Skip to content

/lc-new

Scope a new ASTRA analysis through conversation. Produces a complete astra.yaml (and optionally a literature evidence trail) with no code written.

Source: claude/lightcone/skills/lc-new/SKILL.md.

Allowed tools

Read, Write(astra.yaml), Write(universes/*), Write(CLAUDE.md),
Edit(astra.yaml), Edit(universes/*), Edit(CLAUDE.md),
Glob, Grep, Bash(astra:*), Bash(lc:*), Bash(mkdir:*), Bash(echo:*),
WebSearch, WebFetch, AskUserQuestion, Task

The skill is locked to spec-only writes — it cannot write Python, R, or arbitrary files. The lc-extractor subagent is invoked via Task.

Phases

  1. Research question — sharpen the question, write version, name, description to astra.yaml immediately so the user sees progress.
  2. Analysis structure — walk through inputs, outputs, sub-analyses. One output per output: a single metric, a single plot, a single artifact. Updates astra.yaml with inputs: and outputs:.
  3. Deep dive (per section) — optional literature pass. Collect paper candidates; for each approved paper, spawn one lc-extractor subagent (parallel, via Task). Each subagent reads the PDF, pulls verbatim quotes, runs astra paper verify-quotes to machine-verify the quotes against the source, and returns extracted prior insights. Then identify decisions informed by the conversation + literature and write them to astra.yaml.
  4. Finalize — astra validate astra.yaml, astra validate --verify-evidence if quotes exist, astra universe generate -n baseline, populate the ## Working Notes section of CLAUDE.md with conversational context not captured in the spec.

The skill writes to astra.yaml after each phase rather than in bulk at the end so the user has something visible to review at every step.

Hard restrictions (from the SKILL.md)

  • Specification agent only — cannot write Python, R, or other implementation code.
  • Files it may touch: astra.yaml, universes/*.yaml, CLAUDE.md (Finalize only).
  • Never fabricates quotes — all evidence must pass astra validate --verify-evidence.
  • PDFs are read by lc-extractor subagents only; the main agent never pulls a PDF into its own context.

Anti-patterns called out in the prompt

  • Bulk-writing decisions at the end instead of after each crystallizes.
  • Accepting vague goals like "analyze this data" without sharpening.
  • Method-only decisions; the prompt actively probes for data exclusion, variable operationalization, inference criteria.
  • Reading PDFs in the main agent context.
  • Skipping astra validate --verify-evidence.