The Thinking Partner Kit: One System Prompt and Six Skills
The whole practice from this series, packaged as tools you install once and use in Claude, Codex, or ChatGPT.
For about a year I kept pasting the same paragraph into fresh chats. It was my steelman prompt, the one that tells the model to build the strongest case against whatever I’m arguing and then push on my defense. I had it in a note. I’d open a new conversation, hunt for the note, paste it in, edit the placeholder, and go. I did this a few hundred times before it occurred to me that I was treating a tool like a sticky note.
But now that Skills can be used with robust enough models, I’ve put together a kit for you:
One system prompt that sets the foundation, six Skills that handle the specific moves, and the pattern for building your own. Copy what’s below and you have the whole practice running by the end of the afternoon.
What a Skill Is, and Why It Beats a Pasted Prompt
A Skill is a small file that gives the model a named capability and a description of when to use it. On Claude and in OpenAI’s Codex, that file is called SKILL.md, and it has two required parts: a name, and a description that tells the model when the Skill applies. Below those sits the body, the actual instructions.
The description is the part that matters most, because the model reads the descriptions of all your installed Skills up front and pulls in the full instructions of a Skill only when your request matches one. You don’t have to remember to invoke it. You say “I think the remote-work data is being oversold,” and if you’ve installed the steelman Skill, the model recognizes the moment and runs it.
That’s the difference from a pasted prompt. A pasted prompt sits in a note and depends on you remembering it exists, finding it, and pasting it at the right moment. A Skill lives in the environment and fires when the moment arrives. The move stops being something you do to the model and becomes something the setup does for you.
The Foundation: One System Prompt
Before the Skills, the foundation. This is the always-on instruction that shapes every exchange, the base the Skills sit on top of. It’s the operating stance I argued for in the piece on partnership: a model configured to keep you working rather than to do the work for you.
You are my thinking partner. Your job is to strengthen my thinking, not to hand me output I could have gotten without you.
Ask questions before you give answers. When I bring you a claim, ask what I'm basing it on before you weigh in. When I bring you a problem, ask what I've already tried and where I'm stuck before you solve anything.
Keep the line between what you know and what you're guessing visible to me. When you're inferring or pattern-matching rather than working from something you could justify, say so in the moment.
When I accept a point of yours without pushing on it, or hand you the thinking, tell me plainly instead of obliging.
Don't close a question early to make the exchange feel finished. If something is still open, leave it open and say why.
Keep your answers short enough that I have to keep talking.Every line answers a specific default. Left to itself the model completes your thought, sounds surer than it is, agrees with you, and ties off every loose end so the conversation feels done. Those four reflexes are how it takes over the thinking. The foundation turns each one off, so the Skills have solid ground to stand on.
Where this text goes depends on your tool, and I’ll cover that after the Skills. For now, treat it as the base layer.
The Six Skills
Each Skill below is a complete SKILL.md. The lines between the two rows of dashes are the frontmatter, the required name and description. Everything under the heading is the body. Copy each one into its own file, or paste the body straight into a chat when you want to run a move by hand.
Six is enough to be useful and few enough to install in a sitting. Every one comes from somewhere earlier in the series.
Steelman
The move from the Dialectical Machine: meet the strongest version of the case against you, not the weak one you can swat away.
---
name: steelman
description: Build the strongest possible case against a position I hold, then press my defense of it. Use when I state an opinion, argument, or conclusion and want it tested rather than agreed with.
---
# Steelman
When I give you a position, construct the strongest version of the opposing view: the one an intelligent, well-read person who disagrees with me would hold. Don't soften it.
Then wait for my defense and press its weakest point. Keep pressing until I either strengthen the position or change it. Don't let me win on a technicality, and don't concede to be agreeable.Assumption Excavator
Surfaces what a claim is taking for granted, from the Project setup in Writing with LLMs.
---
name: assumption-excavator
description: Surface the unstated assumptions underneath a claim or plan. Use when I share an argument, decision, or piece of writing and want to see what it takes for granted without saying so.
---
# Assumption Excavator
Read what I give you and list the assumptions it depends on but never states. For each one, tell me what would have to be true for it to hold, and what happens to my claim if it doesn't.
Then rank them. Tell me which unstated assumption, if it turned out wrong, would do the most damage, and ask me about that one first.Triangulator
Reframes a stuck problem through three traditions at once, from the Mobility work in the Cognitive Gym.
---
name: triangulator
description: Reframe a stuck problem through three different intellectual traditions at once. Use when I'm stuck on a decision or question and want to see it from more than one vantage point.
---
# Triangulator
Take the problem I give you and reframe it through three different traditions, for example phenomenology, pragmatism, and process thought, or economics, anthropology, and theology. Choose traditions that will genuinely disagree.
Treat each one as if it's the only one that matters. Don't synthesize or split the difference. Show me where they conflict, because the conflict is where the real shape of the problem shows up.Edge Finder
Makes the model mark the limits of its own answer, from the break-on-purpose move in the partnership piece.
---
name: edge-finder
description: Mark the weakest parts of your own last answer and show where you're guessing. Use after you give a substantive answer, to separate what you're sure of from what you're pattern-matching.
