Skip to content

Open-source skills for designing and reviewing software with AI agents

5.8 KB
Raw Blame

Muse integration

Muse loads EGM through its Claude-compatible plugin path: no Muse-native manifest is needed. The plugin validates with all seven skills and all nine agent adapters supported: seven reviewers, the synthesizer, and the bootstrap orchestrator.

The skills route Codex and Claude sessions to their guides, but not Muse sessions. A Muse session doesn't read this file unless you point it here.

Install and locate the plugin

Install the whole plugin, including its personas and standards, from a local path:

git clone https://github.com/google/elephant-goldfish-model.git
muse plugins install ./elephant-goldfish-model --scope user
muse plugins approve egm

marketplace add on this repository discovers zero plugins: Muse reads the Codex catalog (.agents/plugins/marketplace.json) first and skips its entry, so it never falls through to the Claude catalog. Install from a local path instead, and keep the source free of symlinks (remove scratch directories such as .venv/ first; the installer rejects symlink entries). Approval is required once: skills stay active without review while the plugin is enabled, but the nine agent definitions need approve before activation.

Start a fresh session after installation. The skills surface as plugin:egm:egm-* with user-invocable-only activation; confirm with muse skills list --source plugin. That prints plugin:// URIs, not filesystem paths; get the installed path from the cache= field of muse plugins inspect egm. Do not assume a particular cache directory, version, or checkout location.

From <plugin>/skills/<skill>/SKILL.md, the plugin root is three parent directories above the file. Resolve bundled agents/, resources/, and docs/ paths against that root. Resolve references/ against the active skill's directory. Resolve user documents and generated artifacts against the user's working directory. Never substitute similarly named files in the user's project for bundled personas or standards.

Offline pre-check without installing: muse plugins validate <path> reports the plugin valid with 7 supported skills and 9 supported agents. Warnings are expected (multiple-manifests picks the Claude manifest over the Codex one; Claude-only fields such as presentation metadata, color, and unknown agent frontmatter keys are ignored). Installing or discovering the plugin alone does not establish end-to-end workflow support.

Fresh-context reviewers

Muse imports the nine Claude adapters as supported agent definitions (egm/egm-design, egm/egm-comprehension, and so on), but every agent tool projection is unsupported: Claude tool allowlists do not restrict Muse reviewers, so do not assume reviewers are read-only. Inspect the available native subagent tool schema before dispatching; tool names can be namespaced and vary by Muse version.

Use a named reviewer agent when the session exposes it, and pass absolute paths to the persona, standards, target document, and explicitly selected reading list. Do not pass the parent transcript, private reasoning, or answers the document is supposed to explain. When named agents are unavailable, fall back to a general-purpose subagent that reads the persona file first. Match it by name frontmatter, not filename: egm-design is agents/design_critic.md. On Muse 1.4.14 the named egm/egm-* definitions report trusted_enabled in the plugin inventory but spawning one is rejected (agent_definition_unavailable), so the general-purpose fallback is the working path in practice.

Require the reviewer to read its persona, target, and every supplied standard before returning findings. A missing read or inaccessible file is an incomplete review, not a clean result. Verify native tool results when the host exposes them; a parent-written completion summary is not equivalent evidence.

Wait for actual reviewer completion before synthesizing findings. Propagate failures and timeouts; a launched reviewer is not a completed review. If nesting is unavailable, dispatch reviewers directly from the parent while preserving the same isolation. If the available tool cannot establish a fresh context, stop the Goldfish review and report that limitation.

Fresh conversation context does not isolate the filesystem, configured memory services, or project instructions. Do not retrieve conversation history or memory to supply missing design context. An isolation test should use a unique parent-only marker and check the actual child result, not the parent's claim.

Questions and artifacts

The shared skills' question payloads describe the intended interaction, not a literal Muse API. Use a structured question tool only when it is available and permitted in the current mode, mapping to its actual schema. Otherwise ask one plain-text question and wait for the answer. Preserve the skill's readiness and section approval gates; tool unavailability is not approval.

Link generated artifacts with absolute-path Markdown links such as [design draft](/absolute/project/design_draft.md). For paths containing spaces use [design draft](</absolute/my project/design_draft.md>). Do not use file:// links in Muse chat.

Verification

Run the offline packaging checks with pytest tests/unit. The end-to-end workflow fixture (brainstorm, draft, and uninterrupted review through the installed plugin) passed on Muse 1.4.14 on 2026-10-30 with the general-purpose reviewer fallback described above. The prompt named this guide and it ran headless with approvals simulated up front, so skill routing, interactive approval gates, and chat link rendering are untested; see the harness test record. Installing or discovering the plugin alone does not establish end-to-end workflow support.