Shadow Code — Summary

Shadow Code is a VS Code extension (also available for Cursor) that introduces "Shadow Coding" — a novel AI coding technique where developers write pseudocode in a .shadow file that is transformed into target-language production code by a language model. Rather than prompting in natural English ("add a login button"), developers write pseudo-syntax that is structurally closer to code ("implement LoginButton with OAuth2 provider, error states, redirect to /dashboard on success"), reducing the semantic distance between input and output and producing more deterministic results than vibe coding. The extension opens a split-view "Shadow Mode" panel when editing any code file; the .shadow file is the user's pseudocode scratchpad; the original code file receives the AI-generated output. The context() function in shadow files enables selective context injection (only the files relevant to the task), preventing hallucination from irrelevant codebase noise. At version 0.7.4 with 79 stars (GPL-3.0), Shadow Code is the only VS Code extension in this batch and the only framework based on pseudocode as input rather than natural language or structured templates. Compared to seeds, it maps to no seed directly — it is a distinct input modality (pseudocode → code) rather than a workflow orchestration framework.

Shadow Code — Overview

Origin

Created by adifyr (adifyr/shadow-code). GPL-3.0 license. Version 0.7.4. Pushed 2026-03-12. Publisher: adifyr. TypeScript VS Code extension.

Philosophy

From README:

"Developers think in code, not paragraphs."

Shadow Coding is built on the premise that developers communicate naturally in code-like structures, not English prose. When developers describe what they want to build, they naturally think in terms of types, function signatures, control flow, and data structures — not paragraphs. By letting developers write pseudo-syntax close to the intended output, the AI receives an input that is structurally similar to what it must produce, reducing ambiguity and hallucination.

From README:

"With Shadow Coding, developers can control how predictable they want their AI-generated code to be based on the detail & expressiveness of their pseudocode." "Shadow Coding puts developers back in the driver's seat."

Key design decisions

1. Pseudocode as a language extension: Shadow Code enables developers to define custom pseudo-syntaxes per language (e.g., firestore class Payment { ... } for Dart ORM generation). The system prompt explains the custom syntax; the AI applies it consistently.

2. Selective context injection: The context() function at the top of a shadow file specifies exactly which source files the AI sees. This prevents the model from guessing based on unrelated files and produces sharper output.

3. Dependency auto-install: When the generated code uses undeclared dependencies, Shadow Code automatically installs them.

4. Language skills: Users can create .shadows/.skills/<LANGUAGE>.md files to teach the AI project-specific coding patterns for any language, effectively extending the language's syntax for their project.

5. Package-file auto-injection: Shadow Code automatically includes pubspec.yaml (Dart) or package.json (JS/TS) so users don't need to list project dependencies in context().

Manifesto-style quotes

"Shadow Coding produces consistently high-quality code, that is closer to what the developer intends, faster and with less tokens, using cheaper/open-source models."

Breaking change in v0.7.2

import() function renamed to context() to avoid confusion with native language import statements.

Shadow Code — Architecture

Distribution

VS Code Extension (also Cursor)

• Publisher: adifyr
• Extension ID: adifyr.shadow-code-ai
• Version: 0.7.4
• Install: VS Code Extensions Marketplace / Cursor Extensions Marketplace

Required runtime

• VS Code >= 1.108.0 (or Cursor equivalent)
• GitHub Copilot or other VS Code Language Models API provider
• Target language runtime (Node.js for JS/TS, Dart SDK for Flutter, etc.)

