AndrewDryga

Author, validate, test, distribute, and certify a custom Emisar action pack for a customer fleet. Use when wrapping internal CLIs or services as declared actions, writing or reviewing pack/action YAML, testing a pack locally, rolling a pack out to runners, hosting a private pack registry with packctl, publishing a new pack version, retiring an unsafe version, or diagnosing pack trust and hash-mismatch issues.

Connect an LLM or MCP client to Emisar and certify the connection end to end. Use for installing the emisar-mcp stdio bridge, registering a local client (Claude Code, Cursor, Zed, Copilot CLI, and other stdio clients), wiring a cloud connector (Claude.ai, ChatGPT), repairing a broken client registration, or proving that discovery, action execution, and audit work through the configured client. For installing the on-host runner itself, use the install-emisar skill.

Install, configure, repair, and certify the Emisar runner on a host end to end for customer onboarding. Use for first-run setup, runner installation, guided pack discovery and selection, host capability coverage, custom-pack recommendations, pack credentials, upgrades, supervised-service migration, authenticated MCP dispatch, or any request to diagnose and report runner health across the host, control plane, registry, action execution, and audit trail. After runner setup, reuse the current authenticated Emisar MCP connection or offer to connect the operator's agent through connect-llm; when a required host job has no suitable pack, offer the public author-pack workflow.

Investigate, contain, and permanently remediate production incidents through Emisar. Use when an operator asks an agent to investigate production, explain why a service is degraded, stop the bleed, stabilize an environment, or fix an infrastructure incident; gather evidence, test hypotheses, use the smallest authorized action, fall back to explicitly approved break-glass access only when Emisar has no viable path, turn the root cause into a source-controlled application or infrastructure-as-code fix, verify the operator's deployment, iterate until recovery, and write the incident report. An investigation-only request authorizes observation, never a production-changing or break-glass action.

Write and edit clear, distinctive, human-quality content for emisar across marketing pages, product UI, documentation, blog posts, announcements, SEO pages, CTAs, and positioning. Use when copy sounds corporate, generated, vague, overcomplicated, over-marketed, tonally flat, or inconsistent; when a surface needs a confident, direct, friendly professional voice; or when approved writing samples should become a reusable brand voice.

Drain a project's .agent/tasks/ queue autonomously and run-to-completion, taking EACH task to a ship-ready bar — `coop tasks claim` a 00_todo/ task, build it, gate it green, self-review it against the house rules (portal Iron Laws / runner security posture / pack conventions + the .agent/rules KB) AND every hat (security, UX, marketing, docs, code quality, tests), ITERATE until it's clean, then COMMIT IT ON ITS OWN and `coop tasks done` — without quitting early. A Stop hook scoped to this skill blocks quitting while any task is actionable. Use to "work all the tasks" / drain a backlog to a high bar / run an unattended sweep.

Put on the SEO/marketing hat for the emisar marketing site and positioning — clear honest value prop, crawlable server-rendered pages, titles/meta/structured data, internal linking, and sitemap. Use when editing controllers/marketing_html (home, pricing, security, use-cases, compare, docs, packs), writing positioning/copy, or improving how pages rank and convert.

Orchestrate an agency-grade redesign of the emisar marketing website: positioning, creative territories, art direction, content architecture, visual system, implementation loop, and review. Use when redesigning or substantially changing marketing pages, landing pages, positioning, launch pages, or any public site work that must feel distinctive rather than template-generated.

Drain this repo's .agent/tasks/ queue autonomously and run-to-completion, taking EACH task to a ship-ready bar — claim a 00_todo/ task (`coop tasks claim`), build it, gate it green, self-review it against the .agent/rules KB and every hat, ITERATE until clean, COMMIT it on its own, then `coop tasks done` — without quitting early. Use to "work all the tasks" / drain a backlog / run an unattended sweep.

Cut a versioned release of coop — derive the semver bump from the `## Unreleased` changelog, finalize + annotated-tag it, then (only on explicit confirmation) push the `v*` tag that triggers the public GoReleaser build. Refuses a docs-only no-op release. Use when asked to "cut / ship / tag a release", "publish a release", or "bump the version".

The heavyweight pre-merge review — convene a board of expert hats (correctness, this repo's own rules, security, plus PM/UX/maintainer and any hat the change earns) as parallel reviewers, then synthesize ONE ranked verdict and an ordered plan to fix everything. Reads the project's OWN `.agent/rules/` and runs its OWN gate — no hardcoded laws. Read-only. Works on local changes, a commit, a range, a branch/ref, or a PR. Use before landing anything substantial, or when you want a thorough multi-perspective review.

