I had Nick Beaird's ambition skill installed in Claude Code when I came across be-ambitious by RoboNuggets, and I wanted to know which one was better. I'd read both. I still couldn't tell you.
So I tested them, blind, on 9 real tasks from my own week, against the same agent with no skill at all. On the first 6 tasks neither skill scored higher than the plain agent. Then I wrote my own version. In the final round it beat the original on all 9 tasks. The skill and the test harness are in Kompas/ambition, with the full results.
Where the skills come from
In October Felix Rieseberg wrote How to be more ambitious: A tutorial for humans with models. It's worth reading. His point is that you usually ask a model for the task you already had in your head. That task is often a guess at what you want.
He describes 4 kinds of ambition. Speed is the same work, faster. Altitude aims at the outcome, so "fix this crash" becomes "keep the app crash-free". Depth means working like the best in the field, starting from the raw data, using their instruments and letting something other than the model check the result. Range brings in other fields, including an adversary, and borrows their methods.
Nick turned this into ambition (write-up). It gives a baseline, a ranked table of upgrades that each name an outside check, and a rewritten brief. Then it stops and you pick. RoboNuggets made be-ambitious, which writes an ambitious plan and then starts on the work.
Felix also says a model shouldn't grade its own work. That seemed like a good rule to apply to the skills themselves.
How I tested them
I took 6 requests from 1 week of work, from the chat with my business partner Tachmy and from the channel where 2 of our agents coordinate. Each came with the context an agent would have had at that point.
- A missing field in a matching API, after a customer batch matched only 54%
- Checking whether a new country works
- 1 in 6 jobs getting lost in large batches
- An idea for a new service
- A YouTube video that feels too AI-made
- 1 small question, "Where do we still use Atlassian?"
Each skill got the same prompt, "Be ambitious with this task", and so did the same agent with no skill as a control. Without a control you can't tell whether a skill does anything. Every combination ran 3 times, in a fresh Claude Opus 5.5 session with no tools and none of my own settings. Here's roughly the command for 1 run.
claude -p --model opus --system-prompt "$SKILL" --tools "" \
--setting-sources project --strict-mcp-config --no-session-persistence \
< task.txtI didn't want Claude grading Claude, so the judges were GPT-6-Astra (through Codex) and Grok. They saw all answers to a task at once, shuffled and unlabeled. Each answer got a score on 5 things, and the judges named its biggest weakness.
- Would the owner adopt it
- Each upgrade has an outside check
- It sticks to the given context
- It makes the task deeper without making it bigger
- A busy person can act on it right away
The tasks and the prompt were in Dutch, so the quotes below are translated.
The skills didn't beat the plain agent
| Variant | Overall (1-10) | Outside check | Scope | Usable |
|---|---|---|---|---|
| no skill | 5.67 | 2.50 | 3.22 | 3.14 |
ambition | 5.64 | 3.56 | 2.78 | 2.72 |
be-ambitious | 4.92 | 3.19 | 2.39 | 1.69 |
ambition did what Felix asks, and of the 3 it scored highest on outside checks. 1 run on the lost jobs even questioned the task's own assumption. Maybe the product codes Google UK doesn't know were burning through all 15 retries, and maybe that retry storm is what pushed the server into its time limit.
That idea sat in an answer of about 1,300 words. For "Where do we still use Atlassian?" the 2 skills wrote 550 to 800 words, with an access audit and a migration plan. The answer was 1 line. Bitbucket, for 1 old repo.
v2 changed what the agent writes down
The judges kept naming the same weaknesses. I kept Felix's method and changed only the output.
