brainstorm

name	brainstorm
description	Strategic brainstorming for newsletter quality. Analyzes the pipeline holistically, identifies gaps in source coverage, and proposes new features, strategies, and scorer dimensions. Use for high-level quality improvement ideas.
disable-model-invocation	true
allowed-tools	["Bash","Read","Grep","Glob","Agent"]

/brainstorm — Strategic Quality Brainstorming

You are running a strategic brainstorming session for Astral Index. Your job is to analyze the current pipeline holistically and propose high-impact improvements to source coverage, pipeline stages, strategies, and evaluation.

Arguments

$ARGUMENTS may contain a focus area:

• sources — focus on source coverage gaps
• pipeline — focus on pipeline stage improvements
• strategies — focus on new authoring strategies
• eval — focus on evaluation system gaps
• (empty) — cover everything

Step 1: Current State Inventory

Read the following to understand the current state:

# List configured sources
uv run --package astral-ingest astral-ingest sources

# List available strategies
uv run --package astral-author astral-author strategies

Read these files:

• ARCHITECTURE.md — roadmap and design overview
• packages/ingest/src/astral_ingest/sources.yaml — all configured sources
• packages/author/src/astral_author/pipeline.py — strategy definitions

Check for recent drafts and eval results:

ls -t data/drafts/*.md 2>/dev/null | head -3
ls -t data/eval/*.json 2>/dev/null | head -3

If recent drafts or eval results exist, read them for context on current output quality.

Step 2: Reference Analysis

Read ARCHITECTURE.md for the description of The Orbital Index's editorial DNA and the project's aspirations.

Identify gaps between current output and the reference:

• Link density: The Orbital Index typically had 15-25 links per issue. How does the current output compare?
• "Paper of the Week": A signature feature — deep technical analysis of one academic paper. Does the current pipeline support this?
• Technical depth: Per-item analysis vs. surface-level summaries
• Editorial voice: Consistent, knowledgeable, opinionated vs. neutral aggregation
• Issue consistency: Predictable structure readers can rely on across hundreds of issues

Step 3: Source Gap Analysis

Analyze sources.yaml for coverage gaps:

By organization type

• Space agencies: NASA, ESA, JAXA — check which are present. Look for missing: ISRO, CNSA/CASC, Roscosmos, KARI, CSA, ASA, UKSA
• Commercial: SpaceX, Blue Origin, Rocket Lab, etc. — check blog/press feeds
• Government/regulatory: FAA (launch licenses), FCC (spectrum filings), GAO reports
• Academic: arXiv, NASA ADS, specific journal RSS feeds

By geography

• Are non-US/European sources adequately represented?
• Asia-Pacific space programs (JAXA, ISRO, KARI, CNSA)
• Emerging space nations

By content type

• Launch manifests and schedules
• Financial/investment news (space SPACs, funding rounds)
• Policy and regulatory developments
• Earth observation and climate applications

Source health

If item data exists, check for sources that produce very few or no items:

ls data/items/ 2>/dev/null | head -20

Step 4: Pipeline Proposals

For each pipeline stage, propose specific improvements:

Ranker (rank.py)

• Authority-weighted scoring (original reporting > aggregator rewrites)
• Breaking news time bonus (items < 24h old get a boost)
• Source tier system (primary > secondary > aggregator)
• Geographic diversity bonus

Clusterer (cluster.py)

• Cross-category narrative threads (e.g., "Artemis update" spanning policy + launch + science)
• "Paper of the Week" as a dedicated section type
• Configurable section templates per strategy
• Better handling of small categories (the "In Brief" problem)

Summarizer (summarize.py)

• Fact-checking layer (compare summary claims against source text)
• Trend identification (connect current item to previous coverage)
• Technical depth levels (brief/standard/deep for different section types)
• Multi-source synthesis (combine overlapping items into richer summaries)

Drafter (draft.py)

• "In Brief" section with editorial prose (not just bullet points)
• Configurable output formats (full newsletter, digest, breaking-news flash)
• Section-level editorial voice calibration
• Statistics/data callout formatting

Eval (scoring.py, llm_judges.py)

• Timeliness scorer (are recent items prioritized appropriately?)
• Factual grounding scorer (do summaries match sources?)
• Information density metric (insight per word)
• Link quality scorer (are links to primary sources, not rewrites?)
• Reader engagement proxy (headline quality, hook effectiveness)

Step 5: Strategy Proposals

Propose 2-3 new strategies with specific configurations:

deep-analysis

• Fewer items (top 10-15 instead of 20-30)
• Longer prose per item (3-4 sentences instead of 1-2)
• Trend identification enabled
• Higher min_group_size to force only strong clusters
• Target audience: industry professionals who want depth

breaking-news

• 24-48h lookback window
• Minimal prose (headline + 1 sentence)
• Aggressive recency weighting (w_recency > 0.6)
• Fast turnaround, lower quality bar
• Target audience: people who want to know what happened today

research-focus

• Prioritize arXiv papers and agency technical reports
• "Paper of the Week" section mandatory
• Technical depth setting = deep
• Lower weight on news aggregators
• Target audience: researchers and engineers

For each strategy, specify the config values that would differ from baseline.

Step 6: Prioritized Output

Rank all proposals by:

• Impact (H/M/L) — how much would this improve newsletter quality?
• Effort (H/M/L) — how much work to implement?
• Dependencies — does this require other changes first?

Format as an actionable list:

### High Impact, Low Effort (do first)
1. ...

### High Impact, High Effort (plan for)
2. ...

### Medium Impact, Low Effort (quick wins)
3. ...

### Ideas for Later
4. ...

For each item, note which /iterate target it maps to (scorer name or pipeline stage).

Important Notes

• This skill is read-only — it does NOT modify any files or code.
• All output is conversational, presented to the user for decision-making.
• Be specific and actionable. "Improve the ranker" is not useful. "Add a source_tier field to sources.yaml and weight items by tier in rank.py" is.
• Ground proposals in what you actually see in the code, not hypotheticals.
• If you identify a source that should be added, include the actual URL if you can find it in the existing sources.yaml patterns.