Skip to content

/lc-new

Scope a new ASTRA analysis from a research question, through conversation. The output is a complete astra.yaml and (optionally) a literature evidence trail. Implementation comes later — /lc-new writes spec, not code.

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), Edit(index.md),
Glob, Grep, Bash(astra:*), Bash(lc:*),
WebSearch, WebFetch, AskUserQuestion, Agent

Writes are locked to the spec surface — no Python, no R, no arbitrary files. The lc-extractor subagent is dispatched via Agent.

Phases

  1. Research question. Sharpen the question, then write version, name, and description to astra.yaml so the user has something visible to react to from the first turn.
  2. Analysis structure. Walk through inputs, outputs, and any sub-analyses. One output per output: one metric, one plot, one artifact. inputs: and outputs: land in astra.yaml as they crystallize.
  3. Deep dive (per section). An optional literature pass. Collect paper candidates with the user; for each approved paper, dispatch one lc-extractor subagent in parallel. Each subagent reads the PDF, pulls verbatim quotes, runs astra paper verify-quotes to machine-verify the quotes against the source, and returns prior insights. Decisions then fall out of the conversation and the literature together.
  4. Finalize. astra validate astra.yaml; astra validate --verify-evidence if quotes exist; astra universe generate -n baseline. Author a short description: on the root analysis (one or two paragraphs orienting a reader), then append a ## Project Notes section to CLAUDE.md with conversational context the spec doesn't carry, and update the scaffolded intro's "has not been scoped yet" framing. The scaffolded MyST report (index.md) still references the boilerplate ids (example_method, main_result); swap those for a real decision ID and output ID from the finalized spec so the report doesn't ship with dangling references.

Writes happen at the end of each phase, not in bulk — the user always has something visible to review.

Hard restrictions (from the SKILL.md)

  • Specification agent only. No Python, no R, no implementation code.
  • Touchable files: astra.yaml, universes/*.yaml, CLAUDE.md (Finalize only), and index.md (Finalize only, reference IDs only).
  • Quotes are never fabricated; every evidence entry must pass astra validate --verify-evidence.
  • PDFs stay inside lc-extractor subagents — the main agent never pulls one into its own context.

Anti-patterns called out in the prompt

  • Bulk-writing decisions at the end instead of after each crystallizes.
  • Letting vague goals like "analyze this data" pass without sharpening.
  • Method-only decisions. The prompt actively probes data exclusion, variable operationalization, and inference criteria.
  • Reading PDFs in the main agent context.
  • Skipping astra validate --verify-evidence.
  • After /lc-new, ask the agent to implement the spec through the normal Claude Code workflow.
  • /astra — astra.yaml schema, decision criteria, prior insights / findings, universe management.
  • claude/lightcone/agents/lc-extractor.md — the literature extraction subagent definition.