Comparison with the original compound agent-native skills
What the original skills did well
The deprecated Compound package was opinionated in a good way. It made a strong claim that agent support is not a bolt-on chatbot: the user and the agent should share action paths, context, and workspace state.
The most valuable original ideas preserved in v2 are:
- Action parity: meaningful user actions need agent paths unless intentionally human-only.
- Context parity: agents need the same domain objects, state, recent activity, and constraints a human operator sees.
- Shared workspace: agent work should land in durable, inspectable records/files/branches/drafts rather than hidden sandboxes.
- Primitives over workflows: expose composable capabilities; use workflow tools only for justified atomicity, safety, or external orchestration.
- Dynamic context injection: static prompts are not enough; runtime state and capabilities should be injected.
- Noun test: for every domain object, verify the agent can understand, discover, inspect, act where appropriate, and verify.
- Human-only exceptions: CAPTCHA, MFA, OAuth consent, legal acceptance, biometric prompts, and platform trust gates are not parity failures.
- Mobile/offline and self-modification guardrails: checkpoints, local files, review, rollback, and safe evolution matter.
What was narrow or brittle in the original package
The original package focused mainly on app architecture and code review patterns. It was excellent for reviewing UI/tool parity inside a product, but weaker for:
- Public web/docs agent readiness.
llms.txt, markdown negotiation, robots/sitemap/content policy, API catalog, and OAuth metadata.- CLIs/TUIs as first-class agent surfaces.
- HTTP APIs, SDKs, schemas, error envelopes, pagination, idempotency, and versioning as agent contracts.
- Agent Skill packaging quality, trigger evals, scripts, templates, and progressive disclosure.
- File/workspace, mobile/offline, and long-running job patterns as reusable audit surfaces.
- Runnable local scanners, validators, scaffolding scripts, schemas, and example outputs.
- Separating “documentation says this exists” from “implementation evidence shows this exists.”
How v2 improves it
Version 2 turns the original principles into a cross-surface audit/design skill. It keeps action/context parity as the central architecture principle, then adds surface-specific contracts for websites and docs, CLIs and TUIs, APIs and SDKs, MCP/A2A/tools, Agent Skills and repository instructions, app UIs and shared workspaces, files, mobile/offline, long-running jobs, checkpoints, and self-modification.
It also adds operational artifacts the original lacked:
- A local scanner with implementation-vs-doc evidence separation and generated-report exclusions.
- A web readiness checker for common agent discovery/capability endpoints, including
llms.txt, API catalog Linkset, OAuth/OIDC metadata,/.well-known/mcp.json, A2A cards, and agent skill indexes. - An
llms.txtgenerator. - An action-parity inventory script that turns UI/action/tool evidence into a candidate capability map.
- An agent-asset scaffolder.
- A skill/asset validator.
- JSON schemas and templates.
- Trigger and output-quality eval seeds.
- Example validation and self-audit outputs.
The result is not just “review a product for agent-native architecture”; it is “audit or design any useful agent-facing surface and give engineers patchable fixes.”
What did not change
The skill still treats generic advice as low quality. Findings should be evidence-backed and implementable. Human approval is allowed and often required. Safety protocols are part of parity, not obstacles to it.
v2 correctness fixes over v1 of this package
The v2 pass fixed concrete packaging and tooling defects: stale command references, validation path handling for . skill directories, missing output-directory creation, recursive self-audit evidence from generated reports, hidden .well-known discovery being skipped, ad-hoc API catalog JSON, missing /.well-known/mcp.json, and one-size-fits-all web scoring that penalized content-only sites for not publishing API/app endpoints.