Skills, Agent Packs, and Knowledge Assets

Overview

This section collects the agent-facing skill packs that teach matrix-core how to reason about accessibility, agent evaluation, API design, architecture, autonomous operation, and adjacent knowledge workflows. The source is split into paired assets: a prose guide in SKILL.md and a MatrixScript manifest in SKILL.mtx. Together they define what the skill means, when to activate it, what the compiler should resolve, and how the skill should fail or clarify when the target is uncertain.

The most important pattern across this section is that the markdown files describe the human-readable policy and workflow, while the .mtx files provide the compiler-facing contract. That contract is intentionally narrow: a required slot.target, optional slot.constraints, a typed result, and explicit handling for unknown or ambiguous targets. The result is a small but consistent family of agent packs that can be composed for engineering, API, architecture, testing, and audit work.

Shared Skill Pack Format

All of the visible skill manifests follow the same shape:

• §SKILL declares identity and routing fields such as id, version, display, description, mcl.verbs, determinism, and seed_policy.
• §INPUTS defines the target artifact and optional constraints.
• §CORTEX declares the memory categories the skill may read.
• §TOOLS and §SUB_SKILLS are set to none in the visible packs.
• §PROCEDURE resolves slot.target with cortex.find, then clarifies if confidence is too low or the target cannot be identified.
• §OUTPUTS always includes a required slot result: ArtifactRef and optional slot unknowns: Unknown[].
• §FAILURE_MODES consistently uses target_not_found, ambiguous_after_clarify, policy_violation, and budget_exceeded.

Across the manifests, the compiler-facing behavior is reproducible: the packs are seedable, use per_intent seeding, and are organized around a single target artifact rather than open-ended conversation.

Common Execution Flow

Accessibility Skill Pack

skills/accessibility/SKILL.md frames accessibility as WCAG 2.2 Level AA work grounded in POUR: Perceivable, Operable, Understandable, Robust. It focuses on semantic mapping, accessibility-tree behavior, focus management, and labeling or hints for screen readers, switch controls, keyboard navigation, and native accessibility traits. The visible example shows a search form with a visually hidden label, a search input, and a submit button with aria-label set to Submit Search.

skills/accessibility/SKILL.mtx turns that guidance into a two-verb pack with find and build. The manifest expects slot.target: ArtifactRef, optionally accepts slot.constraints, and emits slot.result: ArtifactRef. If the target is unclear, it blocks with a clarification prompt instead of proceeding with a weak guess.

Agent Evaluation and Regression Assets

skills/agent-eval/SKILL.md is a lightweight comparison tool for coding agents such as Claude Code, Aider, and Codex. Its core goal is to make agent comparisons reproducible by defining tasks in YAML, then judging pass rate, cost, time, and consistency across repeated runs. The workflow is explicitly evaluation-first: define tasks, run them, and compare the metrics rather than relying on intuition.

skills/ai-regression-testing/SKILL.md complements that with regression protection for AI-assisted development. It starts with mandatory automated checks, specifically npm run test and npm run build, and then moves into a bug-check review flow that looks for sandbox and production consistency, API response shape mismatches, rollback gaps, and optimistic-update races. The visible helper functions show how test requests are built: createTestRequest can add x-sandbox-user-id and JSON content type, while parseResponse returns the response status and parsed JSON.

Agent Engineering Skill Packs

Harness Construction and Introspection

skills/agent-harness-construction/SKILL.md is about improving how an agent plans, calls tools, recovers from errors, and converges on completion. Its guidance is explicit about action-space design, observation formatting, error recovery contracts, context budgeting, and the hybrid ReAct plus function-calling pattern. It also defines the observation fields the tool should return: status, summary, next_actions, and artifacts.

skills/agent-introspection-debugging/SKILL.md is a self-debugging workflow for repeated failures, runaway retries, or context drift. It divides the process into capture, diagnosis, contained recovery, and report generation. The visible failure patterns include ECONNREFUSED, 429, stale diffs after writes, and tests that still fail after a supposed fix, with each pattern tied to a minimal corrective action.

