Improve agent behavior: investigate before speculating

[RockyWearsAHat]

Added action word dictionary to hopefully force copilot into a certain mode of thinking quicker and more efficiently then debating what the user is actually requesting.

Additionally:

• Add investigation-first approach to all prompt templates
• Replace 'minimize questions' with 'investigate immediately' requirement
• Never speculate without verification using available tools
• Add action word analysis system for request categorization
• Enhance completion criteria with root cause fixing & verification
• Add semantic understanding for ambiguous terms

Ideally this will assist in more complete requests that shouldn't require nearly as many follow up prompts. In testing I was able to generate a a simple 3d car driving game in ThreeJS in 6-10 requests.

EDIT: Didn't realize intention was already getting extracted, essentially doing what I was trying to do but better, it would be more productive to edit those files to extract better intentions rather than implementing this whole new "Action Service"

[vs-code-engineering]

📬 CODENOTIFY

The following users are being notified based on files changed in this PR:

@bhavyaus

Matched files:

• src/extension/prompts/node/agent/anthropicPrompts.tsx

@bryanchen-d

Matched files:

• src/extension/prompts/node/agent/anthropicPrompts.tsx

[RockyWearsAHat]

@microsoft-github-policy-service agree

[Copilot]

Pull request overview

This pull request aims to improve agent behavior by adding investigation-first prompting patterns and creating an action word analysis system. The changes add comprehensive guidelines across multiple AI model prompts to encourage root cause investigation, verification testing, and semantic understanding of user requests.

Changes:

• Added investigation and verification guidelines to all agent prompt templates (OpenAI GPT-4, GPT-5, GPT-5.1, Claude 4.5, and default prompts)
• Created ActionWordService and actionWordAnalysis module to categorize user requests by action words (fix, create, modify, etc.)
• Added semantic understanding guidance with specific domain examples to reduce clarifying questions

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
src/extension/prompts/node/agent/openai/hiddenModelBPrompt.tsx	Added 5 new guidelines emphasizing verification testing, root cause fixing, semantic understanding, and immediate investigation over speculation
src/extension/prompts/node/agent/openai/gpt5Prompt.tsx	Added identical 5 new guidelines for GPT-5 model prompts
src/extension/prompts/node/agent/openai/gpt51Prompt.tsx	Added extensive completion criteria, semantic understanding section, and quality standards (27 new lines) for GPT-5.1
src/extension/prompts/node/agent/openai/defaultOpenAIPrompt.tsx	Added investigation and verification guidelines to the keep-going reminder section
src/extension/prompts/node/agent/anthropicPrompts.tsx	Added structured completion_criteria and semantic_understanding_and_quality tags with comprehensive guidelines for Claude models
src/extension/prompt/common/actionWordService.ts	Created new service interface and implementation for analyzing user requests based on action words
src/extension/prompt/common/actionWordAnalysis.ts	Implemented action word detection, categorization, and completion criteria logic with dictionaries of action/subject words
src/extension/extension/vscode/services.ts	Registered the new ActionWordService in the dependency injection container