AI Guides › Workbench
By Nigel Guy · 8 min read
A "harness" is everything wrapped around the model so it doesn't start from nothing each time: standing instructions, memory, project files and reusable skills. Most people get it wrong in one of two ways. Either they build nothing and re-explain their business at the top of every chat, or they watch someone's 40-file setup video and copy all of it. The copied setup feels thorough. In practice it fills every session with rules written for someone else's work, and the model follows the ones that matter less reliably.
The rule: interview yourself before you install anything. Build only the layers your answers ask for, starting with one page of standing instructions, and add a layer only when the same gap shows up three times.
| Layer | What it does | Cost / free tier at time of writing | Best for | The catch |
|---|---|---|---|---|
| Standing instructions (ChatGPT custom instructions; Claude's instructions and memory) | Applies the same few rules to every chat | Free on all plans. ChatGPT allows 1,500 characters on Free and Go, 5,000 on paid plans | Who you are, how you want answers, hard "ask first" rules | It applies everywhere, so one project's rules leak into everything |
| Memory | The assistant keeps facts it learns from your chats | Claude: on by default on Free, Pro and Max. ChatGPT: varies by plan and region | Small durable facts you don't want to repeat | It fills itself, including with one-off details you'd rather it forgot |
| Projects | A workspace with its own instructions, files and chat history | Claude: free accounts get up to five projects. ChatGPT: free accounts get 5 files per project, Go and Plus 25, Pro and Business 40 | One client, product or recurring piece of work | People make one giant project, which defeats the point |
| Skills | A packaged procedure Claude loads when the task matches | All Claude plans, with code execution switched on | A job you do the same way every week | A vague description means Claude never picks it up |
| CLAUDE.md (Claude Code only) | A file of instructions loaded at the start of every coding session | Needs Claude Code on a paid plan | Repositories and folders of working files | Anthropic advises keeping each file under 200 lines. Past that, it uses more context and gets followed less reliably |
On price: when I checked, Claude's pricing page listed Pro from $17 a month in US dollars (about £13 at time of writing). I couldn't confirm the sterling price for Claude or ChatGPT from the official pages, so check the figure, including VAT, at checkout.
How to use it. In ChatGPT on web or desktop, open Settings > Personalization and fill in Custom Instructions. On iOS and Android it's Settings > Customize ChatGPT. In Claude, the durable facts about you go into memory (next layer). Rules for a single piece of work go into that project's instructions.
Write it last, not first. The interview below produces it for you.
How to use it. In Claude, open Settings > Memory and check that Generate memory from chats is on. The entries under Topics show what Claude currently believes about you. Click into one to edit or delete it, or tell Claude in a chat what to remember or forget. Each Claude project keeps its own separate memory. In ChatGPT, open Settings > Personalization > Memory. There, saved memories (things you asked it to keep) are separate from referencing chat history. Deleting a chat does not delete a memory saved from it.
Beginner trap. Treating memory as the whole harness. It gets the "who I am" facts right and is poor at holding a procedure.
How to use it. In ChatGPT, select New project in the sidebar, name it, then open the project's ••• menu and choose Project settings to add instructions. When you create the project, you can also choose project-only memory, which keeps what it learns inside that project. Shared projects are always project-only. In Claude, create a project, add its instructions and upload reference files as project knowledge.
Make one project per client or per recurring job, not one per topic you find interesting.
How to use it. On Free, Pro or Max, switch on Settings > Capabilities > Code execution and file creation. On Team and Enterprise, an owner does this under Organization settings > Plugins & skills. Your skills then live under Customize > Skills. To add your own, click +, then + Create skill, then Upload a skill, and upload a ZIP.
The quicker route is Anthropic's own skill-creator skill. It is built interview-first: before writing anything, it asks what the skill should do, when it should trigger and what the output looks like, then goes into edge cases and success criteria.
How to use it. In Claude Code, run /init inside a project folder to generate a starting CLAUDE.md. If you set the environment variable CLAUDE_CODE_NEW_INIT=1 first, /init asks follow-up questions and shows you a proposal before it writes any files. Personal rules for every project go in ~/.claude/CLAUDE.md. Run /memory to see and edit what's loaded, including the notes Claude takes for itself (auto memory).
This is the step everyone skips, because installing feels like progress. Paste the prompt below into a fresh chat outside any project. Replace the bracketed items with what you know and leave the rest blank; it will ask.
You are a careful setup consultant. Your job is to work out the smallest AI harness that fixes my real repeat problems. Do not recommend any tool or file until the interview is finished.
About me so far: [YOUR_ROLE_AND_BUSINESS_IN_ONE_LINE]
Which assistant I use: [CLAUDE / CHATGPT / BOTH]
My plan: [FREE / PRO / PLUS / TEAM / OTHER]
Interview me in rounds of no more than three questions. Wait for my answers before moving on.
Round 1: What do I re-explain most often at the start of a chat? Which three tasks do I repeat every week?
Round 2: For each repeated task, what does a good result look like, and where do the inputs live (files, inbox, a website, my head)?
Round 3: What has gone wrong before: wrong tone, invented facts, forgotten context, or something else? What must you never do without asking me?
If an answer is vague, ask one follow-up question rather than assuming.
When the interview is done, give me:
1. A table with columns: Problem | Layer that fixes it (standing instructions, memory, project, skill, CLAUDE.md) | Why this layer and not a bigger one.
2. A draft of my standing instructions, under 1,200 characters, in plain sentences.
3. At most two projects or skills to build now, plus a "not yet" list for everything else.
Before answering, check that every recommendation traces back to something I actually said, that nothing assumes a plan or feature I haven't confirmed, and that the "not yet" list is longer than the "build now" list. Remove anything that fails.
Fill in your role, which assistant you use and your plan. Expect it to ask you around nine questions in all.
Say you're a sole-trader bookkeeper who keeps re-explaining UK spellings, your client list and your preferred tone for chaser emails. A plausible interview result: standing instructions for spelling and tone, a separate project for each of your two largest clients, and one chaser-email skill. Everything else goes on the "not yet" list, including the CLAUDE.md you don't need because you don't use Claude Code.