| What the judges flagged | What v2 does instead |
|---|---|
| No answer at the top | A fixed order. The answer, what's known, up to 3 upgrades, a brief, what it needs from you |
| Long answers | Works through Altitude, Depth and Range without printing them |
| Invented facts, like "the existing golden set" | A Known block. Anything not in it is labeled "(assumption)" |
| The table says you pick, the brief adopts everything | The brief holds the baseline plus upgrade #1. The rest get 1 line each |
| Fixes before the cause is known | Measure first when the cause is unknown |
| Targets the data says are out of reach | Targets must fit the numbers in Known |
| 600 words for a 1-line question | Small task. Answer in line 1, under 80 words |
The skill itself went from 1,885 to 909 words.
v2 won round 2 with 6.06, against 5.42 for no skill and 5.25 for ambition. Its scope score went to 3.92 while ambition stayed at 2.64. On outside checks it scored lower than ambition, 2.89 against 3.56. The weaknesses showed why. v2 often listed "me, today" as the check, which is the agent grading itself again.
Then v2 lost on new tasks
I wrote v2 from the mistakes on the same 6 tasks it then won on. That's sitting a test when you already know the questions. So I added 3 tasks I hadn't looked at while writing it. A nightly spike of useless requests, a monitoring service exposed to the internet, and how to get our new site ranking in Google.
On those, ambition scored 6.44 and v2 6.06. No skill got 4.78. v2 was shorter and easier to use. ambition had the better ideas.
Without those 3 tasks I would have published v2 as the winner.
v3 added 2 rules
The first rule is that the check is never the agent. "I'll review the logs" doesn't count. "The error count in Grafana for that batch" does, and the skill asks who reads it, when, and what counts as a pass.
The second is that a requested feature gets built. Measuring first is for a problem with an unknown cause. When someone asks for a feature, the agent delivers it and the upgrades build on it.
The final blind round had every variant on all 9 tasks.
| Variant | 6 original tasks | 3 new tasks | Median length |
|---|---|---|---|
| v3 | 6.28 | 6.44 | 487 words |
| v2 | 6.08 | 6.06 | 494 words |
ambition | 4.83 | 6.00 | 573 words |
| no skill | 5.06 | 4.67 | 858 words |
v3 scored higher than ambition on all 9 tasks, by 1.1 points on average (95% interval 0.7 to 1.6). Both judges, separately, ranked v3, v2 and ambition in that order.
Take the Atlassian question. ambition opens with a header and a baseline section that does mention Bitbucket, and then roughly 500 more words follow. v3 answered in 76 to 87 words. In the first run, line 1 says that as far as it knows only Bitbucket is in use, for 1 repo. Then 1 upgrade, because the real question is probably whether the subscription can go, so put the billing lines next to the last activity per product.
All 3 v3 runs on the lost jobs opened with the same advice. Find out where those jobs fail before fixing anything.
What this doesn't show
The judges are models. They judged blind and come from other labs, but I haven't ranked the answers myself yet. It's also 1 company, 1 week and 1 model, with 9 tasks and 3 runs each.
The agents had no tools, so I measured plans, not what happened when someone ran them. That hurts be-ambitious most, since it's built to start working right away. Nick's own eval does run the brief against test fixtures, and mine doesn't.
I also wrote both v3 and the scoring questions. Those questions reward a short answer at the top, which is what v3 was built to give. The 3 new tasks protect against fitting to the tasks. They don't protect against fitting to the rubric.
2 results don't flatter v3. On the new tasks ambition scored highest on value and was picked as the best single answer most often. v3 wins on average because it's easier to use, and I'm not claiming it thinks better. v3 also misses its own length limit. The skill says under 350 words, and its answers ran around 490.
Try it on your own tasks
The harness in Kompas/ambition runs any set of skills against a no-skill control, and Codex and Grok judge the answers blind. The full results are there too. What wins on my week may not win on yours.
2 things made the difference for me. The control, because without it I wouldn't have seen that the original skills added nothing on average on my first 6 tasks. And the 3 tasks I held back, because they're the reason I didn't publish v2 as the winner.
The method is Felix Rieseberg's. The first skill is Nick Beaird's, and it was my starting point. The second is by RoboNuggets.