abvx-agent-skills

Define and review concrete agent capability cards, asynchronous operation receipts, decision receipts with revalidation, isolated and trust-graded memory scopes, and evidence-backed provider or tool registry entries. Use when a project introduces named agents, scheduled or long-running work, durable decisions or memory, provider routing, MCP tools, or external-action approval boundaries.

Review tools, MCP functions, CLI commands, SET inputs, and AGENTS.md generator outputs as agent-facing contracts. Use when a tool description, schema, workflow input, or generated instruction may be ambiguous, unsafe, noisy, or hard for agents to call correctly.

Build or review gesture-driven web interactions such as drag, swipe, sheets, carousels, and draggable panels. Use when pointer tracking, interruption, velocity handoff, momentum projection, hysteresis, rubber-banding, spatial continuity, or motion/transparency/contrast fallbacks materially affect the interaction.

Review frontend motion and animation code before shipping, with strict checks for purpose, frequency, easing, duration, transform origin, interruptibility, GPU-safe properties, reduced-motion support, and hover behavior. Use when a diff touches transitions, keyframes, Framer Motion, Lottie, drawers, popovers, toasts, hover or press states, gesture motion, or branded motion assets.

Set and review a bounded visual direction for marketing, editorial, portfolio, and brand-forward frontend work. Use when a surface needs an explicit design read, calibrated composition/motion/density, preservation-first redesign, layout-rhythm review, or desktop/mobile/reduced-motion verification. Route product UI to Lazyweb and UX review instead.

Harden repeated delivery loops with evidence-based harness stripping, immutable runtime-path sprint packets, and broken-window revalidation without automatic revert. Use for Cardputer, browser, CI, runtime, or production work split into repeated sprints.

Review whether a recurring or semi-autonomous agent loop is ready to run. Use before scheduling Codex, Claude Code, Grok, Opencode, CI, issue triage, PR babysitting, dependency sweeping, changelog drafting, or other repeated agent workflows.

Review an implemented product UI for hard visual defects, design-system incoherence, and repeated template patterns without treating subjective taste as universal law. Use for pre-ship screenshots or browser review when a landing page, product surface, dashboard, or redesign may be generic, clipped, inaccessible, visually inconsistent, or overbuilt from familiar AI-generated patterns.

Capture and classify minimal red-to-green evidence for a diagnosed bug and scoped fix, with command hashes, Git/environment identity, broader-check records, risk-based approval, route evidence, and optional cpat linkage. Use after diagnose establishes a credible reproducer or when a reported regression needs an auditable proof packet; do not use as an autonomous bug scanner.

Debug coding failures through reproduction, ranked hypotheses, narrow fixes, and verification. Use when a bug, regression, flaky test, performance failure, or inconsistent behavior is broken, slow, throwing, or hard to reproduce and the task needs disciplined diagnosis instead of guessed fixes.

Create or review minimal Git-native project context using only adr, rfc, rule, spec, plan, rnd, and cpat documents; human-gated draft/accepted/rejected lifecycle; implements, depends_on, extends, and related relations; and evidence-backed code-change patterns. Use when a project needs durable decisions, specifications, research, plans, incident lessons, or the SET git-native-context capability without installing a context runtime.

Execute feature work and bug fixes with a tracer-bullet red-green-refactor loop. Use when building behavior that should be guided by tests, when fixing regressions, or when the user wants TDD or test-first execution. Prefer public interfaces, vertical slices, and behavior-focused tests that survive refactors.

Coordinate an opt-in Planner, Reviewer, and bounded Executor workflow with stable findings, at most five review rounds, disjoint write ownership, explicit route evidence, and root-owned verification. Use when the user asks for plan review, multi-agent orchestration, routed executors, parallel implementation, or the SET bounded-orchestration capability.

Design small evaluation gates for agent outputs, skill changes, prompts, and tool contracts. Use when a change needs rubric checks, pairwise comparison, position-bias mitigation, regression fixtures, or LLM-as-judge discipline without building a heavy eval platform.

Review agent work for context poisoning, lost-in-the-middle failures, distraction, context clash, and stale carryover. Use before long-session handoffs, SET runner handoffs, skill changes, memory writes, or any task where too much or conflicting context may degrade behavior.

Use files as durable context without polluting the live prompt. Apply when a task needs scratchpads, plan persistence, retained outputs, proposal folders, handoffs, or repo-local context that agents can discover and audit later.

Generate diverse hypotheses, explanations, or solution candidates before evidence review. Use for research, incident analysis, market or product reasoning, adversarial review, and SkillOpt proposal generation when first-answer mode collapse would hide alternatives.

Enrich a Markdown knowledge base with durable tags, source provenance, existing-note links, and reviewed wiki synthesis while keeping raw notes read-only by default. Use when organizing Obsidian-style notes, refreshing generated wikis, or preparing source-grounded context for agents.

Run risky or multi-file agent work as a reviewable proposal before mutating the target workspace. Use when a task should produce retained output, inspectable diffs, and an explicit select/apply/discard decision instead of direct edits.

Evolve Codex skills and agent instructions with a SkillOpt-style loop. Use when creating, refining, or auditing local skills, AGENTS.md rules, prompt playbooks, or reusable agent workflows after repeated task experience, especially when the user asks to "improve your skills", "capture this lesson", "make this reusable", or "update agent instructions".

