Planned work, with the reason it is next. None of this exists today.
An eval harness for the skills themselves
They are instructions to a model, so they can regress silently when a prompt is reworded. Right now nothing catches that, which is exactly the gap AIKYA closes on its own pipeline.