Skip to content
Bernhard GötzendorferAI automation that runs on its own · Vienna
MeasuredBehind the Scenes

Deeper, not bigger: a prompt in my planning skill, measured

Felix Rieseberg's depth prompt has been in my planning skill since 19 October. 43 past surprises: old flow 0 hits, the panel 11. Plus what it cannot do.

TL;DR

Felix Rieseberg works on Claude at Anthropic. In mid-October he published a prompt: four questions, then the model rewrites the task to be deeper, not bigger. I did not wire it up as a slash command. I put it into the planning step of my Session Orchestrator. Before that, I measured it: three old plans, 43 things that only showed up later. The old flow saw 0 of them coming, the new brief 7. Together with a brief that tries to refute the plan, it was 11. What the prompt cannot do is further down. And to be honest: for a week the built-in step did not run at all, and I kicked off the first real run by hand.

The prompt, and what Hacker News made of it

On 13 October Felix Rieseberg published How to be more ambitious: A tutorial for humans with models. He separates four kinds of ambition. Speed is the obvious one: done sooner. Altitude is the goal behind the goal. Depth means working like someone who leads their field, with a judge from outside. Range borrows methods from other fields.

For depth he gives a prompt. Here it is verbatim:

"I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I'm not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that."

The core is the fourth question. More on that in a moment.

On Hacker News the post got 30 points and 6 comments (as of 26 October). I read the objections as fair questions. Doesn't this read like an ad for Claude? Where is the evidence that ambitious prompts work better? And one commenter said, roughly: the more ambitious he gets, the less Claude follows the plan. I think the last one matters most. That is why there is a bypass further down.

Why I did not build it as a slash command

On GitHub I found seven other skills that wrap the same prompt. Five of them simply put it behind a command: sharpen in pstack, deeper by sammcj, be-ambitious by DannyMac180, better in openagents and better-prompt by davidondrej. Two go further and add an exception for small tasks and a check: ambitious-model-use by sammcj and ambition-calibrator, which comes from the author of a Reddit post. All of that works. None of it measures anything.

In my setup the prompt sits where a task is born. Session Orchestrator is my open-source plugin for Claude Code. Since 19 October its planning step /plan sends two briefs over a finished plan in Phase 4.5. Brief A is called "Refute" and tries to break the plan. Brief B is called "Deepen" and is Felix's prompt, applied to the plan.

Two briefs instead of one, because they find different kinds of mistakes. Refute mostly hit security and structure. Deepen hit missing oracles, meaning places with no outside check: a "measurement failed" that would have passed as "fixed", or a lifecycle that never ran.

The measurement: three old plans against their own future

Before I merge anything into the planning step, I want to know whether it holds up there. The recipe is easy to copy. I take three old PRDs, product plans, and clone their repos with truncated history. The clone has no commit after the PRD. Then each one is scored: the PRD as it originally shipped, and the same PRD with each brief.

The right answers come from the plans' own future. From the later git history and the issues I collected 43 surprises, things that only showed up after the plan was written. A blind scorer, itself a Claude model, read every version of the plan against that list, and which version was called X and which Y was assigned at random.

VersionHits out of 43
Original PRD flow0
Refute8
Deepen7
Refute and Deepen together11

11 out of 43 is not a clean sweep. But the starting point was 0, and the 11 are things the old flow did not see coming. It did get 15 half right.

Two findings came along on the side. In 3 of 3 PRDs, the real acceptance had never happened. And 0 of 9 PRDs in the whole corpus had a section on success or KPIs.

Three plans are not many. If you tune on three plans, you easily end up fitting exactly those three. So I held back a fourth plan and only checked it at the end. It had 15 surprises: Refute caught 7, Deepen 3, both together 9.

One thing no brief caught. The fourth plan blew its time box by more than four times. That is hard to predict, so I did not solve it as a prediction but as structure. The plan template now has a line "First Real Use / Kill": when the first real use happens and at what point I stop.

The sentence that carries it: what, other than you or me, shows it is right

The fourth question is the cheapest line in the whole prompt. Since 19 October the planning step requires an Oracle: line for every acceptance criterion. It says what, other than the author, shows the criterion is met. The author's own unit test against a fake does not count, because it only checks what the author already believes.

On top of that there are two separate states, "implemented" and "accepted". The three PRDs from the back-test were implemented and never accepted. If acceptance cannot run inside the session, it becomes its own issue instead of quietly going nowhere.

One example from the held-back plan: the process ID in a registry belonged to a hook process that was already dead at that point. A test against a fake would never have shown it. Only the Deepen brief saw it, because it asked how you would tell from outside that the entry is correct.

If you have read my post on loop engineering, this is the same question, asked at the start instead of the end. And My orchestrator now only counts what actually ran shows what happens when a test never starts and still reports green.

What the prompt cannot do

It widens tasks if you let it. The Hacker News objection hits exactly that. So my rule is: deeper means the same scope. For single tasks there has also been /deepen since 26 October, the short form of the Deepen brief. If the task contains a ceiling word like "only" or "quick", or it is a one-step edit, /deepen rewrites nothing. I took that idea from the ambition-calibrator, which lets small tasks through and asks for confirmation when a rewrite changes scope or cost.

It is not a judge either. Question 4 asks for an outside judge; it does not replace one. Where none exists, the result is labelled "unverified", not "looks good".

And only part of it is measured. The back-test shows that the planning step foresees more of the later surprises. Whether deeper tasks end up producing better software, I have not measured. That would be the answer to the second Hacker News question, and I owe it.

Then the uncomfortable part. On 26 October I checked the log of skill invocations. Since the merge there had been 0 calls to /plan. The step was built, but not switched on, which is exactly the pattern from Shipped, but never switched on.

The same day it ran for real for the first time. It checked the plan for a new orchestrator feature, before a line of it was built. Each brief came back with 20 predictions. Only what could be checked in the code was adopted: 11 from Refute, 10 from Deepen, 14 distinct ones together. Three assumptions in the issue itself did not hold. One bug only Deepen found: after an idle break, a new session would have taken the old session's goal as its own.

These numbers do not compare to the back-test. There, a real future did the scoring; here it was only a check against the code. And the run was started by hand, not through /plan. On 29 October the log still shows 0 calls. The step is switched on only once it runs without me.

What I take from it

  1. A prompt is a tool, not a mantra. Back-test first, then adopt.
  2. Two different lines of questioning beat one that is twice as long. Refute and Deepen find different things.
  3. The question about an outside judge costs one line in the plan and separates "implemented" from "accepted".
  4. Deeper means the same scope. Anything that makes the task bigger becomes its own issue.

Where Session Orchestrator sits in my work is described under tooling. If you try the back-test on your own plans, write to me and tell me what came out of it.

Thanks

Thanks to Felix Rieseberg, who works on Claude at Anthropic, for the post and the prompt. And to Paul Krügel, an AI consultant in Vienna who builds his own software with agents. He works with the prompt too.

session-orchestrator

Your task becomes tested work, and the next session picks up where the last one stopped.

No language model
in the tool. Your coding agent writes the code.
The code
gives the agent its workflow: plan, work in rounds with tests after each one, wrap up. The state is kept as plain text in your repository.

session-orchestrator.com (opens another website)Product page for session-orchestratorWhat I build for you