Surface hidden assumptions in plans, specs, AGENTS.md, SKILL.md, SET bundles, and workflow contracts before implementation or handoff. Use when an artifact looks plausible but may rely on unstated environment, dependency, behavioral, temporal, or success assumptions.

Review whether an artifact's confident claims are backed by evidence, tests, contracts, or operational safeguards. Use before shipping, publishing docs, accepting generated plans, or trusting a workflow that may present fragile assumptions as certainty.

Choose and run a compact pre-implementation, post-implementation, or ship-readiness gate for agent work. Use when a task needs ordered review lenses, proof gates, or release confidence without adopting a heavy multi-agent pipeline runtime.

Use local Rabbithole for human-in-the-loop exploration of AGENTS.md, SKILL.md, SET plans, repomaps, and long repo docs. Use when the user wants to inspect, question, branch, or review dense agent-facing documents interactively before changing durable docs or workflow contracts.

Design, audit, or improve agentic systems and agent harnesses. Use for agent MVP blueprints, tool design, permission models, planning and goal loops, context compaction, memory, skills, connectors, observability, evals, safety, and provider-neutral agent architecture.

Plan and run large Codex tasks with compact dynamic workflow packets, risk gates, integration, and verification. Use when the user explicitly asks for dynamic workflows, packets, subagents, swarm, orchestration, long-running audits, multi-track coding/research work, job or client search agent design, migrations, repo-wide reviews, or tasks where independent discovery, implementation, review, and verification tracks would reduce drift.

Design or maintain a local personal AI workspace with a root router, isolated domain folders, user-triggered memory, decision logs, and sparse routing corrections. Use when a user wants one agent workspace spanning multiple projects or life/work domains without mixing context or bloating every session.

Review a codebase for architecture deepening opportunities: shallow modules, weak seams, scattered domain logic, low testability, high coupling, and AI-navigability friction. Use when the user asks for architecture improvement, refactoring opportunities, modularity, test seams, or maintainability review.

Stress-test a plan, feature, or design against the repo's existing docs, domain language, ADRs, and context files. Use when a task is ambiguous, terminology is fuzzy, architecture trade-offs matter, or the user wants alignment before implementation. Ask one question at a time, inspect the codebase when answers are already present, and sharpen the project's language as decisions crystallize.

Review and route social content before publishing. Use when drafting, adapting, auditing, scheduling, or monitoring posts for LinkedIn, X, Threads, Telegram, newsletters, or other distribution channels where brand voice, factual claims, platform fit, approvals, and no-spam guardrails matter. Enforce draft -> audit -> approval -> publish -> monitor, and never post or message externally without explicit user approval.

Route an idea, vague feature request, or agent-built product change through delivery gates from intent to release. Use when work needs more discipline than a readiness check: clarify the spec, slice the work, protect architecture boundaries, gather proof, converge through review, and prepare release without turning the task into a heavyweight process.

Route reusable agent learning after an error, success pattern, review finding, or repeated workflow. Use when deciding whether an improvement belongs in the model prompt, memory/context note, durable docs, SKILL.md, checklist, script/tool, eval, golden fixture, or should be rejected as overfit.

Run a baseline preflight before long implementation, autonomous execution, or publishing a PR. Use when a task has multiple phases, mandatory commands, high risk of thrashing against an already-broken repo state, or should be checked in an isolated push gate before reaching the real remote. Deduplicate the minimum useful checks, surface red baseline conditions early, and require an explicit user decision before proceeding past known breakage.

Keep repo-local agent context accurate after code and workflow changes. Use when docs/ai, AGENTS.md, runbooks, or context entrypoints have drifted from the codebase, after major feature work, after architecture changes, or when an agent keeps rediscovering the same repo facts. Refresh durable context, split groups only when needed, and preserve compact read order.

Compile a raw agent goal into a bounded loop harness before execution. Use before launching long Codex, Claude Code, MiMo Code, or local-LLM coding loops when the goal needs clearer stop rules, verification rubric, judge prompt, budget policy, retry limits, memory policy, and portable YAML/JSON/Mermaid artifacts.

Route authorized defensive security and reverse-analysis tasks by scope, target type, user intent, and available toolchain before taking action. Use for APK review, static binary triage, frontend JavaScript security review, firmware inventory, malware/YARA analysis, API/LLM/supply-chain audit, security report planning, or security issues that need evidence-led classification without offensive exploitation.

Convert books, papers, long documents, or prior reading notes into compact reusable agent skills. Use when the user provides a PDF, EPUB, DOCX, HTML, Markdown, text file, article bundle, or structured notes and wants reusable workflows, mental models, frameworks, checklists, anti-patterns, or a public/private SKILL.md artifact rather than a summary.

Verify frontend changes in a real browser, including layout, console errors, responsive states, network failures, screenshots, and interactions. Use after local or deployed web UI changes, visual fixes, forms, navigation, asset updates, canvas or 3D work, or whenever browser verification should replace static guessing.

Evaluate whether Doc-to-LoRA is the right path for turning documents into parametric memory, and guide a safe proof-of-concept before investing in a full plugin or training pipeline. Use when a user wants to internalize long documents, reduce repeated context costs, compare Doc-to-LoRA against RAG or long-context prompting, or assess hardware/model/data risks for document-to-adapter workflows.

Research and answer evidence-sensitive questions with source, date, unit, and operand discipline. Use for web research, PDF/document QA, policy or finance facts, tables, time series, calculations from retrieved evidence, exact-answer extraction, and any task where wrong source selection, stale facts, nearby table confusion, or formatting errors would matter.