Directory tree (extension source)

adifyr/shadow-code/
├── src/
│   ├── extension.ts         # Extension entry point
│   ├── commands/
│   │   ├── shadow_mode.ts   # Open Shadow Mode (split view)
│   │   ├── convert_shadow_code.ts  # Trigger AI transformation
│   │   ├── select_model.ts  # Choose LLM
│   │   └── copy_code.ts     # Copy generated code
│   ├── providers/           # Language provider implementations
│   ├── services/            # AI service integration (VS Code Language Models API)
│   └── utils/
├── syntaxes/               # Shadow file syntax highlighting
├── language-configuration.json  # .shadow file language config
└── esbuild.js              # Build script

User-facing directory structure (in project)

<project>/
└── .shadows/               # All shadow files
    ├── src/
    │   └── components/
    │       └── Login.shadow  # Mirrors code file location
    └── .skills/
        ├── DART.md          # Language skills
        ├── PYTHON.md
        └── GO.md

Shadow file example

context("src/components/chat_message.tsx", "assets/styles/message.css");

// The rest of your pseudocode here...
component LoginForm with {
    fields: email, password
    error_state: display inline below failed field
    on_submit: validate → POST /api/auth → redirect /dashboard
    loading_state: disable submit button, show spinner
}

Language support

Modular. All languages use the same system prompt template; Shadow Code injects language name and config files. Basic support for all popular languages. Extended via .shadows/.skills/<LANGUAGE>.md.

AI integration

VS Code's native Language Models API (requires GitHub Copilot or compatible provider). Shadow Code does not make direct API calls — it delegates to the IDE's LLM infrastructure.

Shadow Code — Components

VS Code Commands (4)

File Types

Context Injection

context(<file1>, <file2>, ...) — special syntax at top of shadow file specifying which source files the AI should read. Auto-injected: pubspec.yaml (Dart) or package.json (JS/TS).

Language Skills

User-created markdown files in .shadows/.skills/ that define custom pseudo-syntax for specific languages:

Example (Dart ORM skill):

firestore class Payment {
    id?: string,
    payer_id: string,
    method: enum(cash, card, online),
    amount: double,
}

→ AI generates complete Dart class with Firestore ORM boilerplate.

VS Code Contribution Points

No workflow framework primitives

No CLAUDE.md, no skills, no hooks, no commands outside VS Code extension system. Shadow Code is a VS Code extension, not an AI coding framework in the traditional sense.

Shadow Code — Uniqueness

Differs from Seeds

Shadow Code maps to no seed directly. All 11 seeds are workflow orchestration frameworks (skills, commands, agents, hooks, task graphs). Shadow Code is a VS Code extension that changes the input modality for AI coding — pseudocode instead of natural language. The closest seed in spirit is kiro (IDE-integrated, changes how developers interact with AI coding) but kiro is a full coding environment whereas Shadow Code is a thin transformation layer that works within existing VS Code setups. Shadow Code is the only framework in this batch or the seed set that treats pseudocode as a first-class input format with its own file extension, syntax highlighting, and custom skill definitions.

Positioning

Shadow Code fills a niche none of the other 10 frameworks in this batch address: the input quality problem. Other frameworks assume the developer can describe what they want in natural language and focus on how to process that description into code. Shadow Code questions this assumption — developers think in code-like structures and should express requirements in code-like structures. The result is more deterministic output with fewer iterations.

Notable: Language extension pattern

The ability to define custom pseudo-syntax per language (.shadows/.skills/DART.md) effectively lets developers create new domain-specific language extensions via natural language skill files. The Dart Firestore ORM example demonstrates that repeating a boilerplate pattern once in a skill file produces consistent output on all future uses — a form of team-level prompt engineering persisted as a project artifact.

Observable Failure Modes

1. Requires GitHub Copilot (or equivalent): VS Code Language Models API requires an active model provider. Free users without Copilot cannot use Shadow Code.
2. Response filtering: README warns about Failed To Stream AI Response: Response got filtered for code matching public patterns. Requires GitHub's "Suggestions matching public code: Allowed" setting.
3. GPL-3.0 license: Copyleft license limits commercial embedding or modification without sharing source.
4. No AI coding workflow: Shadow Code produces code but does not manage tasks, specs, PRs, or testing. Users need a separate framework for SDLC management.
5. Cursor-only outside VS Code: The extension works in VS Code and Cursor only, not Claude Code, Gemini CLI, or other AI coding tools.
6. Manual context declaration: Developers must manually maintain context() calls in shadow files as the codebase evolves.