agents

Read credentials and secrets from Proton Pass through pass-cli with scoped agent sessions and non-disclosing handoffs. Use when retrieving passwords, TOTP codes, pass:// references, attachments, or other vault fields; logging into a website or application; running commands with Proton Pass secrets; inspecting vault, item, share, agent, or access-token metadata; using Proton Pass SSH keys; or diagnosing pass-cli installation, authentication, session, and permission failures.

Read-only review of a PR, branch, diff, or worktree changes for correctness, requested behavior, regressions, scope creep, proof gaps, and merge readiness. Use when asked to review changed code or decide whether work is safe to merge. Route over-engineering-only reviews to whittle-review and repository-wide structural scans to improve-codebase-architecture.

Review for over-engineering without applying fixes: find what to delete, replace with stdlib/native, or shrink. Use when asked to review a diff or audit a repo for unnecessary complexity, bloat, deletion candidates, or YAGNI violations. Correctness, requested behavior, merge readiness, security, and performance are out of scope; route ordinary code review to code-review.

Write, rewrite, critique, and manage high-taste technical program and engineering documents. Use when asked for better docs, specs, RFCs, decision memos, launch or readiness docs, completion records, Google Docs-ready drafts, less LLM-y writing, senior TPM or staff-engineer polish, document pruning, audience fit, decision clarity, or durable knowledge management.

Build, refine, and synthesize ChatGPT Pro consultation prompts. Use when the user asks for a paste-ready Pro prompt, wants Pro to audit, review, or critique a repo or plan, needs a follow-up after a Pro response, wants Pro feedback checked against local evidence, or wants a manual Pro consultation loop without browser monitoring.

Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression.

Design, critique, polish, or rework a frontend interface -- websites, landing pages, dashboards, product UI, components, forms, onboarding, or empty states. Use for UX, visual hierarchy, accessibility, performance, typography, layout, color, and motion review; making bland designs bolder or loud ones quieter; live browser iteration on UI; or ambitious visual effects. Not for backend-only or non-UI tasks.

Break an approved plan or parent GitHub issue into vertical-slice child issues when the user explicitly asks to slice, split into issues, or create grabbable tickets. Do not use during initial alignment before the user approves a breakdown.

Internal guidance for shaping focused Codex prompts and preserving deliberate model and reasoning choices when Claude forwards coding, review, diagnosis, or research work

Verify Xcode 27, Command Line Tools, Apple SDKs, simulators, devices, XcodeBuildMCP, Xcode native agent access, and Icon Composer 2 before Apple platform work. Use when setup, DEVELOPER_DIR, SDK, runtime, simulator, device, MCP, or toolchain drift could affect macOS, iOS, iPadOS, watchOS, tvOS, or visionOS builds.

Design and implement App Intents, App Entities, App Shortcuts, Spotlight, Siri, widgets, controls, and app-routing handoffs for Apple apps. Use when exposing app actions or content to system surfaces across iOS, iPadOS, macOS, watchOS, widgets, Shortcuts, Siri, or Spotlight.

Build, run, test, launch, log, screenshot, and debug Apple apps with Xcode 27, XcodeBuildMCP, xcodebuild, SwiftPM, simulators, devices, and macOS paths. Use when launching an app, proving a build, inspecting runtime behavior, wiring a Codex Run button, or diagnosing Apple build/test/runtime failures.

Design and implement Apple Foundation Models features with availability checks, guided generation, tool calling, prompt/transcript hygiene, privacy boundaries, and Instruments proof. Use when adding on-device Apple intelligence, local model summarization, extraction, classification, or tool-backed reasoning to an Apple app.

Localize and accessibility-test Apple apps with string catalogs, assets, previews, locale/layout proof, Dynamic Type, VoiceOver, contrast, motion, keyboard, switch/control access, and Xcode 27 localization workflows. Use when preparing Apple app UI for multiple languages, accessible interaction, App Store quality, or regression-proof visual behavior.

