Ascend Skills Library
Core

Ascend build loop.

Spec, let a second model attack the plan, let it build, then judge the result yourself.

v2.1.0·updated 18 Sep 2026·stable
Use the Ascend Skills: Ascend build loop

Paste that line into your AI and it fetches this skill, whole and current, then asks you for what the skill needs. Not connected yet? How to connect.

Skill details

What it does

A feature or system taken from spec to built and approved, with a second model attacking the plan before any code is written and building it afterwards, and you judging the result against the spec on evidence.

Who uses it

operatorengineerin Core

What you need

You bring:

  • the ask in one line
  • the Codex companion plugin installed
  • a git repository to change

How it works5

The steps, in order. The method itself is served whole through the connector, current on every use.

  1. 1

    Write the spec and the plan

  2. 2

    Let Codex attack them until approved

  3. 3

    Transfer and let Codex build

  4. 4

    Judge the diff against the spec

  5. 5

    Approve and deliver, or escalate at the round cap

What it doesn't do

  • Research, content writing or reviews of existing work
  • Anything outward-facing, which still needs a human

When your AI reaches for it

Use when the operator invokes /ascend-build-loop or asks to take a feature/system from idea to built with a second model gating the work — "spec it, get Codex to tear it apart, then build", "run the build loop on this", "Codex reviews, you approve". For build tasks with code output. NOT for pure research, reviews of existing work, or content writing.

Latest change

v2.1.018 Sep 2026
Every work product now lands in the standard output tree (docs/SKILL-FORMAT.md, "Where a skill writes"): the spec, plan and review log for a run move from `<project-root>/.build-loop/<slug>/` to `.ascend/build-loop/<slug>/`, the agent-state directory under the repository being changed. No change to the loop.

See all 4 releases →

Related skills