Google Releases Elephant-Goldfish Plugin: Write the Spec Before the Code

From "Write Code Directly" to "Write the Spec First": Google's New Experiment

AI coding agents have improved noticeably over the past while, but their failure modes are becoming clearer too: they write code at breakneck speed, yet veer off course from the very start. Google's newly released Elephant-Goldfish Model plugin tries to invert this — first having the agent write requirements as a spec, then introducing multiple rounds of zero-context review, and only then allowing it to touch the code.

According to MarkTechPost, the plugin's core workflow is: the coding agent first drafts a specification document, then goes through 5 zero-context review stages, and only after passing all of them does it enter the coding phase. "Zero context" means the reviewers do not inherit the drafter's conversation history or reasoning chain — they face only the spec text itself. This design approach is worth unpacking.

Why Specs Need Review More Than Code

Context Contamination Is a Hidden Flaw of Coding Agents

Coding agents typically work within a long session: reading requirements, examining code, reasoning, and writing implementations. The problem is that the same context contains both faulty assumptions and correct judgments, and the agent struggles to distinguish which is fact and which it made up. Once an early misunderstanding takes hold, all subsequent code is built around that deviation — and the more it writes, the more confident it becomes.

Zero-context review cuts this contamination chain. Reviewers see only the spec, free of the psychological inertia of "I've already thought this through," making them more likely to spot gaps at the requirements level.

"Context contamination" borrows from the context contamination / context rot phenomenon in LLM applications: as conversation turns accumulate, the context window mixes in outdated information, faulty assumptions, repetitive self-confirmation, and irrelevant tool return results, causing the model's judgment of the current task to gradually drift. For coding agents, there are three main sources of contamination: early misreading of requirements, one-sided conclusions drawn while exploring the codebase, and intermediate reasoning generated by the model itself being cited as established fact in later steps. Because LLMs tend toward self-consistency bias, once a faulty assumption enters the context, subsequent generation keeps finding justifications for it, forming a self-reinforcing loop. The value of zero-context review lies in letting reviewers start from a clean state and judge solely on the spec text, thereby bypassing this self-reinforcing path.

Multi-Stage Review Targets Different Risk Points

Although the original information does not enumerate the specific responsibilities of the 5 stages, the "spec-first" positioning suggests that reviews typically cover several categories of issues: whether requirements are complete, whether boundary conditions are defined, whether acceptance criteria are testable, and whether there are assumptions that conflict with existing systems. Splitting review into multiple independent stages is equivalent to having different "perspectives" each serve as a gatekeeper, rather than letting a single comprehensive review draw conclusions all at once.

This structure is especially important for agent systems — a single review step is easily swayed by the spec's wording, while multiple independent rounds of review enable cross-validation.

"Zero-context review" is engineering-wise close to a multi-agent review pattern: independent model instances take on the reviewer role, sharing no session state with one another and communicating only through the explicit artifact of the spec document. Its key difference from common self-refine and self-critique methods lies in "independence" — self-review within the same context tends to follow its existing reasoning framework and struggles to break out of prior assumptions, whereas an independent instance amounts to a cold-start check. For such review to be truly effective, it usually needs to be paired with structured output (e.g., scoring by dimension, listing blocking items), an explicit review checklist, and a traceable record of conclusions; otherwise independent review can easily degenerate into perfunctory rubber-stamping.

Spec-driven development is not a new concept — it has long-standing counterparts in requirements engineering and design by contract: first clarify input/output contracts, pre- and post-conditions, and acceptance criteria, then implement the code. AI coding agents have brought this idea back into the spotlight because the marginal cost of agent-generated code is extremely low while the cost of rework and review is relatively fixed, so the economics of moving the quality gate forward to the spec stage improve substantially. The spec here plays the role of a "contract" between the agent and the human: it is both a constraint on code generation and an objective basis for judging whether a task is complete, and it can also serve as a source for deriving test cases.

Practical Impact on Developer Workflows

The plugin makes the specification document a first-class citizen. For teams using coding agents, this means the center of gravity of the workflow shifts earlier: rather than reviewing after the agent has generated a large amount of code, it is better to intercept problems at the spec stage. The cost of code review is linear, whereas the cost of spec review is leverage-based.

Another implicit change is the auditability of agents. Specs are natural language or structured text that humans can read, archive, and review within a team; the agent's internal reasoning process, by contrast, is often invisible. Requiring the agent to produce a spec first amounts to forcing it to externalize its intent, which is valuable for both team collaboration and post-hoc accountability.

Caution Is Warranted

The publicly available information is currently quite limited: the plugin's specific form, the agent frameworks it supports, the decision criteria for each of the 5 review stages, and which model performs the zero-context review are all undisclosed. It is too early to treat this as a mature engineering solution.

A more realistic view is that this is a directional signal — the competitive focus of coding agents is shifting from "generation speed" to "gatekeeping before generation." Spec-first plus multiple rounds of independent review is a solution that aligns with engineering intuition, but its actual effectiveness depends on whether the review steps can be truly independent and where the quality threshold for the spec itself is set.

For teams building coding agent pipelines, this approach can be borrowed directly: have the agent first write clearly what is to be done and how completion is measured, then use reviews unburdened by history to poke holes, and only then greenlight code generation. As for how Google's plugin actually performs, more empirical information is still needed.

One-line Summary

Google's new plugin has coding agents write the spec first, pass 5 rounds of zero-context review, and only then write code.

Paragraph Summary

The article interprets Google's Elephant-Goldfish Model plugin, whose core is to invert the coding agent workflow from "write code directly" to "write the spec first." The agent first drafts a specification document, then goes through 5 zero-context review stages in which reviewers do not inherit the drafter's conversation history and face only the spec text; only after passing all of them does it enter coding. The article notes that context contamination in long sessions causes early faulty assumptions to self-reinforce, and zero-context review cuts this contamination chain; multi-stage review targets different risk points such as requirements completeness, boundary conditions, and acceptance criteria, forming cross-validation. For teams, making the spec a first-class citizen moves the problem-interception point earlier, turns review cost from linear to leverage-based, and improves the auditability of agent behavior. However, the plugin's form, review criteria, and execution model are all undisclosed, so its actual effectiveness remains to be verified.