/create-skill-autoresearch

name: create-skill-autoresearch
version: 0.2.1
license: MIT
description: >-
  Factory skill that creates production-grade, benchmarked, autonomously improved,
  and verified agent skills. Orchestrates a 5-phase pipeline: interview the
  user to discover purpose and gold standards, research domain materials with parallel
  subagents, draft the skill with a design-first approach, invoke autoresearch to
  iterate against gold-standard-driven LLM-as-judge evaluation, and verify quality
  through multi-agent consensus with a devil's advocate. Use when building a new
  skill, creating a skill from existing materials, or upgrading a skill to
  production quality with benchmarking and autonomous improvement.

Create Skill via Autoresearch Factory

A factory for forging production-grade agent skills through gold-standard-driven autoresearch, multi-agent verification, and structured consensus.

The factory orchestrates 4 agent roles through 5 phases:

| Phase | What Happens | Agent Role | |-------|-------------|------------| | 1. Interview | Discover purpose, gold standards, scope | ORCHESTRATOR | | 2. Research | Study domain materials, build dossier, propose rubric | RESEARCHER (N parallel) | | 3. Draft | Design structure, generate SKILL.md, measure baseline | BUILDER | | 4. Autoresearch | Iterate skill against gold standards (LLM-as-judge, or an objective real-world metric for procedural skills — see 3.4) | BUILDER + autoresearch skill | | 5. Verify | Premortem, panel scoring, consensus, ship/iterate | PANEL (3 subagents) |

Key constraint: BUILDER and PANEL never share context. Panel receives only the skill output, gold standards, and rubric -- no bias from the building process.

Relation to create-skill

This factory **extends** the official single-pass skill creators (Anthropic's Skills best-practices and `skill-creator`; Cursor's `create-skill`) rather than replacing them. It adds what a one-shot generator cannot: a research dossier, gold-standard benchmarking, an autonomous improvement loop, and independent multi-agent verification. The skills it produces follow the same official conventions -- see [references/skill-authoring-best-practices.md](references/skill-authoring-best-practices.md).

Companion skills

The factory orchestrates these sibling skills at runtime: **autoresearch** (Phase 4 improvement loop), **premortem** (Phase 5 risk pass), and **handoff** (cross-session continuity); the Phase 5 panel/consensus design draws on **llm-council**. In this harness they are vendored under `.agents/skills/`. If you install this skill standalone, install those alongside it. The factory's craft layer ([references/skill-craft-principles.md](references/skill-craft-principles.md)) is distilled from **writing-great-skills** ([mattpocock/skills](https://github.com/mattpocock/skills), MIT), which the harness vendors under `.agents/skills/`. Phase 5 will delegate to **tribunal** when it is installed (see 5.2).

---

Phase 1: Interview

Discover what the user needs through structured questions. Do not assume -- ask. Ask them one topic at a time and record the answers; the exact question wording and follow-ups are in [references/pipeline-phases.md](references/pipeline-phases.md).

1.1 - 1.5 What to discover

| Topic | Ask about | Record | |-------|-----------|--------| | **Purpose and domain** | the problem it solves, the domain, which agent will use it, what "success" looks like | `SKILL_PURPOSE`, `DOMAIN`, `TARGET_USER`, `SUCCESS_CRITERIA` | | **Gold standards** | examples of "what good looks like" — input/output pairs, reference artifacts, previously solved problems, existing quality reports; where they are, what format, how many | `GOLD_STANDARD_SOURCE`, `GOLD_STANDARD_FORMAT`, `GOLD_STANDARD_COUNT` | | **Study materials** | docs, existing code, transcripts, design docs, reference implementations, specs, style guides | `STUDY_MATERIALS` | | **Scope and constraints** | conventions to follow, skills to integrate with, anti-patterns to avoid, target line count (default < 500), and invocation mode — model-invoked (pays permanent context load) or user-invoked (`disable-model-invocation: true`) | `CONSTRAINTS`, `INTEGRATION_SKILLS`, `ANTI_PATTERNS`, `INVOCATION_MODE` | | **Existing skill** | is there one for this domain already? If so it is both a study material and a baseline — research it, measure it against the rubric, improve it rather than starting over | `EXISTING_SKILL`, and set mode to **upgrade** rather than **greenfield** |

Minimum 3 gold standards. Fewer is a risk -- warn the user and offer alternatives: create synthetic examples, or find additional reference materials.

1.6 Confirm and Create Workspace

Summarize all parameters in a table. Ask the user to confirm.

Once confirmed, create the build workspace at `builds/<skill-name>/` with three ownership zones:

• `input/` -- where the user drops gold standards + study materials, in **any** structure
• `work/` -- everything the factory generates: `manifest.yaml`, `research/`, `evaluation/`, `experiments/`, `handoffs/`
• `output/<skill-name>/` -- the finished skill (`SKILL.md` + `references/`) in its own named dir, publish-ready

Do **not** ask the user to hand-author a manifest. Scan whatever is in `input/`, classify each item as a gold standard (exemplar input/output pair or reference artifact) vs a study material, and write your derived index to `work/manifest.yaml` with train/validation/test tags. Present the derived manifest for the user to confirm or correct. See [references/pipeline-phases.md](references/pipeline-phases.md) for intake formats and the manifest schema.

---

Phase 2: Research

Study the domain thoroughly before writing any skill code.

2.1 Spawn Researcher Subagents

Cluster study materials by relatedness, then launch one `explore` subagent per cluster. Clustering heuristic:

• **By source type**: existing skills in one cluster, gold standard outputs in another, planning docs in a third
• **B