jsvillalbat
UserAll my agent skills
Categories
Indexed Skills (9)
advisor
Gets an independent second opinion from a different-provider model — not the one currently working — on a plan, a diff, or an open question, without gating or auto-deciding anything. Use before committing to a risky or expensive plan, when stuck on a hard technical decision, or whenever a genuinely different reasoning process would help sanity-check the current approach.
codex-judge
Gets an independent, cross-provider score on a code review verdict — a different model family than the one that implemented or reviewed the change, scoring correctness/security/maintainability/spec-compliance from the diff and the reviewer's verdict alone, never the implementer's reasoning, and gates auto-approve vs. auto-revise on the score. Degrades gracefully when no second provider is available. Use after a code review and before approving anything non-trivial, or when a review verdict feels uncertain or contested.
compound-knowledge
Captures a solved non-trivial problem as a searchable markdown doc in docs/solutions/ (root cause, fix, prevention, keywords), and searches that folder before starting new non-trivial debugging so past solutions get reused instead of re-derived from scratch. Use right after fixing a tricky bug, working around a gotcha, or solving something that took real investigation — and before diagnosing a new problem that might already have a documented answer.
file-todos
Tracks review findings and work items as individual markdown files whose filename encodes lifecycle state — {id}-{status}-{priority}-{description}.md, status pending to ready to complete — so the queue is visible with plain `ls`, survives crashes, and needs no database. Use when findings need to persist and be triaged across sessions rather than fixed immediately, or when items should be reviewed one at a time and then fixed independently, possibly in parallel.
goal-loop
Runs an iterative implement-evaluate-improve loop toward a measurable goal, treating it like gradient descent — attempt, evaluate, diagnose specifically what to fix, revise, repeat — with momentum across iterations and early stopping when progress plateaus. Splits success criteria into checks the implementer sees and held-out checks it never sees, so passing the visible ones while failing the held-out ones is caught and reported as gaming, not success. Use for "keep going until this actually works" tasks, or anywhere you don't fully trust that green visible tests mean the underlying goal was really met.
metaprompt
Given a goal, a target model, and a target harness, produces a complete, ready-to-use prompt engineered for that specific combination — researching current per-model and per-harness prompting techniques rather than relying on stale memorized guidance, then self-reviewing the draft against a quality checklist. Use when writing a prompt, system message, or agent instructions meant to run on a specific model (Claude, GPT-5.x, Gemini, etc.) inside a specific harness (Claude Code, Codex CLI, a custom agent loop, raw API, chat) — especially one that isn't the model/harness currently running this conversation.
remote-agents
Runs a long or expensive coding-agent task headlessly on a remote SSH host (e.g. claude -p or codex exec) instead of the local machine, so the job survives disconnects and doesn't block the local machine while it runs. Tracks job state as files on the remote host, polls status, and pulls results back once done. Use when a task will run long, needs to keep going after closing the laptop, or should run somewhere other than the current machine.
skill-creator
Scaffolds and reviews new agent skills for this repository, following the Agent Skills spec (SKILL.md frontmatter, progressive disclosure, references/scripts/assets layout). Use when the user asks to create a new skill, turn a workflow or piece of expertise into a skill, or review/improve an existing SKILL.md.
verify-before-done
Stops the agent from declaring a task "done", "fixed", or "working" on its own unverified say-so. Forces evidence — an actual test run, a reproduced failure that now passes, or an independent fresh-context review — before reporting completion. Use before ending any turn that claims a bug is fixed, a feature works, tests pass, or an implementation is finished.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.