Execute an approved plan step-by-step, with the repo's gate between steps and no scope creep. Use when implementing a planned change, working a checklist, or the user says "go" / "implement it" / "do the plan". Stops and reports on the first red gate.

Design an opinionated, boring-by-default implementation plan before writing code. Use when a change spans more than one file or context, when the approach isn't obvious, or when asked to plan/design/think through a feature. Produces a small-step, verifiable plan to approve before /work.

Root-cause a crash, panic, failing test, or wrong behavior — reproduce, read the real error, trace to the actual cause, then propose the minimal fix. Use when something fails or misbehaves and you need the cause, not a guess. Never edit a test to make a real failure pass.

Confirm a function, argument, option, type, callback, or CLI flag actually exists with the signature you think — before you call it. Use whenever you're about to use an API you're not certain of, or any time you'd otherwise be guessing. Stops hallucinated calls before they're written.

Design an opinionated, boring-by-default implementation plan before writing code. Use when a change spans more than one file or context, when the approach isn't obvious, or when asked to plan/design/think through a feature. Produces a small-step, verifiable plan to approve before /work.

Drain this repo's .agent/tasks/ queue autonomously and run-to-completion, taking EACH task to a ship-ready bar — claim a 00_todo/ task (`coop tasks claim`), build it, gate it green, self-review it against the .agent/rules KB and every hat, ITERATE until clean, COMMIT it on its own, then `coop tasks done` — without quitting early. Arms the Stop-hook sentinel. Use to "work all the tasks" / drain a backlog / run an unattended sweep.

Put on the product-manager hat for OnlyTTY — decide what to build, what to cut, and what order; define the smallest valuable slice and what "done" means. Use when a request is vague or oversized, when scoping a feature, prioritizing, or deciding whether something is worth building at all. Pairs with /spec.

Put on the SEO/marketing hat for the OnlyTTY site and positioning — clear honest value prop, crawlable server-rendered pages, titles/meta/structured data, internal linking, sitemap, and the /control/:slug long-tail tool pages. Use when editing server/lib/relay_web/site (page.ex, tools.ex), writing positioning/copy, or improving how pages rank and convert.

Put on the UI/UX + visual-design hat for OnlyTTY — the dark-first marketing site and the mobile terminal viewer. Make it beautiful, calm, fast, and trustworthy: protect the design system in site.css, design every empty/loading/error/offline state, and hold a Linear/Vercel/Stripe-grade aesthetic bar. Use when designing or critiquing any page, screen, component, or flow in server/lib/relay_web/site or server/priv/static (viewer.html, app.js, site.css).

Execute an approved plan step-by-step, with the repo's gate between steps and no scope creep. Use when implementing a planned change, working a checklist, or the user says "go" / "implement it" / "do the plan". Stops and reports on the first red gate.

Run an agency-grade art-direction pass for Offloader's public and customer-facing surfaces — creative territories, visual concept, layout grammar, type/color/motion system, and a rendered implementation loop that stays distinctive rather than template-generated. Use when designing or substantially changing positioning, the eventual portal marketing site, pilot collateral, diagrams, or the ROI report. Pairs with content-director for the words.

Write and review top-tier positioning and customer-facing copy for Offloader — messaging, docs narrative, pilot collateral, ROI-report framing, page copy, proof sections, objection handling, and anti-AI voice. Use with art-director on any public or customer-facing surface, and whenever copy must be precise, credible, and provably true rather than generic, templated, inflated, or AI-generated.

The heavyweight pre-merge review — convene a board of expert hats (correctness, this repo's own rules, security, plus PM/UX/maintainer and any hat the change earns) as parallel reviewers, then synthesize ONE ranked verdict and an ordered plan to fix everything. Reads the project's OWN `.agent/rules/` and runs its OWN gate — no hardcoded laws. Read-only. Works on local changes, a commit, a range, a branch/ref, or a PR. Use before landing anything substantial, or when you want a thorough multi-perspective review.

Drain this repo's .agent/tasks/ queue autonomously and run-to-completion, taking EACH task to a ship-ready bar — claim a 00_todo/ task (`coop tasks claim`), build it, gate it green, self-review it against the .agent/rules KB and every hat, ITERATE until clean, COMMIT it on its own, then `coop tasks done` — without quitting early. Arms the Stop-hook sentinel. Use to "work all the tasks" / drain a backlog / run an unattended sweep.