Skip to content

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

10.2 KB
Raw Blame

Claude Code integration

Use Claude Code 2.1.271 or later for omitClaudeMd support. Older versions must not be treated as verified Goldfish isolation environments.

Invocation and resources

claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm

Enable the installed plugin with claude plugin enable egm@egm. Start messages with /egm:egm-brainstorm, /egm:egm-draft, or /egm:egm-review. Skills retain disable-model-invocation: true: their absence from automatic skill discovery is expected. The Skill tool refuses these skills, so to run another EGM skill (for example egm-review-loop or egm-resolve during review Stages 7 and 9), read its SKILL.md and follow it in the same thread. Do not use the Skill tool for EGM skills. Keep drafting and orchestration in the main context. Claude Code has no artifact comments. To request draft revisions, edit design_draft.md directly or add <!-- --> comments in it, or give the notes in chat; the model reads the draft again before each revision. Write briefs, drafts, and reports to the user's working directory. The per-session scratchpad is not the artifact directory: it belongs to one session, and the deliverables must outlive it. If the user's prompt or agent instructions file (CLAUDE.md or AGENTS.md) has a line EGM artifact directory: <path>, write them to that folder instead. Resolve a relative path against the working directory. If a path from an instructions file resolves outside the working directory, after following symbolic links, ask the user to confirm it before creating the folder or writing there. If they decline or no one can confirm, use the working directory and say why. Create the folder if it does not exist. In Claude Code, egm:egm-synthesizer may spawn reviewers (nesting is allowed up to 3 levels by default), or the main context may dispatch reviewers directly. In Claude Cowork, subagents cannot spawn further subagents, so dispatch reviewers directly from the main context instead of launching a synthesizer child. In both cases the main context collects findings and writes the report. Keep review stages 5, 6, 7, 8, and 9 sequential. Parallelize only reviewers within the current stage. Wait for every reviewer in that stage, write its report, and complete required remediation before dispatching the next stage. Uninterrupted mode skips human checkpoints, not these stage dependencies.

Claude substitutes ${CLAUDE_PLUGIN_ROOT} in plugin skill and agent bodies. Use that resolved absolute root for agents/, resources/, and docs/; use the invoked skill's directory for its references/. This standalone guide is read as a file, so placeholders here are explanatory: use the root already resolved in the invoked skill. Never guess a cache version or search another installed copy. Review targets and generated artifacts belong to the user's workspace, not the plugin cache. Pass absolute resource and target paths to each reviewer.

Personas and standards live in the plugin cache, outside the project. EGM does not install a permission-approval hook. Approve resource reads through Claude Code's normal permission prompts; current versions can surface background reviewer prompts in the parent session. For headless runs, the user must grant read access to the resolved plugin's agents/, docs/, resources/, and skills/ folders using Claude Code's permission settings. Keep write access limited to the project, the intended output directory, and the session or private scratch directory used for temporary files. A mktemp -d path can't be named in a rule in advance, so a headless run can't use one without a rule for its parent folder. Do not disable permission checks or grant access to every installed plugin. A reviewer that cannot read its persona or standard must report the limitation and stop; it has not reviewed.

For example, after resolving the actual installed plugin directory, add only its resource folders to user-approved read rules in /permissions:

Read(//absolute/plugin/path/agents/**)
Read(//absolute/plugin/path/docs/**)
Read(//absolute/plugin/path/resources/**)
Read(//absolute/plugin/path/skills/**)

On Windows, write the absolute path either with a leading // and a lower-case drive letter without the colon, or as a native path:

  • Read(//c/Users/<you>/.claude/plugins/cache/egm/egm/<version>/agents/**)
  • Read(C:\Users\<you>\.claude\plugins\cache\egm\egm\<version>\agents\**)

The form //C:/Users/... does not match, and the reads are still denied.

A marketplace install lives in a version folder: .../plugins/cache/<marketplace>/egm/<version>/. A rule that names one version stops matching after an update. Replacing only the version segment with * keeps matching across versions and still covers only this plugin's folders. For example, on Windows:

Read(//c/Users/<you>/.claude/plugins/cache/egm/egm/*/agents/**)
Read(//c/Users/<you>/.claude/plugins/cache/egm/egm/*/docs/**)
Read(//c/Users/<you>/.claude/plugins/cache/egm/egm/*/resources/**)
Read(//c/Users/<you>/.claude/plugins/cache/egm/egm/*/skills/**)

Replace the example path; do not copy a guessed cache location. These read rules also cover Grep and Glob on those files on a best-effort basis. Adding a directory with --add-dir extends edit access too, so it is not a read-only substitute.

