Practical scaffolding patterns for using OpenAI Codex, Claude Code, LLM coding agents, LREs, LIMs, agentic workflows, and human-in-the-loop AI agent workflows inside VS Code.
This repository is a public knowledge base for prompt engineering, automation prompts, custom AI skills, project memory files like AGENTS.md and CLAUDE.md, verification loops, and repeatable operating methods that make agent-assisted software development more reliable.
Note
This repo is independent community/educational material. It is not affiliated with, endorsed by, or sponsored by OpenAI, Anthropic, Microsoft, GitHub, VS Code, or any other company mentioned here.
- What This Repo Is
- Why This Exists
- Core Ideas
- The Three-Layer Scaffolding Method
- Automation-Based Prompts
- Custom Skills
- VS Code Agent Workflows
- Repository Map
- Getting Started
- Roadmap
- Contributing and Feedback
- Image Note
- Independence Note
VS-Code-LLM-LRE-LIM-GUIDE-HELPER is a practical guide to working with AI agents in a local developer environment, especially VS Code. It focuses on reusable patterns that help humans and coding agents coordinate across specs, prompts, verification steps, project memory, and workspace guardrails.
The working terms used here are:
| Term | Working meaning in this repo |
|---|---|
LLM |
A large language model used for reasoning, writing, planning, or code generation. |
LRE |
A local reasoning environment: the user's project, files, tools, terminal, editor, memory, and constraints. |
LIM |
A local intelligence/memory layer: durable project context, instructions, skills, references, and reusable prompts. |
AI agent |
A tool such as Codex, Claude Code, or another coding assistant that can inspect files, run commands, edit code, and follow project rules. |
These are practical labels, not claims that the acronyms are universal industry standards.
Most AI coding failures are not caused by a lack of model effort. They come from missing context, weak acceptance criteria, unclear tool boundaries, and unverified success claims.
This repo collects patterns for making agent work more inspectable and repeatable:
- encode the human goal before asking an agent to build;
- define verification before declaring success;
- preserve useful context in the environment so every session starts stronger;
- turn one-off prompts into reusable automation prompts or custom skills when they repeat;
- use VS Code as a stable operating surface for humans and agents.
| Idea | Why it matters |
|---|---|
| Human understanding is the source of truth | The agent can help execute and analyze, but the human owns goals, stakes, and context. |
| Prompts should become artifacts | Useful prompts should be refined, versioned, reused, and verified like code-adjacent assets. |
| Verification is part of the design | Tests, review loops, second-model critique, and ground truth should be planned before work starts. |
| Memory files shape behavior | AGENTS.md, CLAUDE.md, and similar files turn repeated guidance into durable workspace context. |
| The environment should compound | Skills, guardrails, repository structure, and prompt libraries should improve future sessions. |
The central method in this repo is:
Spec -> Verifier -> Environment
Also written as Spec → Verifier → Environment.
| Layer | Purpose | Practical artifacts |
|---|---|---|
| Spec | Encode the human's goal, context, constraints, assumptions, decisions, and acceptance criteria. | Project briefs, implementation specs, issue templates, task prompts. |
| Verifier | Define checks before success is claimed. Include tests, review loops, second-model critique, and external ground truth where possible. | Test commands, lint checks, review rubrics, screenshots, diffs, checklists, validation reports. |
| Environment | Make the workflow compound across sessions through structure, memory, skills, guardrails, and repeatable patterns. | AGENTS.md, CLAUDE.md, custom skills, prompt folders, workspace rules, repo maps. |
The long-form method document is here:
Automation prompts are prompts designed to produce repeatable, inspectable work products rather than one-time chat responses. They are useful when the same workflow should run across multiple projects, assistants, or iterations.
Good automation prompts usually include:
- a clear goal and consumer;
- explicit scope and non-goals;
- environment assumptions;
- allowed and forbidden operations;
- output requirements;
- a verification plan;
- a final report format.
Current example:
prompts/VS_Code_Python_Project_Template.md- a reusable prompt for generating a Windows VS Code Python project bootstrap.
Custom skills are reusable task handbooks for AI agents. A skill should teach an agent how to perform a specific kind of work inside a specific environment, including what context to gather, which files to inspect, what commands are safe, and how to verify the result.
A strong skill usually contains:
- when to use the skill;
- what to read first;
- the expected workflow;
- allowed tools and guardrails;
- known failure modes;
- verification criteria;
- final reporting expectations.
Planned material will cover how to design skills for Codex, Claude Code, and other agentic coding workflows without burying the model in broad, unfocused instructions.
VS Code works well as an AI-agent operating surface because it can bring together source files, terminals, tasks, extensions, workspace settings, prompt libraries, and project memory in one inspectable environment.
This repo will focus on patterns such as:
- organizing prompt and skill folders;
- keeping local bootstrap files out of public repos;
- writing
AGENTS.mdandCLAUDE.mdfiles that are concise but enforce useful behavior; - separating public knowledge from local machine setup;
- asking agents for verifiable final reports instead of vague summaries;
- using human review, tests, diffs, and checklists together.
Current public-facing structure:
VS-Code-LLM-LRE-LIM-GUIDE-HELPER/
|-- README.md
|-- .gitignore
|-- docs/
| `-- images/
| `-- README.png # add this image before or after launch
|-- prompts/
| `-- VS_Code_Python_Project_Template.md
`-- three-layer-scaffolding-method-for-LLM-LRE-FIM-integration.md
Local-only scaffolding such as _workspace/, src/, and one-click launcher .cmd files is intentionally ignored by Git for the public launch.
- Read the Three-Layer Scaffolding Method.
- Review the existing automation prompt in
prompts/. - Adapt the Spec, Verifier, and Environment pattern to one real workflow in your own repo.
- Add or refine your own
AGENTS.md/CLAUDE.mdmemory files with concrete rules, not vague reminders. - Treat the first version as a draft: run it, verify it, then improve the environment for next time.
Tip
Start with one recurring task. Turn that task into a spec-backed prompt, add a verifier, then decide whether it deserves a custom skill.
Planned public material includes:
- examples of strong
AGENTS.mdandCLAUDE.mdproject memory files; - custom skill design patterns and review rubrics;
- prompt templates for code review, repo launch prep, bug triage, and documentation work;
- verification loop examples using tests, diffs, screenshots, and second-model critique;
- VS Code workflow patterns for agent-assisted development;
- before/after examples showing how vague prompts become durable automation prompts.
Feedback is welcome through GitHub issues or discussions once the public repo is live. Useful contributions include:
- clearer terminology;
- better examples;
- additional verification patterns;
- prompt templates that are reusable across projects;
- corrections where a pattern overclaims or needs sharper boundaries.
Please keep contributions general-purpose and avoid including private local paths, secrets, proprietary setup details, or vendor branding that implies official affiliation.
The README hero image should be placed at:
docs/images/README.png
An animated PNG/APNG may be used at the same path if it renders properly on GitHub. Until the image exists, the README is already wired to reference it.
Product and project names such as OpenAI Codex, Claude Code, VS Code, GitHub, Microsoft, OpenAI, and Anthropic are mentioned only to describe practical developer workflows. This repository is independent educational material and does not claim endorsement, partnership, or official status.