Agentic Engineering and Payment Controls

skills/agentic-engineering/SKILL.md is the higher-level operating model for AI-led implementation work. It emphasizes defining completion criteria before execution, breaking work into agent-sized units, routing model tiers by task complexity, and measuring success with evals and regression checks. It also tracks cost discipline with model, token estimate, retries, wall-clock time, and success or failure.

skills/agent-payment-x402/SKILL.md adds autonomous payment execution over x402. It explains budget controls, non-custodial wallets, and spending policy enforcement for agents that need to pay for APIs, services, or other agents. The visible example centers on set_policy, check_spending, preToolCheck, and a fail-closed entrypoint via main().catch((err) => {}). The pre-tool logic explicitly handles invalid input, transport failures, tool errors, parse failures, and budget exhaustion before the paid action runs.

API Design and Documentation Skill Packs

API Design

skills/api-design/SKILL.md is a REST design guide for consistent resource naming, status codes, pagination, filtering, sorting, error responses, versioning, and rate limiting. The document includes interface examples for ApiResponse<T> and ApiError, plus a class example for UserViewSet. The examples show how request validation, serializer choice, and response shaping fit together in a controller-style API.

The visible symbol evidence for the embedded examples is summarized below.

ApiResponse

ApiError

UserViewSet

skills/api-design/SKILL.mtx packages that content as a build-oriented manifest. It uses a single target artifact, resolves it with Cortex, and then asks for clarification if confidence is low. The markdown file’s examples are design patterns, while the manifest turns the pattern into a compile-time skill shape.

API Documentation

skills/api-documenter/SKILL.md is the documentation counterpart. It is focused on OpenAPI 3.1, complete endpoint coverage, request and response examples, error documentation, authentication guides, versioning, and interactive documentation experiences such as try-it-out consoles and environment switching. It also spans documentation forms beyond REST, including GraphQL, WebSocket, gRPC, webhook events, SDK references, CLI docs, and integration guides.

skills/api-documenter/SKILL.mtx exposes the same style of compiler contract as the other packs. The visible example payload around the documentation workflow includes progress fields such as status, progress, endpoints_documented, examples_created, and sdk_languages, which makes the pack suitable for documentation runs that need measurable output.

API Connector Builder

skills/api-connector-builder/SKILL.md is for adding a repo-native connector or provider without inventing a new integration architecture. Its workflow is explicit: inspect at least two existing connectors or providers, map the house style, narrow the target integration, build the connector in the repo’s existing layers, and validate the result against the source pattern. The visible guidance calls out config, auth, retries, pagination, registry hooks, and tests as the pieces that must line up.

skills/api-connector-builder/SKILL.mtx is the build manifest for that workflow. It is a target-driven pack with the same Cortex resolution and clarification behavior used by the other manifests.

Architecture and ADR Skill Packs

skills/architecture/SKILL.md is the monorepo architecture guide. It focuses on structure, package relationships, import rules, and component organization. The document defines platform-specific file naming, workspace import boundaries, and a strict hierarchy that prevents circular dependencies. It also sets an analysis protocol: assess scope impact, verify patterns, check architecture integrity, and evaluate performance impact before modifying code.

skills/architecture-decision-records/SKILL.md captures why decisions were made. It activates when a decision is being chosen or documented, or when someone asks why a past choice was made. The ADR format is lightweight and the lifecycle guidance is explicit: record the original date for past decisions and keep superseded decisions linked to their replacements.

Autonomous Operation and Audit Packs

skills/autonomous-agent-harness/SKILL.md describes how to turn Claude Code into a persistent autonomous agent system using native crons, dispatch, MCP tools, memory, and task queuing. The architecture snippet shows command and hook surfaces feeding an MCP server layer with services such as memory, github, exa, supabase, and browser-use. The prose also shows how task queues are structured with active and completed work, which makes the skill useful for continuous monitoring and scheduled operations.

skills/automation-audit-ops/SKILL.md is an evidence-first operator skill for inventorying automations before changing them. It separates configured, authenticated, recently verified, stale or broken, and missing surfaces; it starts read-only; and it refuses to claim a tool is live just because a skill or config references it. Its workflow begins with a real-surface inventory and only then moves to keep, merge, cut, or fix-next recommendations.

Analysis Support Assets

skills/ast-grep-code-analysis/SKILL.md is a structural analysis skill for complex codebases. It is intended for security vulnerabilities, performance issues, and structural patterns where manual inspection is too shallow. The visible guidance prefers AST-based matching over line-by-line reading and notes a prerequisite around ast-grep installation and configuration.

skills/ast-grep-code-analysis/SKILL.mtx packages that analysis workflow as a manifest with the analyze verb, target resolution, and clarification when the target is ambiguous. It belongs in the same family as the other agent-oriented skills even though its output is analysis rather than implementation.

Source-Backed Skill Inventory