Claude Code Warning: Default agents/ folder is ignored

In Claude Code (/plugin > Errors), you may see the following advisory warning:

Default agents/ folder is ignored because the manifest sets "agents"

This specific advisory is expected with EGM's explicit adapter list. It does not mean the registered adapters failed to load; investigate other errors separately.

EGM maintains two agent directories:

  1. agents/: Contains canonical, harness-agnostic reviewer persona markdown prompts loaded by Antigravity and other harnesses.
  2. .claude-plugin/agents/: Contains Claude Code adapters. Reviewer adapters use omitClaudeMd: true and tools: [Read, Grep, Glob]; the synthesizer and bootstrap orchestrator have their own tool lists, described below.

Because .claude-plugin/plugin.json explicitly lists the adapters in "agents", Claude Code skips auto-loading the root agents/ folder. The shared personas do not declare Claude's read-only tool restrictions. Auto-loading them instead would lose the adapter-specific restrictions; available tools would still be subject to the host's permissions. Registering both copies would also duplicate reviewer names. Keep the explicit adapter list rather than removing it to silence this notice.

Reviewer isolation

Spawn a new native Agent of the plugin's registered reviewer type for each independent review, for example egm:egm-zoom-out. Do not fork the conversation or resume a previous reviewer. Provide only the task, target, persona/standard paths, and explicitly supplied Goldfish context; do not copy parent deliberation.

Claude adapters in .claude-plugin/agents/ declare tools: [Read, Grep, Glob], model: inherit, and omitClaudeMd: true. Each adapter reads its canonical persona from agents/ through ${CLAUDE_PLUGIN_ROOT}. Keep Claude tool names out of shared persona frontmatter: other hosts can interpret tools themselves. Reviewers cannot edit files or run shell commands, and no persistent memory is configured. The synthesizer adapter declares tools: [Agent, Read, Grep, Glob, Write] so it can spawn reviewers and write reports but cannot run shell commands or use the session's MCP connectors. A plugin subagent's tools field cannot restrict Write to specific paths, and Claude Code ignores hooks and permissionMode in plugin subagents. The adapter's rule to write only the requested report or brief is therefore an instruction, not a host-enforced limit. Writes still follow the session's permission rules. When you spawn egm:egm-synthesizer, name the artifact directory you are using in its task. Otherwise it never sees a line from the user's prompt, and it drops a folder outside the working directory that only agent instructions name. It has no shell, so it checks that path as written and does not follow symbolic links. The bootstrap orchestrator adapter allows only Agent, Bash, Read, Grep, Glob, Write, and Edit for local discovery and documentation. It routes workers to the bootstrap SKILL.md so they return the JSON subtree_index instead of prose. The parent handles fixes. Bash remains subject to normal permission rules and is not a filesystem or network sandbox. Reviewers cannot open URLs. Do not pass documentation URLs in GoldfishContext to Claude reviewers: save the raw page as a file in the workspace and pass that path instead.

Normal subagents exclude parent conversation history. omitClaudeMd additionally excludes user/project/local instruction files, but managed policies still apply. This is context and tool isolation, not filesystem isolation: reviewers can read accessible files. Pass any required review constraints explicitly and keep unrelated history, logs, and memory out of the review task.

For local testing, load a uniquely named copy with claude --plugin-dir <copy> to avoid confusion with an already installed egm plugin. Validate it with claude plugin validate <copy>/.claude-plugin/plugin.json.

Claude Cowork

Claude Cowork uses the same .claude-plugin package as Claude Code. Install the plugin from claude.ai/plugins or the Cowork plugin panel via marketplace google/elephant-goldfish-model. For a candidate package, Customize > Plugins > Add plugin > Upload plugin accepts a ZIP containing the Claude manifest, adapters, and shared runtime files. Install it and choose Try in chat. Record whether the resulting task runs locally or in the cloud; the tested desktop surface used cloud tasks. Verify resource resolution and reviewer tools in the actual Cowork session. Installation alone does not establish that Claude Code's subagent settings or ${CLAUDE_PLUGIN_ROOT} substitution work on that surface. Local testing with claude --plugin-dir is supported in Claude Code only.

References: plugin paths, skill invocation, and subagent context and tools.