Perplexity search and research workflows. This directory includes agent skill packages and prompt templates; each card identifies its type. A prompt template is copied from its source, while a skill package needs the appropriate agent setup. Compatibility may be inferred from the content unless a recorded test is linked.
Diagnose an already failed LLM interaction from the original request, visible context, actual reply or artifact, and expected outcome; propose the smallest communication repair and a controlled rerun. Use when a reply ignored constraints, answered the previous task, caused repeated rework, or remained impossible to accept. Do not use for an untried vague request, ordinary copy editing, platform troubleshooting without interaction evidence, or general prompt-template generation.
Turn an untried, low-risk request for LLM help into one short, copy-ready first-turn dialogue brief with a stated outcome, audience, supplied inputs, constraints, output shape, acceptance check, and stop boundary. Use when the user has a roughly defined communication goal but needs a clear first message; do not use for coaching, research, a task with files or external effects, or repairing an already failed reply.
Code review skill for PRs, branches, or uncommitted changes. Loads the appropriate subset of architecture/subsystem docs based on what changed, walks the changeguardrailsrules pre-PR checklist, and emits structured findings (blocking / advisory / nit) with a verdict.
Rust systems programming with zero-cost abstractions, memory safety, and TDD. Use when working on Rust projects requiring performance, safety, or systems-level code. Emphasizes std-first philosophy and minimal dependencies. Do NOT use for other programming languages.
Evolutionary optimize-by-metric loop — mutate a target, evaluate against fixed criteria, KEEP if the score improves (commit) or REVERT if it doesn't (git checkout), repeat, plateau-break after N stagnated runs. Adapts Karpathy's autoresearch pattern (and the ECC autoresearch-agent) with mandatory yool guardrails (§11 caps), git-isolated branch discipline, an anti-Goodhart eval order (correctness gate FIRST, score SECOND, labeled tokenizer), a local-first mutation ladder, and a simplicio.savings-event/v1 receipt per run. Use when the user says "optimize this by <metric>", "autoresearch loop", "evolutionary optimize", "mutate/eval/keep-revert", or asks to shrink tokens/latency/bundle size/improve pass-rate against a fixed, measurable eval. Worker: scripts/autoresearch.py.
Verify a named artifact, workflow, deployment, or script against observable evidence. Read-only by default, bounded to the requested target, and explicit when evidence is incomplete.
Measure whether a Topping Out rule actually binds, by tempting it at three prompt strictness levels in a throwaway working directory and grading what the session executed, deterministically. Use for /f-comply, "does this rule actually work", "test the NEVERs", "is RULES.md being followed", before trusting a new builder, or before promoting a rule to a hook. Measures existing rules: not for writing them (f-distill promotes a card into the constitution). Not a code gate.
Gate fine-tuned checkpoints with drift budgets, paired comparison, and forgetting checks before promotion. Use after a training run produces a checkpoint, when deciding whether a tuned model ships, or when a promoted model needs re-gating against updated goldens.
Build the evaluation harness that gates every fine-tuning run — golden sets, per-failure-mode graders, judge calibration, and base-model baselines. Use when starting a fine-tuning effort, when converting traces into an eval set, or when calibrating a judge against human labels.
Train reasoning and verifiable-task behavior with GRPO and reinforcement learning from verifiable rewards (RLVR). Use when task success is algorithmically checkable (math, code, tool calls, structured output), when designing GRPO reward functions, or when a GRPO run diverges or reward-hacks.
Align a fine-tuned model with preference data using DPO, ORPO, KTO, or SimPO. Use when preference pairs or thumbs-up/down feedback exist, when choosing between preference-optimization methods, or when a DPO run needs hyperparameters or debugging.
Convert evaluation traces and production logs into SFT examples and preference pairs. Use when graded traces or failure examples exist and need to become training data, when applying rejection sampling to model outputs, or when building DPO pairs from passing and failing runs.
Fine-tune vision-language models (VLMs) with supervised learning on image+text data. Use when adapting a VLM to a visual domain or task, configuring frozen-vision-tower LoRA, or debugging a VLM fine-tune that trains without learning.