sylphx-skills

Assess whether disabled users can complete the selected tasks, and what the current results support saying about conformance. Use when the user asks to review accessibility, audit a conformance claim, or check a VPAT or accessibility statement. Do not use to design the system under review or to fix accessibility defects.

Assess account recovery — alternative authentication, recovery proof, takeover handling, and enterprise operations. Use when the user asks to review password reset, recovery flows, account takeover handling, or SSO recovery. Do not use to design the recovery flow.

Estimate a causal effect with an explicit estimand, identification assumptions, design, diagnostics, and sensitivity. Use when the user asks whether a change caused an outcome, for difference-in-differences, IV, or target-trial emulation. Do not use to register a product experiment, forecast an unresolved future event, or critique a present explanation.

Stress-test a claim, plan, diagnosis, or design by writing rejection conditions before interpreting the evidence. Use when the user asks to critique, steelman, red-team, find the kill criterion, or what would change your mind. Do not use for a domain go/no-go review, a causal effect estimate, a resolvable forecast, or ordinary implementation.

Explain behavior over time through stocks, flows, feedback, delays, and adaptive actors. Use for growth, decline, oscillation, bottlenecks, or policy resistance. Do not use to estimate a single causal effect, register an experiment, or forecast a scoreable event without a dynamic structure.

Design an app product whose promise is a recurring job — thesis, journeys, system shape, platform behavior, and validation. Use when the user asks to design an app, define an app concept, or shape a mobile, desktop, or web app. Do not use for a game, a SaaS web platform, a single UI flow, or a pricing catalog.

Write or revise agent-facing instructions: an Agent Skill, CLAUDE.md, AGENTS.md, system prompt, subagent prompt, or reusable prompt. Use when creating a skill, writing a subagent, tightening a trigger, encoding a gotcha, rightsizing CLAUDE.md, or rewriting prompts that still use MUST/NEVER recipes a capable model no longer needs. Do not use for user-facing product copy, store listings, or status updates.

Bound a request into one objective, one terminal (local, landed, released, or live), in-scope work, adjacent independently shippable outcomes, and hard cut lines. Use when the user mixes several deliverables, scope is expanding, or it is unclear what done means. Do not use to pick the next backlog item or to finish an already accepted multi-step objective.

Prepare and verify a signed product package for its distribution channels — install, update, uninstall, and live readback. Use when shipping a binary, installer, package, or store build. Do not use to write store listing copy or screenshots.

Design and verify production payments — checkout, provider events, ledger, entitlements, refunds, disputes, settlement, and finance close. Use when the user mentions Stripe, app-store billing, webhooks, invoices, chargebacks, or taking money in production. Do not use for seller marketplace payouts, first-time packaging, or a live price increase.

Build a missing product capability end to end on a real user path. Use when an accepted direction needs a usable slice across code, interface, data, and delivery. Use the product's current identity, storage, job, and event providers; do not invent a second runtime.

Design or implement a Telegram bot on the current Bot API. Use when the user mentions Telegram, Bot API, slash commands, inline keyboards, callback queries, Web Apps, sendRichMessage, or rich messages. Do not use parse_mode HTML/MarkdownV2 as the product text path. Do not use for Slack, Discord, or generic product UI.

Curate an Agent Skills repository by improving, merging, splitting, transferring, or retiring packages. Use for portfolio-wide overlap, routing, package quality, or historical recovery. Use author-skill for one package.

Assess a daily reward or streak system — clocks, eligibility, claims, missed days, and reward tables. Use when the user asks to review daily rewards, streaks, check-in bonuses, or retention loops. Do not use to design the system.

Fulfill one access, export, correction, deletion, restriction, or appeal request with proof and reconciliation. Use when a user asks to fulfill a data-rights request, a DSAR, a GDPR or CCPA request, or a specific access/delete ticket. Do not use to design the privacy lifecycle, and not for legal advice.

Design in-product ad placements and rewarded-ad grants. Use when the user asks about rewarded ads, mediation, ad-free purchase, or interrupting the core loop. Do not use for buyer checkout or store listings.

Design an offline sync and conflict protocol that preserves user intent across retries, multi-device edits, deletion, and partial failure. Use when offline is a declared product capability. Do not use for ordinary online CRUD or for a live incident.

Design a live subscription price increase so contract, notice, consent, billing channel, and no-action default stay truthful. Use when raising prices on existing contracts. Do not use for first-time packaging or for implementing checkout.

Turn a product uncertainty into a pre-registered experiment protocol with an estimand, assignment, guardrails, and a decision rule. Use when the result must be able to change a decision. A staged rollout is not automatically an experiment. Do not use to estimate an already-identified observational effect or to run a demo without assignment.

Design a referral or invite loop where attribution, qualification, and grant are separate facts. Use when growth through relationships must not become spam, coerced contact upload, or unearned reward.

Design the customer and entitlement consequence of a refund, cancellation, dispute, or chargeback. Use when spent value, access, and support action must stay consistent with payment truth. Do not use to own the money ledger.

Design a semantic taxonomy as an evolvable concept graph with durable identifiers, relations, and lifecycle. Use when shared meaning must survive rename, reparent, split, and merge.

Drive one already accepted engineering or product objective to its requested terminal. Use when completion matters more than a plan. Do not use to choose what to do next from a backlog.

Replace a predecessor with one destination and retire the old path in the same delivery. Use for system, API, package, instruction, installer, configuration, or data migrations that require a single final authority.

