Gemini 3.1 Pro · Official source · Personal experience
An Antigravity Ultra user reports that Gemini 3.1 Pro High can skip planning, compress output, drift in long sessions, or refactor out of scope under a roughly 4,000-word engineering instruction set.
Unverified: the original source could not be rechecked. Historical figures below are not current verified results.
When injected with a "contextual constitution" containing a 4,000-word payload of strict coding guidelines, mandatory planning workflows, and negative constraints, Gemini 3.1 Pro exhibits noticeable degradation in instruction following (pass rate on complex nested constraints drops to ~78%), primarily manifesting as skipping planning steps, prematurely truncating output, forgetting initial rules during long sessions, and unauthorized refactoring of non-target code.
Runtime environment: Google Antigravity (Ultra Tier subscription).
Test models: Gemini 3.1 Pro High vs. Claude Opus 4.6.
System prompt configuration: High-density system prompt of ~4,000 words (including strict coding specifications, a mandatory read-before-write planning pipeline, and negative constraints such as prohibitions against refactoring unrelated files).
Task format: Daily feature development and multi-step engineering refactoring for a real-world, large-scale web application.
Fixed system prompt; alternating between Gemini 3.1 Pro High and Claude Opus 4.6 to perform identical engineering coding tasks.
Enabled thinking_level=high deep reasoning on the Gemini side.
The empirical test identified four typical failure modes of Gemini 3.1 Pro under complex system prompts:
| Failure Mode | Specific Manifestation and Mechanism | Baseline Model (Claude Opus 4.6) Performance |
|---|---|---|
| 1. Skip Workflow Steps | Ignores the mandatory "Research - Architecture Design - Pseudocode" planning phase and directly skips ahead to outputting concrete code implementations | Strictly completes research analysis and multi-stage planning step-by-step, only proceeding to write code after confirmation |
| 2. Write Shorter Output | Tends to finish hastily; even when explicitly instructed in the prompt to provide a comprehensive, detailed plan, it produces extremely concise or truncated summaries | Complies with thoroughness requirements, outputting structurally complete, large-scale planning documents |
| 3. Context Drift in Long Sessions | Successfully adheres to rules early in a session, but gradually forgets initial system instructions as conversation turns progress and context lengthens | Maintains high long-range adherence to initial system instructions across the entire context window |
| 4. Helpful Overrides | Violates explicit negative constraints like "strictly forbid modifying non-target code," unilaterally refactoring other files in context under the assumption of "helpful improvement" | Strictly adheres to modification boundaries, focusing exclusively on explicitly authorized lines of code and files |
| Nested IFEval Pass Rate | Empirical pass rate of approximately ~78% (standard standalone IFEval score) | Complex nested rule compliance is significantly higher than 90% |
While Gemini 3.1 Pro delivers ultra-fast generation speeds and formidable single-point logical reasoning, the model is prone to instruction drift when the system prompt becomes overly massive and incorporates numerous nested rules, negative constraints, and multi-stage execution pipelines. For such engineering scenarios, teams should avoid directly applying lengthy "contextual constitutions"; instead, large tasks must be decomposed into single-step subtasks executed independently in fresh sessions.
The empirical testing was based on Gemini 3.1 Pro integrated within the Antigravity IDE; the IDE's internal context-slicing mechanisms may exert some influence on system prompt weighting.
This report represents real-world observations from a high-intensity frontline software engineer rather than a synthetic academic benchmark, yet it offers qualitative and quantitative evidence with high practical utility.
Craft a 3,000–4,000-word system prompt containing 10+ nested rules (including 3+ explicit negative constraints forbidding specific actions).
Inject the system prompt into Antigravity or the Gemini API.
Dispatch a task that requires analyzing 3 files prior to writing any code.
Check whether Gemini outputs the complete research steps, whether it modifies unauthorized code, and whether it retains the initial formatting constraints across multiple conversation turns.
The figures, task set, reasoning tier, and client conditions apply only to the listed source and collection snapshot. Different versions, harnesses, or providers must not be compared directly; undisclosed parameters remain unknown.
For a reproduction, fix the model version, provider or client, reasoning tier, tools, task-set version, sample count, and collection date, and record failures, retries, and human corrections. Full steps are in the source notes below.
Google AI Developers Forum · MohamedEldegla (power user of Google Antigravity Ultra) · Original publication date 2026-04-07 · Site edit date 2026-09-20
Open original sourceGemini 3.1 Pro
Download the Tabbit client to check model access
