git:20261106.f0caf8b to git:20261106.82f9a81

---

name: hillclimb

description: 'Improve an existing artifact round after round toward a goal — a page, a prompt, a document, code whose speed matters, a simulation — keeping a change only when it measures better, until rounds stop paying. Runs figure-out, define and do; quick by default, thorough when asked or when the stakes call for it. Use when the user asks to hill-climb, optimize, keep improving or make something as good as it can be for its goal, or to continue an earlier climb.'

argument-hint: '<target> [goal and measure] [--autonomous]'

user-invocable: true

---

Climb the target the arguments name: a file, a folder, a page, a prompt, a component. With no

target given or inferable from the conversation, halt with usage. The rest of the arguments say

the goal and, if the requester has one, the measure, in their own words; results from a test on

real use, such as an earlier A/B test, may come with them.

This command holds no loop of its own. figure-out, define and do run the climb, and their task

files for climbs and measured comparisons say how: what counts as better, how rounds try changes,

when a gain is confirmed, and when to stop. Tell every handoff this is a hill-climb on the target, so each loads its task file for

climbs, and pass the request as given, the

requester's words on the goal and the measure, and these limits:

- **The assignment** is to improve the target toward the goal, inside what the climb protects.

It does not authorize a push, a pull request, publishing, or a test that puts versions before

real people, unless the requester asks for it in the request, their rules or their context file.

- **Quick or thorough.** A climb is quick unless the requester asks for a thorough one or the

request shows what is at stake: a long run, such as many rounds or one that keeps running

unattended under a scheduler or coordinator; a wrong result that is costly or hard to undo; or

objectives that pull against each other. `--autonomous` alone does not make a climb thorough.

Pass your reading on; defining may revise it as it learns the artifact.

**Without `--autonomous`:**

- A **thorough** climb: invoke `figure-out` with the climb request, then `define`, with the

requester taking part — they settle what counts as better, whether the measure reflects the

goal, how objectives trade off, and what must not change.

- A **quick** climb: ask the requester only when the goal cannot be inferred from the request and

**With `--autonomous`:** invoke `figure-out` with `--autonomous` where shared understanding is

missing, then `define` with `--autonomous`. Nobody is asked from start to end.

The request to climb is the request to run it: when `define` reports the Manifest path, invoke

`do` with that exact path in the same turn, unless the requester asked to stop at the plan. No

path → stop and report. Before the first round, tell the requester the line opening the Manifest's Initial Approach:

quick or thorough and why, how rounds try changes, and how the climb measures. They change any

of it in plain words; a change goes to `define` as an amendment.

When `do` finishes, an attended climb asks whether to push the result, unless the requester

already said. An autonomous climb with no instruction to push keeps its commits local and says so.

**A re-run** on a target already climbed continues that climb: find its Manifest — among the

Manifests whose execution logs name this target, or through a record the requester names — and

pass its path to `define` to amend, with any results from real use the requester supplied as the