Managing Claude & ChatGPT: the practical guide
Most people use these tools like a search box and wonder why the output is mediocre. The people who get real work out of them treat them like a junior teammate you brief well and manage tightly. Here's how.
This guide is original writing informed by hands-on use and formal AI coursework. It teaches the how, not any one course's material.
The one mental model that fixes everything
An LLM has no memory of you and no access to your intent beyond the words in its context window. Everything good comes from controlling that context. Vague in, vague out. The skill isn't "knowing the magic words" — it's consistently giving the model the right role, the right constraints, the right examples, and the right information, every time.
Three levers do most of the work:
- Role & standing instructions — who the model should be, every time, without you re-typing it.
- Context — the specific material the task needs (the doc, the data, the examples), and nothing that distracts.
- Output contract — exactly what shape you want back, and how you'll judge it.
Claude
Standing instructions that persist
- Projects (claude.ai) hold a persistent set of instructions + reference files that apply to every chat in the project. Use one project per recurring job (e.g. "Newsletter drafting," "Home Assistant configs"). This is the single biggest upgrade over one-off chats.
- Styles let you save a writing voice and reuse it. Define your voice once.
- In Claude Code, a
CLAUDE.mdfile is the equivalent — durable project memory the model reads every session.
Structure prompts with tags
Claude responds especially well to lightly-structured prompts. Wrapping the parts of your request in simple tags removes ambiguity about what's instruction vs. data:
<role>You are my copy editor.</role>
<task>Tighten the draft below without changing meaning.</task>
<constraints>Keep it under 200 words. Preserve the headings.</constraints>
<draft>
...paste the draft...
</draft>This isn't superstition — separating instruction from content is what stops the model from "answering" text that was meant to be edited.
Work in the right surface
- Artifacts — for anything you'll iterate on (a document, a tool, a page). You get a live, editable result instead of a wall of chat text.
- Claude Code — for building, editing files, and automating real work on your machine. The power-user tier.
- Long context — Claude handles large inputs well; paste the whole doc rather than a lossy summary when the detail matters.
Manage the context window
Long chats drift and pick up cruft. When a conversation gets muddy, start a fresh chat and re-state the goal — cheaper and better than fighting a polluted context. Keep one chat to one task.
ChatGPT
The three features that change everything
- Custom instructions — your standing "about me" + "how to respond." Set these once (what you do, your tools, your preferences, how blunt you want it). Every chat inherits them.
- Projects — group chats with shared files and instructions, like Claude's.
- Memory — it remembers facts across chats. Useful, but audit it — wrong or stale memories quietly degrade answers. Know how to view and prune it.
GPTs and tools
- Custom GPTs are reusable, pre-briefed assistants for a repeated task — worth it only if you actually reuse them.
- Turn on the right tools for the job (web, code/data analysis, image) and turn them off when they add noise.
Prompting that works on both
The durable patterns live in prompting-patterns.md. The short version:
- Give a role and a goal, not just a question.
- Show one example of the output you want (one good example beats three paragraphs of description).
- State the constraints (length, format, audience, what to avoid).
- Ask for the format you'll actually use (table, JSON, bullet list, a doc).
- Iterate deliberately — critique the output and ask for a specific revision, rather than re-rolling and hoping.
Which model, when
A rough, honest heuristic (models change monthly — see the model-watch template):
| You want to… | Reach for |
|---|---|
| Build, edit files, automate real work | Claude Code |
| Long, careful writing / editing | Claude (Projects + Styles) |
| Quick answers with web + data tools | ChatGPT |
| Keep everything private / offline | A local model via Ollama |
| Learn a topic deeply | Either, in "explain like I'm skeptical" mode |
Don't be loyal to one. The pros keep two or three open and route by task.
Running your own (the self-hosted angle)
You don't have to rent every token. For private, offline, or high-volume work, run an open model locally — Ollama is the one-command on-ramp, Open WebUI gives it a ChatGPT-style face, and on Apple Silicon a Mac mini is a shockingly capable, quiet box for it. See the AI & Local LLMs picks.
The strategy lens: don't just use it, deploy it
Prompting well is table stakes. The real question — the one MIT Sloan + CSAIL's Artificial Intelligence: Implications for Business Strategy program is built around — is where AI actually creates value, and where it doesn't. A few principles that separate people who use these tools from people who deploy them:
- Automate the judgment-light, high-volume work first. The wins are the repetitive tasks with clear inputs and checkable outputs (drafting, triage, extraction), not the one-off high-stakes calls.
- Keep a human on the decisions that carry risk. Use the model to prepare and accelerate; keep a person accountable for anything with real consequences.
- Design for where models fail. They're confident when wrong. Build in verification (grounding, source-checking) rather than trusting fluent output.
- The advantage is workflow, not the model. Everyone has access to the same models; the edge comes from how you wire them into a process nobody else has.
This section reflects the author's MIT Sloan + CSAIL certificate in AI: Implications for Business Strategy, written as original guidance.
Part of HomeForge — own your home, your servers, and your AI.
Get the free Starter Kit →