Reference methodologies for a Cordis-native fullstack expert plugin

August 16, 2026 · View on GitHub

Scope. Primary-source review of three locally cloned repositories. Every factual claim below links to an exact Git commit, so citations remain stable.

Sourced facts. Superpowers presents a composable coding-agent methodology: clarify the request, agree on a design, write an implementation plan, execute with subagents or plan execution, use TDD, review, and finish the work. (README, “How it works”; subagent-driven-development skill)

Its skill-writing guidance uses behavioral pressure scenarios: establish a baseline failure, add the smallest guidance that changes behavior, then test again and close loopholes. (writing-skills/SKILL.md)

Cordis recommendation. Use separate, composable runtime capabilities for requirement/design, planning, execution/review, and verification. Test real Host Tool/Service and Client Slot seams—not only prompt text—and ensure stop/update disposes every effect.

mattpocock/skills

Sourced facts. The project describes its skills as small, adaptable, composable, and model-agnostic. It distinguishes user-invoked orchestration from model-invoked reusable discipline. (README; README, installation model)

It emphasizes alignment before implementation, shared project vocabulary, fast feedback, TDD, and deep modules with small interfaces and concentrated implementation. (README, “Why These Skills Exist”; codebase-design/SKILL.md)

Cordis recommendation. Keep explicit commands distinct from low-risk automatic guidance. Expose a few deep operations—inspect context, propose a plan, run a guarded step, verify—while keeping phase machinery private. Store project vocabulary as owned data, never as serialized live Cordis objects.

DietrichGebert/ponytail

Sourced facts. Ponytail's ladder asks, in order: whether work is needed; whether existing code can be reused; whether the standard library or native platform covers it; whether an installed dependency solves it; and only then to write the minimum that works. It says this follows understanding and tracing the real flow, and does not waive validation, error handling, security, accessibility, or data-loss safeguards. (AGENTS.md; skills/ponytail/SKILL.md)

Cordis recommendation. Make minimality a planning and review gate, not a shortest-output rule: check existing Host/Client Services, Builtins, Slots, and platform behavior first. Preserve approval, reversibility, lifecycle ownership, and accessibility.

Cordis-native synthesis (recommendations)

Host owns files, orchestration, model Tools, and durable state; Client owns page state, Slots, themes, and visible progress.
Use explicit commands for high-impact orchestration and automatic events only for low-risk reminders.
Give every meaningful mutation a declared verification check; make cancellation, rollback, approval, and diagnostics normal states.
Keep the public interface small and deep; keep implementation phases private.
Record evidence, inspected contracts, checks, and unresolved assumptions in outputs.

These are recommendations derived from the sources plus Cordis constraints—not claims that any source repository implements or endorses Cordis.

Licenses and MIT obligations

Sourced facts. Each repository includes an MIT license:

The MIT text permits use, copying, modification, publication, distribution, sublicensing, and sale, provided that the copyright notice and permission notice are included in copies or substantial portions. It is provided “as is,” without warranty, and disclaims liability. (Superpowers license text; skills license text; Ponytail license text)

Practical obligation (not legal advice). If this plugin copies source code or substantial expressive text, preserve the relevant copyright and MIT permission notices and document attribution. Check files for additional third-party notices before redistribution; repository-level MIT status does not automatically relicense embedded third-party assets.

Do not copy

The safe boundary is concepts versus expression:

Concepts may be reimplemented. Workflow ideas such as clarify → plan → implement → verify, TDD, YAGNI/native-first checks, composability, deep interfaces, and human approval are general ideas. Re-express them in original wording and implement them through Cordis contracts and lifecycle semantics.
Do not copy implementation text wholesale. Do not lift SKILL.md prose, prompts, examples, checklists, source code, tests, or file layouts into the plugin merely because they are useful. If substantial text or code is copied, apply MIT notices and attribution; prefer an independent implementation.
Do not copy branding or identity. Do not reuse repository names, logos, slogans, distinctive presentation, marketplace identity, or imply endorsement. MIT licensing covers copyright permissions, not trademark rights or endorsement.
Do not copy adapter code or unrelated integration machinery. Do not import multi-harness manifests, lifecycle hooks, installer scripts, telemetry, platform adapters, or configuration designed for Claude/Codex/Cursor/etc. unless independently required, reviewed, and clearly attributed. They are not needed to reproduce the concepts and may create unwanted side effects. (For example, Superpowers documents optional visual-companion telemetry in its README: pinned section.)
Do not copy benchmark claims as guarantees. Ponytail's reported measurements are repository-specific results, not independent validation or a Cordis performance promise. (pinned benchmark section)
Do not copy behavior that conflicts with Cordis. Never bypass approval, access undeclared Services, serialize live runtime objects, or leave effects unowned. Cordis safety and lifecycle rules take precedence.

Bottom line

Borrow Superpowers' compositional workflow discipline, mattpocock/skills' explicit/model-invoked split and deep-interface vocabulary, and Ponytail's safety-preserving YAGNI ladder. Build an original Cordis-native integration with pinned provenance, MIT attribution where applicable, inspectable contracts, reversible effects, and no copied multi-harness bundle.