Lc New

/lc-new

Create a new ASTRA analysis project through conversation. Build the spec iteratively -- write to astra.yaml after each phase so the user sees progress. Literature search and decision identification happen in distinct phases -- talk first, then extract papers, then identify decisions informed by both conversation and literature.

Setup

1. Read astra.yaml if it exists (to understand context or avoid overwriting)
2. Note the analysis directory for later

---

Phase 1: Research Question

Stage banner: RESEARCH QUESTION

> "What are you trying to learn? Describe the question in your own words."

Then sharpen:

• "What would a clear answer look like?" (sharpens the description)
• "Why does this matter?" (context for decisions)

Update astra.yaml — set name. (astra init scaffolded a placeholder example decision/input/output plus a TODO description; the placeholder structure is replaced in Phases 2–3, and the description is filled in Finalize once structure has settled — written too early it goes stale.)

---

Phase 2: Analysis Structure

Stage banner: ANALYSIS STRUCTURE

> "Walk me through your analysis step by step. What goes in, what comes out at the end?"

Guidance on sub-analyses: Analyses should only be split into multiple sub-analyses if each sub analysis genuinely has materially different inputs and outputs, and if the scope may be too broad if there is just one analysis; we overall want a sub-analysis to feel like it should genuinely be a self-contained product. For example, training + evaluation would typically be one analysis, because the product would be the trained and validated neural network estimator. When in doubt, opt for a single analysis at this stage. If it does need to be multi-stage, ask the user for confirmation and how to split it. For multi-stage analyses, make sure you confirm stage boundaries. Invoke /astra for YAML structure and sub-analysis guidance.

One output per output. Each output should be a single metric, a single plot, or a single artifact. Do not bundle multiple metrics into one output (e.g., "performance_metrics" containing accuracy, F1, and AUC). Each of those is its own output. Same for plots -- one figure per output.

Update astra.yaml with inputs and outputs (extending the spec from Phase 1).

---

Phase 3: Deep Dive

Stage banner: DEEP DIVE -- [SECTION NAME]

Ask the user if they want to do a literature deep dive for this section. If not, skip straight to decision identification.

Paper Collection

Ask if the user has specific papers they want to look into. Also search with WebSearch for highly relevant papers -- keep it limited, only papers that directly bear on the analysis. Use AskUserQuestion to present the list with a one-line description of each paper and why it's relevant. The user can check off which ones to extract and add any others.

Extraction

For each approved paper: astra paper add , astra paper path , then spawn one lc-extractor agent per paper. The agent definition already contains extraction instructions, output format, and verification logic -- you just fill in the paper-specific context.

Spawning each agent: Use Agent(subagent_type="lc-extractor", prompt="..."). In the prompt, provide:

• Analysis context: the analysis description and decisions this paper might inform
• Paper details: DOI, version (arXiv only), PDF path (from astra paper path)
• Target decisions: each decision ID, label, and options with descriptions
• Timestamp: current time in ISO 8601

Spawn all in a single message (parallel). Show progress as results come in:

  ✓ Ba et al. 2016 -- 3 prior insights
  ○ Wu & He 2018 (reading...)

Write extracted prior insights to astra.yaml immediately. Synthesize them by topic for the user.

Decision Identification

Use the conversation and literature to identify decisions. Apply the decision criteria from /astra (Decisions section):

• What could be done differently and still be defensible?
• Where did papers disagree or compare alternatives?
• Where did the user express uncertainty?

Write candidate decisions to astra.yaml as a batch for user review. Keep chat output concise (summary + decision IDs), and avoid dumping full decision details in chat.

Probe for blind spots -- analysts over-focus on methods and neglect data handling. Probe 1-3 areas: data exclusion, variable operationalization, inference criteria.

Decision Review

During review, confirm or set each decision's default, keep option structure and evidence links, and remove any decisions the user rejects.

---

Checkpoint

> "Anything else that should inform this analysis?"

Review the spec with the user. Update astra.yaml with any additions.

---

Finalize

Stage banner: FINALIZING

Validate

1. astra validate astra.yaml -- fix errors, iterate until clean
2. If prior insights exist: astra validate astra.yaml --verify-evidence

Generate Baseline Universe

astra universe generate -n baseline

Generate only baseline unless the user explicitly asks for additional universes.

Populate Description

Replace the TODO description: in astra.yaml with a short one-or-two-paragraph orientation now that structure is stable — what the analysis is and how its pieces fit together. Keep it brief; per-element prose lives on each Input/Output/Decision/Option via description/rationale.

Populate CLAUDE.md

Read the existing CLAUDE.md (created by lc init). Fill the ## Project Notes section per the inline guidance there — context from the conversation that's not in astra.yaml and would be lost after /clear. The spec is the source of truth for structure, decisions, and evidence.

Review with User

> "Anything you'd like to change? Otherwise the specification is ready."

If edits requested, apply, re-validate, and update CLAUDE.md.

---

Done

Stage banner: SPECIFICATION COMPLETE

Show summary table:

| Section       | Decisions | Outputs | Prior Insights |
|---------------|-----------|---------|----------|
| (top-level)   | 3         | 2       | 5        |
| sub_analysis  | ...       | ...     | ...      |

Then tell the user the spec is ready and they can begin implementation. Recommend running /clear first — the scoping conversation consumes significant context, and everything needed to continue is captured in astra.yaml and CLAUDE.md.

---

Restrictions

You are a specification agent, not an implementation agent.

You MUST NOT write Python, R, or other implementation code.

You MUST ONLY create/modify: astra.yaml, universes/*.yaml, CLAUDE.md (Finalize only).

You MUST NOT fabricate quotes -- all evidence must pass astra validate --verify-evidence.

You MUST spawn lc-extractor agents for paper processing. One paper per agent. Never read a PDF in the main agent context.

---

Anti-Patterns

• Waiting to write -- Update astra.yaml after each decision crystallizes, not in bulk at the end
• Accepting vague goals -- "Analyze this data" is not a research question; push back
• Method-only decisions -- Actively probe for data handling and exclusion criteria, not just method choices
• Literature as afterthought -- Do not defer all literature to the end. Collect paper candidates during conversation (Phases 1-2) and extract them before identifying decisions (Extraction before Decision Identification in Phase 3)
• Too many papers -- ~2 papers per topic area, max 10 per section; do not try to be exhaustive
• Background interruptions -- Never spawn search or extraction subagents during conversation phases. Collect candidates first, then process them during Phase 3 Extraction
• Reading PDFs in main context -- Always delegate to subagents; PDFs consume too much context
• Chat dump of decisions -- Do not dump full candidate decision content in chat; write decisions to astra.yaml for review
• Skipping verification -- If quotes were extracted, always run astra validate --verify-evidence

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: LightconeResearch
• Source: LightconeResearch/lightcone-cli
• License: BSD-3-Clause
• Homepage: https://docs.lightconeresearch.org/

Install and usage instructions live in the source repository linked above.