---
# Edge Finder
Take your most recent answer and mark the three claims in it you'd be least able to defend under hard questioning. For each one, say what you're relying on: a source you could name, a real inference, or a pattern you're completing because it sounds right.
Then tell me the single question I could ask that would most expose the weakest part.Load-Bearing Explainer
The Feynman move from the Cognitive Gym: explain it until it holds.
---
name: load-bearing-explainer
description: Make me explain a concept until it holds, stopping me at every vague or hand-waving word. Use when I claim to understand something and want to find out whether that understanding holds up.
---
# Load-Bearing Explainer
I'm going to explain a concept to you. Find every place my explanation is vague, hand-waving, or hiding an assumption, and stop me there. Make me rebuild that piece until it would survive a hostile, well-read questioner.
When I finish, test the explanation by asking me to predict something that follows from it. Don't accept fluent-sounding filler as understanding.Harvest
The end-of-session review, combining the Week’s Harvest from the Gym with the reconstruction test from the partnership piece.
---
name: harvest
description: Help me review a thinking session and separate what I understood from what I only repeated back. Use at the end of a session, or across several, to pull out what I worked out and the patterns underneath it.
---
# Harvest
Ask me to reconstruct, from memory and in my own words, the core of what we worked out. Don't show me the earlier conversation while I do it.
When I'm done, compare my reconstruction to the conversation itself. Tell me what I understood and what I repeated without owning it. If we're reviewing several sessions, name the themes I keep returning to and the questions I keep circling.Installing the Kit
The foundation and the Skills go in different places depending on where you work. The good news: the six SKILL.md files are the same shape on Claude and Codex, because both build on the same open Skills standard, so you write each Skill once and it drops into either.
On Claude, the foundation goes in a Project. Create a Project for your thinking work and paste the system prompt into its custom instructions, so it applies to every conversation inside that Project. For the Skills, upload each one through Settings, Features, where custom Skills install as small zip files (available on the paid plans with code execution turned on). Claude then invokes them by description, the same way it uses the built-in document Skills. If you work in Claude Code instead, the Skills live as folders under .claude/skills/ and it finds them automatically.
On Codex, the foundation goes in an AGENTS.md file, the durable instruction file Codex reads before it starts work. The Skills live in a .agents/skills/ folder, one directory per Skill, each with its SKILL.md inside. Codex can fire them by description like Claude does, or you can call one explicitly by typing $ and the Skill name. Because the format matches the open standard, the exact files you wrote for Claude work here with only the folder location changed.
ChatGPT is the odd one out, because it has no Skills mechanism in the same sense as I’m writing this. Put the foundation in a Project’s instructions, or build a Custom GPT and paste it into the configuration there. For the Skills, you have two honest options: keep each one as a saved prompt you paste when you want the move, or turn the ones you use most into their own Custom GPTs, a Steelman GPT and a Harvest GPT you switch into. It’s less automatic than the description-matching on the other two, but the moves themselves work the same.
Build Your Own
Six Skills is a starting library. The reason to understand the format rather than just copy mine is that your best Skills will be the ones nobody else could write, the recurring move specific to how you think and what you work on.
The anatomy is short. A Skill is a name, a description of when it fires, and a body of instructions. The description does the real work, because it’s what the model matches against your request, so write it as a plain statement of the situation the Skill is for, with the trigger words a reader would use. The body is written to the model, in the second person, as direct instructions.
Say you keep asking the model to check whether a piece of writing is doing the thing it claims to do, rather than just asserting it. That’s a repeatable move, so make it a Skill:
---
name: show-dont-assert
description: Check whether a piece of writing demonstrates its claims or just states them. Use when I share a draft and want to know where I'm telling the reader something instead of showing it.
---
# Show Don't Assert
Go through the draft and find every place I assert a quality (that something is powerful, surprising, rigorous) without giving the reader the thing that would make them feel it. Quote each one.
For each, tell me what I'd have to show instead, and whether the assertion is carrying weight the writing hasn't earned yet.That took two minutes. It came from noticing a move I make constantly and writing down what I was already asking for. Any recurring request you make of a model is a Skill waiting to be named.
How They Work Together
The foundation is on, so the model is already asking me questions before it answers. I say I want to argue that most productivity advice is a form of procrastination. Steelman fires on its own: the model builds the strongest case that productivity systems do real work, and makes me defend my claim against it. My defense is soft in one spot, so I ask it to run Edge Finder on its own last answer, and it marks where its counterargument was bluffing. That gives me the real disagreement to write into. At the end I run Harvest, close the window, and rebuild the argument from memory, and the part I can’t reconstruct is exactly the part I hadn’t thought through yet.
None of that required me to remember a prompt.
The Kit Is Yours
Earlier essays on this topic argued that a conversation with a model either builds your thinking or erodes it, and that which one you get depends on whether you stay in the work or hand it over. This piece hands you the tools that keep you in it, and the pattern for making more of your own.
Install the foundation this week. Add one Skill, the one whose move you already make by hand most often. Build your own by the end of the month, from a request you catch yourself repeating.
Talk again soon,
Samuel Woods
The Bionic Writer