load when
skill description -> matching work
A skill is not a runtime. It is a focused instruction file the agent reads when the task calls for that discipline.
corpus skills
Optional agent guides: conditioning stances and code-authoring depth, packaged in the open Agent Skills format.
Install the guide that matches the work. Leave the rest unloaded.
index
install.sh — add one skill
pwrcheckevidence
terminal# list what's available
$ npx skills add jcosta33/corpus-skills --list
# install into the current repo
$ npx skills add jcosta33/corpus-skills --skill persona-skeptic
# or install globally / for one agent
$ npx skills add jcosta33/corpus-skills --skill persona-skeptic -g
$ npx skills add jcosta33/corpus-skills --skill persona-skeptic -a claude-code
# no CLI? copy the folder into your agent's skills directory
$cp -R skills/persona-skeptic <your-repo>/.agents/skills/
Skills name abstract command slots like cmdTest and cmdLint; your repo's AGENTS.md supplies the real commands. That is what makes them portable across stacks.
stances.conf — cognitive postures
Conditioning stances
Cross-cutting cognitive postures that tilt what the agent looks for and refuses — load one alongside a task guide when the work needs a particular lens. These are the stances that apply across several guides; the ones tied to a single kind of work (a spec, an audit, a research note, a doc) already ship folded into that work guide, so you get them by using the guide.
markdown onlyon demand
persona-skeptic
judging another agent's work, deepening an audit, or root-causing — refute by default; the lever is the checks you re-run yourself
persona-challenger
pressure-testing a live proposal before it's built — surface assumptions, steelman the alternative, ground every challenge externally
persona-surveyor
breadth research across patterns or products — three named instances per claimed pattern
empirical-proof
any completion claim — bind it to pasted output or it reads unverified
authoring.conf — change-shape guides
Code-authoring depth
Guides for specific change shapes. Pick the one that matches the task so the agent does not treat every problem like a feature.
implement-task
the full corpus task-packet frame, long form
write-feature
net-new behavior behind a defined surface
write-fix
a reproduced defect with a root cause
write-refactor
restructuring with behavior pinned by tests
write-rewrite
re-implementing code whose behavior changes on purpose
write-migration
moving from API A to API B, green after every wave
write-performance
a measured bottleneck with a target and baseline
write-testing
adding the tests an area should already have
write-documentation
human-facing docs for a reader who hasn't read the code
fix-flaky-test
a test that fails intermittently — diagnose, don't retry-loop
authoring.guide — write your own
How to write a corpus skill
Skills are plain markdown, but their structure is evidence-backed: directive descriptions, self-contained bodies, forced visible output, and the AGENTS.md contract.
security.note — no runtime
Why there is no runtime
A skill is a markdown guide your agent reads when the work matches. You can read it first, pin to a commit, and audit what your agent was told to do. If you want command-line scaffolding or automated checks, that is corpus-cli. It is real, but the command surface is still settling.