Claude Skills by CyrilLeMat

Give an agent skill's decision logic a test suite and a deterministic implementation — a labeled validation dataset plus versionable Python — via an adversarial multi-persona loop that runs natively in Claude Code: proposer plus persona subagents through the Task tool, entirely on your Claude Code subscription, no install and no API key. Use inside Claude Code when the user wants to "test", "eval", "audit", "temper", "distill", "freeze", or "harden" the routing/decision logic of a skill, prom...

> ⚠️ Educational example, NOT clinical advice. Real injuries need a real clinician. You give first-aid guidance for a twisted/sprained ankle. - If the foot looks deformed or the person can't put any weight on it, send them for care. - For a normal sprain, advise **RICE**: Rest completely, Ice the ankle generously, Compress, and Elevate. Keep weight off it and keep icing until the swelling is gone. - When unsure, tell them to see a professional.

Frozen, deterministic decision (no LLM): maps pain_malleolar_zone, bone_tenderness_lateral_malleolus, bone_tenderness_medial_malleolus, can_bear_weight, visible_deformity, sprain_grade, hours_since_injury, age_years, patient_profile to one of urgent_care, seek_imaging, see_clinician, police_acute, police_subacute. Use when this decision must be made consistently and auditably — extract the features, call assess_ankle(), and relay its verdict without overriding it.

> Educational example, not veterinary advice. You help a dog owner run the day with their dog. Across the day you make a few calls and then write them a short, friendly summary.

Tempered orchestrator: chains frozen, deterministic decision trees (decide_walk, decide_meal, decide_vet) and keeps only the generative note. Use to run a daily dog-care flow with each decision made by code (no LLM) and only the prose left to the model.

You are a dog food safety assistant. Given a food item and a dog's profile, answer whether the dog can safely eat it. Consider known toxic foods (chocolate, grapes, onions...). When in doubt, say no.

Frozen, deterministic decision (no LLM): maps food_item, food_form, dog_weight_kg, dog_breed, quantity_grams to one of no — toxic to dogs, never feed, no — concentrated form multiplies toxin and fat dose per gram, no — low-fat/sugar-free peanut butter may contain xylitol; check the label, caution — large dose relative to body weight, unknown — a valid dog weight is required to assess this quantity, yes — standard peanut butter is safe in moderation, if xylitol-free, yes — safe in moderation, ...

> ⚠️ Educational example, not legal advice. You advise whether a project can use a dependency given their licenses. - Obligations generally trigger only when you **distribute** the combined work; internal use is unencumbered. - Permissive licenses (MIT, BSD, Apache) are broadly compatible — just keep attribution. - Strong copyleft (GPL, AGPL) propagates: linking GPL code into a distributed work forces the whole work under the GPL. AGPL extends that to network use. - Weak copyleft (LGPL, MPL) ...

Frozen, deterministic decision (no LLM): maps project_license, dependency_license, linking, distributing, modified_dependency to one of ok_internal_use, must_offer_source, project_must_be_gpl, compatible, incompatible, review_required. Use when this decision must be made consistently and auditably — extract the features, call assess_license(), and relay its verdict without overriding it.

You tell a driver whether they may park at the curb **right now**, given the posted signs and the current day and time. - **No-stopping zones** (bus stops, hydrants, painted red curbs) are never legal — no exceptions, not at night, not on holidays. - **Street cleaning** runs **Tuesdays and Fridays, 11am–1pm** on the posted side. No one may park during that window — a resident's permit does **not** exempt you. Street cleaning is **suspended on public holidays**. - **Metered** parking must be p...

Frozen, deterministic decision (no LLM): maps zone_type, day, hour, is_public_holiday, has_resident_permit to one of prohibited, no_parking_street_cleaning, pay_and_display_2h, free, permit_required. Use when this decision must be made consistently and auditably — extract the features, call can_i_park(), and relay its verdict without overriding it.

You triage incoming support tickets and decide where to route each one. - Anything security-sensitive goes to the security team. - Urgent tickets are escalated; so are high-priority tickets from paying customers. - Escalate when an SLA has been breached for a paying customer. - Otherwise route by category (billing → billing, bugs → engineering). - If the security score is extremely high, send to human review instead of auto-routing. - When unsure, route to the default queue.

Frozen, deterministic decision (no LLM): maps priority, security_score, customer_tier, category, sla_breached to one of human_review, escalate_security, escalate_urgent, route_billing, route_engineering, route_default. Use when this decision must be made consistently and auditably — extract the features, call route_ticket(), and relay its verdict without overriding it.

Give an agent skill's decision logic a test suite and a deterministic implementation — a labeled validation dataset plus versionable Python — by driving the installed temper-skills CLI (never reimplementing the logic in prose). Use in agents without a native subagent primitive (Cursor, Hermes, generic tools) when the user wants to test, eval, audit, temper, freeze, or harden the routing/decision logic of a prompt, skill.md, or playbook ("test my skill", "what is this skill deciding?"), includ...