A thinking partner is a medium, not a source

For the last couple of weeks I've been using Claude as a thinking partner. At the low end that means "help me get started on this task". At the high end, "I can't make a call on this big life decision".

The aim here isn't to get advice, but the opposite. I want a partner that listens well, helps me explore the situation, lets me express my thoughts, and gets me to clarity. It should be more medium than content. The value is in what it pulls out of me, not in the text it writes.

You can already use stock Claude for that, but it has many harmful behaviours out of the box. Not just annoying ticks like saying "load-bearing" all the time. Real blockers to getting what I want in this mode:

AI jumps ahead, where a good coach takes their time.
AI starts offering diagnoses and solutions, instead of helping figure out the problem.
AI sends 5-paragraph messages which are not at all like a real conversation, even a written one.
AI takes 5 minutes to think, so you multitask and lose focus on the topic at hand.
AI hands you its conclusion, so you stop doing the thinking yourself.

So I hand-wrote 800 words of instructions to fix these. The following examples come from one conversation where Claude follows these instructions while helping me get unstuck on a side project.

(1) Asking for a time budget before diving in:

Claude: That sounds like an uncomfortable place to be — putting in effort but not being able to tell if it's adding up to anything.

Happy to think this through with you. Before we start: roughly how much time do you want to give this — a quick 10 minutes, or something longer?

Taivo: 10

Claude: Okay, 10 minutes — let's keep it focused.

When you say "hard to tell whether I'm making progress" — progress toward what? What would this project look like if it were going well?

(2) Narrowing the problem instead of diagnosing it:

Claude: That might be the whole thing, no? The feeling of "lost" — you've been navigating by the codebase, not by the thing you actually care about. Evals, prompt caching, step caps... those give you signals, but they're proxy signals.

Does "lost" feel more like I don't know where I am in the code — or I don't know if any of this matters?

Taivo: 2nd defs

Claude: Then the codebase was never the problem — you know your way around it fine.

So sitting with "I don't know if any of this matters": what would actually tell you? Not a metric from the eval suite — what's the real-world moment where you'd go "okay, this matters"?

(2) Even when I ask it directly for the answer:

Taivo: i dont have concrete use cases

Claude: That's a striking thing to say out loud after months of building — you've made the how quite sophisticated, but the who asks what, and why is still blank.

Do you think concrete use cases don't exist — or that you haven't gone looking for them?

Taivo: i think they do but i ahvent made them concrete. help me do that

Claude: I won't hand you a list — they'd be my use cases, not real ones. But here's an exercise that fits in the time we have:

Think of three actual people you know — with names — who are non-technical Estonians. Your parent, a neighbor, a friend. For each one: what's a moment in the last year where they dealt with the state, a rule, a deadline, a right they had?

Give me the first person that comes to mind.

Here, stock Claude would have generated ten plausible use cases. But with the coach'y approach, ninety seconds later I had four use cases on the page, and they were mine. They had been in my head the whole time.

(5) Leaving the conclusion to me:

Claude: We're about at time. What's the one thing you're taking from this — what will you do differently next session on this project?

Taivo: i think anchoring the whole thing in concrete use cases is important 🙂 and showing it to other people

Claude: That's a solid takeaway — and notice both halves are about contact with reality, not more code. [...]

For what it's worth, the "lost" feeling made sense in hindsight — you can't feel progress when there's no destination to measure against. Now there are four.

Good session. Go show it to someone 🙂

Those four user stories became the seed of an eval set. Instead of going back to the machinery, I spent the following hours defining what the side project should be able to do for a user. The conversation caught an old pitfall of mine, building in isolation without a user in mind, which is fine while hacking and much less fine when you want to ship.

Before writing those instructions, I tried building a voice agent to fix these issues. But after conversing with it for about 3 minutes, I realized the problems were about what was being said, not the modality. So I stayed in text, sometimes dictating my messages for speed and convenience.

Specifically:

(1)-(3) were fixed by prompting. The sheet tells it to ask for a time budget, to say one thing per turn, and to "be a medium" rather than do my thinking for me.
(4) was improved by picking the right model and level of reasoning. I settled on Opus 5 with Low reasoning effort, which follows the instructions better than Fable 5 and doesn't spend much on thinking tokens. A ~25-minute conversation cost only 64k tokens, so even an expensive model is not a problem here.
(5) got better when I fixed (2) and made a decision to think critically and own the outcome.

Beyond using it myself, I've shared it with colleagues at Pactum. The first day's feedback was "really positive & it's helpful" and "asks helpful questions... doesn't try to solve the world's problems every time... token burn is significantly lower... I'm getting a better thought partner that actually challenges me rather than makes every 💩 idea I have sound like the greatest thing ever... a huge win!".

The instruction sheet is on GitHub -- just drop it into Claude or Codex and ask it to save it as a skill for you.

The coach has significant limitations. It starts every session from zero, because it has no memory of what you've said. It can't notice you circling the same question all month, and it can't hold you accountable, which I think is structural rather than a missing feature.

Even though a human coach would provide a much better experience, the coaching skill still does a job that matters to me: helping me think through something, when the alternative is no help at all.