Elephant/Goldfish workflow
Six commands for Claude Code, Codex and Cursor: the main session holds context while fresh agents review the problem and diff.
- 97 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- By the instructions the agent writes command and skill files into config directories and starts subagents that read the project
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add workflows/elephant-goldfishDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- By the instructions the agent writes command and skill files into config directories and starts subagents that read the project
This is third-party code. Review the repository files before installing.
What it does
Elephant/Goldfish builds a workflow on two roles: the main session remembers all context while fresh agents with no conversation history look at the problem, design or diff and find assumptions the first agent stopped noticing. The idea comes from Dave Rensin's article, recent additions come from the ClaudeFlows project. The commands cover work stages: `eg-question` answers codebase questions with citations and writes nothing, `eg-brainstorm` explores approaches, `eg-prd` turns an idea into requirements with tasks for later sessions, `eg-new-feature` reviews a design before building, `eg-fix-bug` asks for an independent diagnosis and `eg-precommit-review` looks at staged, unstaged and untracked changes separately.
Who it is for. Developers who work with agents on hard changes and accept spending extra time and tokens on an independent view.
Good fit when
- A new feature design should be reviewed by a second pair of eyes before any code is written.
- A bug will not yield and an independent diagnosis is needed before fixing.
- Before committing, review staged and unstaged changes in separate passes.
- A raw idea must become requirements with tasks for later sessions.
Not a fit when
- A one-line fix: extra agents cost time and tokens.
- The host cannot start fresh agents: the workflow reports that limitation.
- You want an auto-commit: reviews report gaps and never commit silently.
Example request
/eg-new-feature: add CSV export for the report and have a fresh agent review the design before implementation.Limitations
The reviews cost extra time and tokens, so they pay off where a second opinion matters more than speed. Each agent needs its own adapter and install: Claude Code puts commands in `.claude/commands/`, Codex in `~/.codex/skills/`, and Gemini CLI, Cursor and Grok Build in their own `skills` folders. The install itself through `BOOTSTRAP.md` needs the GitHub CLI. If the host cannot start fresh agents, the process says so and no independent review happens. After installing, the agent session has to be restarted.
How to disable. Delete the installed `eg-*` command or skill files from your agent directory and the `snippet.md` fragment from the instructions if it was added.
FAQ
Which commands exist?
`eg-question`, `eg-brainstorm`, `eg-prd`, `eg-new-feature`, `eg-fix-bug` and `eg-precommit-review`. You can start with any, a PRD is not needed to fix a bug.
Which agents does it work with?
Claude Code, Codex, Gemini CLI, Cursor and Grok Build, each with its own adapter and install location.
How is it installed?
You hand the agent the `BOOTSTRAP.md` for the right adapter through the GitHub CLI and it installs the files while keeping your customizations.
Does a review change code?
No, `eg-question` is read-only and reviews report gaps and rebuttals without committing your work.
Where does the approach come from?
From Dave Rensin's article on elephants and goldfish in software work, with recent parts from ClaudeFlows.
Discussion
No questions or reviews yet.
Related
WorkflowsClaude Code kit with eight phase agents, 21 skills and hooks that keep the agent inside set limits.
A workflow template for Claude Code and five more agents: a backlog task, a diff review and a commit.
Twelve subagents plus a standard-team playbook that carry a Claude Code project from PRD to release, with QA checks at each step.
A 33-skill plugin with gates: question the request, verify facts against sources, test first, review by a second model
14 skills that find the gaps in a task before the code: interviews, plans, assumption tests and proof that a bug fix works
Three skills for Claude Code, Cursor and OpenCode: an interview-driven PRD, task slices and a design doc before any code
README in short
The README gives a table of six commands, an adapter table per agent and a prompt template for installing through BOOTSTRAP. It states the cost in time and tokens separately.