Diagnose Apple app launch/runtime performance, SwiftUI invalidation, hangs, Instruments traces, Top Functions, ETTrace profiles, dSYMs, memgraphs, leaks, retain cycles, and Foundation Models traces. Use when an Apple app is slow, janky, memory-heavy, leaking, hanging, or needs before/after profiling evidence.

Mirror iOS Simulator into the Codex in-app browser and render SwiftUI package previews with hot reload. Use when the user needs browser-visible Simulator proof, interactive iOS app viewing, SwiftUI preview iteration outside Xcode Canvas, or screenshot evidence from the Codex browser.

Diagnose Apple signing, entitlements, capabilities, provisioning, device install, sandbox, hardened runtime, Gatekeeper, notarization, TestFlight, and App Store validation issues. Use when builds or launches fail because of trust, identity, entitlement, provisioning, packaging, or distribution state.

Build and refactor Apple SwiftUI, AppKit interop, Liquid Glass, navigation, windowing, settings, tabs, toolbars, Observation, previews, and platform-native UI structure. Use when shaping Apple app presentation surfaces across macOS, iOS, iPadOS, watchOS, tvOS, or visionOS.

Audit and operate a local Parallels Desktop Windows VM from Codex. Use when testing Codex access to Windows through Parallels, choosing CLI vs GUI automation, tuning VM performance/resource settings, or scrutinizing Mac/Windows integration features for MSP-style work.

Maintain AGENTS.md, CLAUDE.md, CODEX.md, and similar repo instruction networks. Use when asked to create, edit, audit, prune, harmonize, compress, or route agent-facing guidance; when guidance files feel bloated, narrative, stale, duplicated, or too generic; when deciding whether a rule belongs in root instructions, nested owner instructions, a skill, a spec, an issue, ordinary docs, or a memory store; or when the user asks to scrutinize opportunities for new instruction files without defaulting to adding more files.

Force the simplest solution that actually works: question whether the task needs to exist, reuse existing code, prefer stdlib/native features, and keep the diff small. Use when asked to simplify an implementation, take the lazy path, apply YAGNI, or remove over-engineering while doing the work.

Perform language and framework specific security best-practice reviews, reports, and secure-by-default implementation guidance. Use only when the user explicitly asks for security best practices, a security review/report, or secure coding help for supported languages: python, javascript/typescript, or go. Do not trigger for ordinary code review, debugging, or non-security work.

Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.

Create durable user-facing presentation artifacts, usually self-contained HTML, for research synthesis, decision support, planning, code understanding, reviews, dashboards, and briefing surfaces. Use when the user asks for an artifact, dashboard, visual synthesis, decision aid, interactive checklist, or a revisitable surface where Markdown would lose important relationships. Do not use for ordinary summaries, repo source-of-truth docs, production UI implementation, stateful teaching workspaces, skill edits, or agent guidance.

Interview the user about a plan, design, implementation, or decision until the action-changing branches are resolved. Use when the user wants to stress-test an idea, pressure-test tradeoffs, expose blind spots, get grilled on a proposal, compare options, or explicitly says "grill me"; ask no more than two sharp questions at a time and stay project-agnostic while using provided artifacts as evidence when useful.

Write a session handoff to OS temp when the user asks to continue in a new thread or pass work to another agent.

Architecture smell scan when the user explicitly asks for refactor targets, deep-module candidates, or an architecture review. HTML report in OS temp; stop after report unless the user picks a candidate.

Resolve an in-progress git merge or rebase conflict. Use when Codex sees conflict markers, unmerged paths, rebase/merge state, or the user asks to finish a conflicted merge/rebase while preserving both sides' intent and project checks.

Create and maintain a stateful teaching workspace for multi-session learning, with missions, resources, lessons, references, and learning records. Use when the user asks to learn a topic over time, build a course-like lesson set, or continue an existing teaching workspace. Do not trigger for one-off explanations unless the user asks for a durable learning workspace.

Reference for writing, editing, and auditing Codex-compatible skills. Use when creating or refining skill text, pruning descriptions, deciding what belongs in SKILL.md versus references, diagnosing skill failure modes, or when another skill needs quality vocabulary for invocation, hierarchy, no-ops, sediment, sprawl, and predictable behavior.