| name | creative-autoresearch-control-plane |
| description | Use when AI should run an autoresearch-style loop for creative writing, iterating on chapters, reveals, arcs, and webserial hooks against explicit rubrics, benchmark cases, and rewrite decisions instead of stopping at the first plausible draft. |
Creative Autoresearch Control Plane
Use this skill when the task is not just "write a story" but "improve a story through repeated judged experiments."
This is the control-layer skill for the repository's autonomous creative workflow.
Quick Start
Use this skill when you need to:
- iterate on a chapter until it clears a quality bar
- benchmark one draft against another
- choose whether to rewrite prose, scene logic, chapter design, or arc structure
- improve addictive webserial properties such as hook strength, mystery carry, payoff, and strategic protagonist appeal
Pair with:
$strategic-webserial-architecture$faction-power-consequence-engine$strategic-cast-intelligence-engine$mystery-payoff-and-fairness-judge$autonomous-webserial-loop
Read references/system-map.md first.
Core Idea
Treat creative writing the way autoresearch treats model improvement:
- keep the editable surface constrained
- keep the evaluation budget explicit
- score outputs against a stable rubric
- keep or discard revisions based on judged improvement
The evaluator is the product. Generation is only one component.
Operating Loop
- Define the target unit:
- opening chapter
- chapter draft
- reveal beat
- confrontation
- arc scaffold
- Define the objective:
- stronger retention
- better clue fairness
- more strategic character play
- stronger leaderboard fit
- Select or compose the benchmark in
runs/<story-slug>/benchmark.md.
- Select the skill lenses that will judge, guard, and mutate the run.
- Produce or revise the unit under test.
- Judge it with the local and chapter-level rubrics.
- Decide the highest-level failure:
- prose
- scene
- chapter
- arc
- premise
- Apply only the necessary rewrite level.
- Log the iteration under
runs/<story-slug>/loops/<loop-id>/.
- Stop when the unit clears the bar or when more iteration would require new author intent.
Output Contract
Target unit:
Objective:
Benchmark class:
Rubrics used:
Current strengths:
Current failures:
Highest-level repair:
Planned intervention:
Revised output or revised plan:
Judge delta:
Stop or continue:
Rules
- Do not polish prose while structural failure is unresolved.
- Do not approve a draft because it feels smooth if it weakened obsession, tension, or payoff.
- Do not use benchmark rubrics as aesthetic dogma. They are instruments, not sacred law.
- Do not compare against copyrighted stories by imitation. Compare against abstract craft behaviors.
- Prefer measurable progress over vague approval.