Finish an already integrated product to a declared target. Use when the promise is clear and the job is deep resolution of copy, states, feel, and consistency. Do not use to add a capability or to repair a live defect.

Produce an updateable probability or distribution for a resolvable future event using base rates, explicit assumptions, and a declared resolution source. Use when the user asks for a forecast, probability, confidence, or calibration. Do not use to estimate a past causal effect, register an experiment, or critique a present explanation.

Assess a game economy — sources, sinks, affordability, paid acceleration, and randomized rewards. Use when the user asks to review game economy, gacha, IAP, or virtual currency balance. Do not use to design the economy.

Design a game product whose promise is play, mastery, fantasy, or progression — loops, systems, feel, and validation. Use when the user asks to design a game, define a game concept, or shape game loops or systems. Do not use for an app, a SaaS web platform, a single UI flow, or a game economy review.

Assess a game soft launch — bounded real-player exposure, reversible spend, and market transfer. Use when the user asks to review a soft launch, regional test, or limited release plan. Do not use to plan or run the launch.

Govern who may grant, inherit, exercise, review, recover, and revoke administrative authority across customer tenants. Use when identity-provider assertions are inputs to product permission, not a substitute for it.

Safely persist in-progress state and write an actionable resumption brief for another machine, session, or developer. Use when handing off work, switching computers or workspaces, pausing mid-task, saving work-in-progress across environments, or transferring an unfinished objective. Do not use for routine status updates without environment transfer (use write-high-signal-update), fully completed deliverables ready to merge (use drive-to-delivery), or production cutovers (use execute-hard-cutover).

Build or slim a repository's required commit build around fast product-behavior checks. Use when CI is slow, flaky, duplicated, or blocking on prose, naming, or file-layout linters, or when the user asks what should be required on every PR. Do not use for CD, release trains, or incident response.

Assess one release candidate for a go/no-go, hold, or risk decision on the exact build. Use when the user asks whether a release, launch, or rollout is ready, or to review a launch plan. Do not use to plan or execute the launch.

Repair a product defect or regression at its owning layer. Use for live or dogfood failures, flaky paths, and current user harm. Do not use to finish or polish an already integrated product.

Assess marketplace trust operations — policy, evidence, enforcement, appeals, risk queues, and trust economics. Use when the user asks to review marketplace moderation, seller enforcement, risk scoring, or appeal flows. Do not use to design the enforcement system.

Migrate customer data so meaning, ownership, access, and history survive. Use when acceptance is preserved semantics, not row-count equality. Do not use for the authority switch of a live writer.

Model product abuse and a least-harm control ladder. Use when attackers, fraud, or policy evasion can take money, access, or other users' work. Do not publish an exploitation recipe.

Assess a notification strategy — stop conditions, cooldowns, cadence, quiet hours, and suppression. Use when the user asks to review notifications, push strategy, or messaging cadence. Do not use to design the notification system.

Design or improve the support operating model: contact taxonomy, routing, waiting-state clocks, authority, and quality loops. Use for the system, not for answering one ticket.

Operate marketplace seller payouts so money, holds, tax status, and failed transfers stay reconcilable. Use for two-sided payouts, not for buyer checkout.

Translate an operational decision into an explicit model — objective, constraints, uncertainty, and a recomputed solution — then stress it for proxy failure. Use when the user asks to optimize, allocate, schedule, or choose under constraints. Do not use to pick the next backlog item, run an experiment, or critique a claim.

Create a truthful store or channel listing and creative sequence for install, trial, or purchase. Use for metadata, screenshots, and conversion narrative. Do not use for signing or publishing the binary.

Choose SaaS plan architecture and list prices — value metric, seats vs usage, included amounts, overages, trials, and catalog migration. Use when pricing a product, designing a free tier, picking a meter, or repackaging plans. Do not use for raising an already-live price, or for implementing checkout.

Design a privacy lifecycle across collection, purpose, consent, retention, access, and deletion. Use when mapping personal data or building consent, retention, or deletion controls. Do not use to fulfill one specific data-rights request, and not for legal advice.

Produce engine-ready 2D tilesets, prop packs, layered maps, collision metadata, and composed previews. Use when the user asks for tiles, a tilemap, terrain, Tiled/Godot/Unity 2D maps, or seam-checked atlas props. Do not use for character animation sheets or store key art.

Produce engine-ready 2D character, VFX, HUD, and icon sprites, including animation sheets and loop QC. Use when the user asks for a sprite sheet, idle/walk/attack cycle, chroma-key frames, video-to-sprite, or a game atlas. Do not use for map tilesets, 3D models, or marketing screenshots.

Produce finished product media and design assets from a current brief and product truth. Use for screenshots, key art, store media, campaign visuals, motion, audio, localization, and accessible variants. Do not use for engine-ready game sprites or tilesets.

Run the cheapest real probe that can change a product decision, then kill, continue, or pivot from the observation. Use when the user wants a prototype, spike, or cheap test. A demo without a decision criterion is not this job. Do not use to register a production experiment or to polish an already accepted product.

Create or attach an isolated filesystem and shell for an agent — host workspace, sandbox, or VM — with explicit network, secret, lifetime, and cost bounds. Use when the user asks to provision a sandbox, code-execution environment, or throwaway VM. Do not use to write application features in an existing repo.

Structure a multi-participant deliberation and record a decision-ready result — options, strongest reasons, dissent, and the named owner's next action. Use when several perspectives, competing proposals, or material objections need a fair synthesis. Do not use for a single-analyst critique or a domain go/no-go review.