---
name: hill-climb
description: Run a bounded, evidence-driven improvement loop for code, prompts, skills, websites, or generated media. Invoke only when the user explicitly asks for /hill-climb or a measured improvement loop. Freeze a sensitive evaluation, measure baseline noise, change one hypothesis at a time, keep or revert from evidence, and confirm the winner on untouched cases.
license: MIT
compatibility: "Harness independent. Python 3.11+ for the optional offline comparison helper. Actual experiments require the project's runner and relevant runtime or media tools. Typed contract tools require Python 3.11+ and the managed command-schema runtime."
metadata:
  author: Tyler Kendrick
  version: "0.3.0"
  disable-model-invocation: "true"
---

# Hill-climb

<!-- tkstack:execution:start -->
Tool: `.tkstack/tkstack skills hill-climb -- <command>`. Omit the command for deterministic man-page help.

| Executable operation | Contract |
| --- | --- |
| Default/help and man(1) export | [Reserved help](commands/help.md) |
| Declared package version | [Reserved version](commands/version.md) |
| `model` | [Typed parameters](commands/model.md) |
| `verify` | [Typed parameters](commands/verify.md) |
| `native compare` | [Typed parameters](commands/native_compare.md) |
| Evidence, effects, native execution, and semantic boundaries | [Execution contract](references/execution.md) |

`model` reads the explicit contract; `verify` checks retained facts/evidence. Native commands execute existing code. Residual semantic decisions use the focused procedures below; a recorded review is not an independently authenticated judgment.
<!-- tkstack:execution:end -->


Improve an actual outcome through bounded attributable experiments. This skill
does not start an experiment, spend money, or expand merge authority by existing.

| Stage | Read before acting |
| --- | --- |
| Start or resume an optimization | [Full experiment procedure](references/operating-guide.md) |
| Establish the feedback loop | [Instrument and loop](references/instrument-and-loop.md) |
| Compare paired case metrics | [Measurement contract](references/measurement.md) |
| Inspect example plan | [Plan template](templates/plan.json) |

Ground the real workload and prove the instrument can distinguish good/bad
behavior. Freeze inputs, metric, runner, grader, guardrails, splits, threshold,
and budget. Measure unchanged noise before proposing a win. Inspect actual
failures; distinguish a failed launcher or empty trace from model behavior.

Change one causal hypothesis per round; compare with the same incumbent; keep
or revert from actual evidence and retain rejected attempts. Bound independent
work, isolate mutable state, and avoid contention that invalidates measurement.
Confirm a fixed winner on untouched cases. Preserve user targets and stop rules.

Use the real project runner; `scripts/hill_climb.py` checks supported receipts,
not experiment truth. Missing proof stays inconclusive. An active SDLC workflow
and SkillOpt's custom-skill adoption gate retain their guarantees. See
[sources](references/sources.md) and [added synthesis](references/distilled-sources.md).

For session/repository/workspace records, read [local state storage](references/storage.md).