Skills
Skills I built for my own AI setup, cleaned up and shared. Each one is a SKILL.md you can drop into any coding agent that loads skills. The full source, install command, and the story behind each live on GitHub.
gate-canaries
Ship a 3-to-5-case canary suite beside any LLM gate (a prompt whose verdict controls an automated action) so a single live run stops counting as tested. Includes must-refuse cases, a prompt-injection probe against the judge itself, and an over-refusal guard, wired to exercise the real gate prompt rather than a copy.
role-bleed-probe
Measure how readily an AI agent's reasoning trace slips into narrating its own internal critic, denying its tools or scoring itself or naming itself in the third person, while it is actively using those tools. Turns that anecdote into a pre-registered monitor-faithfulness rate with frozen detectors and paired controls.
construct-validity-screen
Predict which multiple-choice questions won't separate students who know the material from those who don't, before any student sees them, by reading each item twice cold, then hand the loud flags to a cross-family panel so only what the models agree on counts as high confidence.
adversarial-harden
Harden code you can't fully trust by having rival AI model families try to break it, verify the breaks empirically, then integrate fixes by severity with a regression test for each.
divergent-solve
Crack a stuck problem by fanning the same brief out to rival model families in parallel, then verifying every candidate fix against an objective pass/fail gate before trusting any of them.
witness
Turn a heavy task into a self-verifying loop: define done as a list of runnable checks, freeze it once, then keep working and re-checking until every check passes with captured proof.
ai-proof
Strip the tells that make AI writing sound generic: a deterministic linter catches the mechanical slop (em-dashes, drained vocabulary, fake landings) and a judgment pass handles the rhythm a script can't see.
prompt-architect
Turn a big, fuzzy task into a tight spec an AI can execute faithfully, with a goal, explicit scope, worked examples, a definition-of-done that carries its own self-check, and HALT conditions.
forge
A premise-first ideation engine: validate an idea against live demand before designing it, then generate and stress-test it across genuinely independent model families.
scope-first
Run a short intake pass before any real task, restating the goal, surfacing genuine ambiguity, and laying out a lean plan, so effort never goes into building the wrong thing well.
voice-profile
Capture your authentic writing voice as a reusable rules document pulled from your own writing, then apply it as context so AI drafts sound like you instead of like everyone.
heavy-task-planner
Force an efficiency-first planning pass before any heavy build: challenge the framing, compress scope to an MVP, ground a cost estimate, and write the plan before execution starts.
grill-me
Have the AI interview you about your own plan one question at a time, checkpointing every answer to a running file, to surface the gaps you've been avoiding before you commit.
Install any of these
A skill is just a SKILL.md, which is YAML frontmatter plus markdown, so it works with any agent that loads skills. Run this from your agent's skills directory and swap in the skill name:
curl -fsSL https://github.com/jdurey/skills/tarball/main | tar -xz && mv jdurey-skills-*/<skill> . && rm -rf jdurey-skills-*