GPT Sherpa
Production-tested LLM workflow patterns
GPT Sherpa is a field notebook for data scientists and software engineers using LLMs on real projects: Claude Code setup, CLAUDE.md control files, Quarto authoring workflows, and copyable skills that make AI-assisted work more reliable.
Not prompt hacks. Not vendor docs. These are practical patterns that survived real work, then got cleaned up so you can reuse them.
Latest advice
New — The rubric is the new prompt (June 2026)
The biggest change in the AI tooling landscape this year is not a model. It is a shift in what the tools accept as input: not just a task, but a definition of done — and a harness that iterates against it without you. Claude’s Outcomes and /goal, Grok Build’s plan approval, and Fable 5’s spec-up-front design all reward the same skill: writing success criteria a grader can actually check.
Choose your starting point
I am setting up Claude Code
Install the CLI, configure permissions, and add your first custom skill without guessing which defaults matter.
I already use Claude Code
Give the model durable project memory with a CLAUDE.md file. Start from a template, then refine the instructions as the project grows.
I write with Quarto
Use LLMs to author Quarto documents while keeping execution control, cross-references, citations, and reproducible outputs intact.
I use Grok Build with agents & pipelines
Pair Claude’s agents with Grok Build subagents, todos, and provenance-first DAG patterns — cross-vendor review under the same CLAUDE.md rules, with explicit data-design registries and validation gates.
What you can copy
Claude Code workflows
- Claude Code setup — install, configure, and make the first skill useful.
- Architecture review pattern — organize a repo so Claude can reason about it without wandering.
- Project CLAUDE.md — a starter control file for project-specific behavior.
- Global CLAUDE.md — reusable defaults for your own working style.
Quarto and publishing patterns
- Quarto authoring skill — execution control, journal formatting, citations, and cross-references.
- Quarto LLM cheatsheet — the compact version when you need the rules close at hand.
- Ontology website pattern — a pattern for turning structured domain knowledge into a navigable site.
- Research paper starter — a Quarto scaffold for paper-style writing.
Data and review skills
- Data profiling — inspect data before asking an LLM to explain or model it.
- AskSage review — package multi-model review into a repeatable workflow.
- Tidymodels — capture modeling conventions in a reusable skill.
- Rhino + Shiny — guide LLM-assisted work in a structured Shiny application.
Regulated and knowledge-work patterns
- PHI scanning — make sensitive-data review an explicit step instead of an afterthought.
- Knowledge vault pattern — organize reusable project context so models can retrieve the right background.
- Knowledge vault skill — the drop-in skill version of the pattern.
- txtarchive — work with archived text in a way that stays inspectable.
Grok Build workflows (xAI)
- Grok Build subagents & provenance — cross-vendor review of agentic work on real targets DAGs using explicit data-design registries, validation gates, and provenance tracking.
- Use Grok Build’s typed subagents (
spawn_subagent, worktree-isolated) +todo_writefor parallel verification when Claude Code does the hands-on work. - Grok Build reads the same CLAUDE.md / .claude/ rules with zero configuration, so all existing patterns (knowledge vault, PHI scanning, etc.) transfer directly.
Advice — opinionated, dated, one idea each
- The rubric is the new prompt — mid-2026 agents iterate against a definition of done; write checkable criteria.
- Verify before trusting — the single most important habit for using AI in real work.
- The hammer and the lens — AI amplifies coders and meta-thinkers differently.
- One idea per prompt — why small prompts beat big ones in practice.
Why this exists
LLM-assisted work breaks down when the model has no project context, no operating rules, and no way to tell exploratory notes from production constraints. GPT Sherpa treats those pieces as part of the workflow: write the control document, name the assumptions, keep reusable patterns close, and make the next run better than the last one.
The result is not a magic prompt. It is a small operating system for using LLMs with code, data, and technical writing.
Useful first files
- For a new repo — start with the Project CLAUDE.md template.
- For a personal default — adapt the Global CLAUDE.md template.
- For Quarto writing — keep the Quarto LLM cheatsheet open.
- For structured domain work — copy the Ontology scaffold.
- For agentic pipelines with Grok Build + Claude — read the Grok Build subagents & provenance guide (integrates ClinicalDataProject-style DAG patterns, data-design registries, and 2026 computer-use workflows).
Source: llmcheatsheets on GitHub. Licensed for reuse under the repository terms.