Superpowers
obra/superpowersA 15-skill development methodology that makes coding agents clarify the goal, write a plan, work through it with subagents and verify before calling it done.
Recommended
Check first Depth 3/5
Elephant-Goldfish Model (EGM) Claude skills
A spec-first workflow: a full-context agent (Elephant) drafts a design doc, then fresh no-context subagents (Goldfish) stress-test it before you approve.
Our verdict Recommended
A disciplined, Google-maintained spec-and-review workflow with a real mechanism (context-free reviewers) rather than just more prompting; Apache-2.0 and actively maintained.
What Elephant-Goldfish Model (EGM) does
EGM is a Google-published plugin used internally for drafting and reviewing design specs before writing code. The main sequence is egm-brainstorm (interactive clarifying questions plus a Goldfish review, producing concepts_brief.md), egm-draft (turns the brief into design_draft.md section by section, pausing for feedback), and egm-review (fresh reviewers check clarity, fit with the code and the implementation plan). egm-review-loop and egm-resolve support iterating on review feedback, egm-drift-report checks whether an implementation drifted from its approved spec, and egm-bootstrap generates and verifies 'Peanuts and Hay' README.md contract files across a directory tree (leaf-level file contracts rolling up into repository blueprints).
The core mechanism is that Goldfish reviewers run in genuinely fresh subagent contexts with no conversation history, so they catch unstated assumptions the main agent (which holds full context) would not question. Skills explicitly stop if the host cannot spawn real subagents, because a review in the same context is not a Goldfish review. Rules are strict: artifact-first delivery (write to a file, never dump long drafts in chat), plain English with no promotional adjectives, and a 'critique first, no glazing' review style that leads with what's wrong. It supports Claude Code, Codex, Antigravity, Claude Cowork and ChatGPT Work, each with its own subagent-spawning convention documented in docs/harnesses.md.
Good for
- Feature work that needs a reviewed design doc before implementation
- Teams wanting an automatic check for assumptions the main agent can't see
- Keeping component READMEs and interface contracts current across a repo
Not for
- Small, clear bug fixes (the README says use your normal workflow)
- Hosts that cannot spawn fresh subagents, since Goldfish review stops rather than degrading silently
What's inside depth 4/5
14 skills averaging 14 KB with dedicated reviewer persona files, templates and (in the skills/ distribution) one test file each, plus harness-specific integration docs; no broader eval suite.
- Skills
- 14
- Reference docs
- 13
- Scripts
- 0
- Tests / evals
- 7
- Templates
- 14
- Avg SKILL.md
- 13.8 KB
Counted from the repository file tree. How we score depth.
Dependencies and costs
- A coding agent that can spawn fresh, context-free subagents (Claude Code's Agent tool, Codex's spawn_agent, Antigravity's invoke_subagent, or equivalent)
Security review
No issues foundWe scanned all 225 files in this repository and read every script. Nothing in it reaches beyond what the skill describes.
A set of Markdown skills and reviewer personas for drafting specs, reviewing designs with fresh-context subagents, and checking implementation drift. The Claude reviewer agents are limited to Read, Grep and Glob; the bootstrap and synthesizer agents can also write files, and bootstrap also gets Bash for local documentation discovery. The package declares no hooks, MCP servers or network calls, and its privacy doc states there is no telemetry; the Node site and Python scripts are maintainer tooling only.
What we checked
The repository (225 files) on Oct 29, 2026. We had every file and read every script.
NVIDIA SkillSpector static scan: 99 pattern hits (Agent Snooping 66, Excessive Agency 9, Prompt Injection 7, Supply Chain 7, Dangerous Code Execution 5, MCP Rug Pull 4). Most hits are HTML comments in Markdown templates, skill-file path mentions, and 'no approval'/'no verification' phrases in docs. The subprocess hits are the opt-in live integration tests; 'secret.txt' is a packaging test fixture proving such files are not shipped. 'npx prettier' is a mention of a project formatter only if the user's project already configures one.
A review, not a certification. Skills change after we check them — read the files before installing.
License
Apache-2.0 — commercial use allowed; not an officially supported Google product.
Can it run unattended?
Not really. Brainstorming and drafting pause for your feedback at each section, and review findings are something you choose to address, not an autonomous loop.
Alternatives to Elephant-Goldfish Model (EGM)
Compound Engineering
EveryInc/compound-engineering-pluginEvery's plugin of 36 skills for a brainstorm, plan, build, review and capture-the-learning loop, with an optional hands-off /lfg pipeline.
Recommended
Broad access Depth 4/5
pstack (standalone mirror)
backnotprop/pstackA standalone mirror of Cursor's pstack skill stack, rewritten to work in any agent harness: 61 skills for rigorous, verified engineering work.
With caveats
Broad access Depth 2/5
Planning with Files
OthmanAdi/planning-with-filesKeeps task_plan.md, findings.md and progress.md on disk and re-injects them via hooks, so a plan survives /clear, compaction and crashes.
Recommended
Broad access Depth 4/5
FAQ
What is the Elephant-Goldfish Model?
It is an open-source Google plugin for designing software with AI agents: an Elephant agent with full project context drafts a specification, and fresh, context-free Goldfish subagents review it to catch unstated assumptions, before you approve implementation.
How do I install EGM?
In Claude Code run claude plugin marketplace add google/elephant-goldfish-model then claude plugin install egm@egm. Antigravity and Codex have their own plugin install commands, and Claude Cowork and ChatGPT Work have a separate setup guide.
Does EGM replace tests and code review?
No. The README states a Goldfish review does not replace tests or human judgment; it is a design-time check before implementation, not a substitute for later code review or testing.
What happens if my agent can't create subagents?
EGM stops the review rather than running it in the same context, because the README treats a same-context review as not a real Goldfish review. Check docs/harnesses.md for your agent's support status.
Sources: GitHub repository (stars, license, file tree, last push Oct 29, 2026). Review based on the README and SKILL.md files as published.