Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add bendrucker/claude --skill hill-climbgit clone --depth 1 https://github.com/bendrucker/claudeWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/bendrucker/claude/hill-climb)<a href="https://agentmods.dev/skills/bendrucker/claude/hill-climb"><img src="https://agentmods.dev/badge/skills/bendrucker/claude/hill-climb/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/bendrucker/claude/hill-climb"><img src="https://agentmods.dev/badge/skills/bendrucker/claude/hill-climb.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00050 | $0.02271 |
| Opus 5.5 | $0.00020 | $0.00908 |
| Sonnet 5.5 | $0.00010 | $0.00454 |
| Haiku 4.5 | $0.00005 | $0.00227 |
Grade A, and why
performance:hill-climb scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 130 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Hill Climb
Goal: a program that is measurably better on the metric the user cares about, where every accepted change carries evidence that it beat run-to-run noise, kept behavior identical, and holds on inputs the climb never tuned against. The user stays in the loop at fixed checkpoints and can reconstruct every decision from one notes file.
This skill is the workflow. Load performance:benchmark to build the harness and compare arms, and performance:profile to find where the time goes. A repo's own benchmark conventions take precedence over the defaults here.
Arguments
<target>: the program or command to climb. Required.<notes file>: where the climb log lives. Default:CLIMB.mdbeside the benchmark harness, ortmp/climb.mdwhen the repo has no harness directory.
Frame
Interview the user until each item below is settled and none rests on your assumption:
-
Ask in rounds with
AskUserQuestion, one question per open decision. Load agrillingskill when one is installed. -
Settle every item before building anything.
-
Write the answers into the notes file's Rules section (references/notes.md has the template), then confirm the section with the user.
-
A dispatching brief settles only the items it states. Ask the rest.
-
When the user does not answer, send a push notification and continue. Record each unanswered item in Rules as
assumed: <value>, and name every assumed item in the first decision report. -
Metric. The number the user cares about, on an exact scenario: the command, its input state, and the machine state. "Wall time of a sync run by hand right after one upstream commit" is a metric. "Make it faster" is not.
- Start from who waits on the result and when: a person after a change, a CI job, a scheduled run. Measure the run the user waits through, even when the code handles another run more often.
- A scheduled run nobody waits on can take on work that speeds the waited-on one.
- Ask where the real workload spends its time. A benchmark input can skip the phase users wait through.
-
Ceiling. The value past which further gains stop mattering to the user. The climb stops there.
-
Fast signal. A cheaper or lower-noise number that moves with the metric: CPU time, instructions retired, work per tick, bytes. Accept on it only after the baseline shows it tracks the metric.
-
Guard. What must stay identical: output, final state, exit code. Name the check that proves it.
- When a candidate can change lookup order (
PATH, module or library search paths), the guard compares what each name resolves to. - Only the user relaxes a guard.
- When a candidate can change lookup order (
-
Side effects. Everything a run touches outside itself: network, remote state, installed files, caches. Decide per item whether to stub it, hold it fixed, or reset it before each run.
-
Scope. Which changes are allowed: code, config, dependencies, external tools, public API. Ask whether parallelism is in scope. When caching is in scope, ask where a cache may live and how it invalidates.
-
Inputs. Which scenarios are
dev(tuned against) and which areholdout(scored once at the end). Prefer real inputs over synthetic ones forholdout. -
Budget. How long the climb may run and how the user wants progress between checkpoints.
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday Changed · +12 lines 037eefcf49f9
- 19d ago Changed · +9 tokens per session 30c2cf932ab7
- 22d ago Changed · +2 lines 5e7cf198de05
- 23d ago First seen · 116 lines · 41 tokens per session scan A b0adaf9a9edb
performance:hill-climb is a skill published in the GitHub repository bendrucker/claude (18 stars, last pushed today), licensed MIT. It adds 50 tokens to every session and 2,271 once invoked, about $0.0002 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-25.
Other skills, from other repositories
ai-observability-promptfoo
Testing and evaluation framework for LLM prompts and applications -- promptfooconfig.yaml, assertions, model-graded evals, red teaming, CI/CD integration, custom providers, and comparative evaluation.
prompt-testing
Use when comparing two prompt variants, defining quality/efficiency/robustness metrics, or deciding whether to adopt a challenger prompt over a baseline.
prompt-optimization-loop
Prompt-engineering expertise for designing test cases, edge cases, adversarial inputs, and iterating on prompts based on eval results.
promptfoo-evals
Design and run promptfoo evaluations for LLM behavior regressions or comparisons across prompts, agent rules, and models. Use for repeatable behavioral cases with explicit assertions, not ordinary unit tests or static configuration checks.
prompt-evaluation-runner
Use when evaluating prompts, LLM outputs, red-team suites, or model behavior with local eval configs and safe provider/cost controls.
deep-plan
Creates detailed, sectionized, TDD-oriented implementation plans through research, stakeholder interviews, and multi-LLM review. Use when planning features that need thorough pre-implementation analysis.