digest-insights

name	digest-insights
description	Answer insight-style questions about recent dev/AI/research news using the code-intel-digest local mirror (hourly-refreshed Postgres). Use when the user asks what's trending, wants a weekly roundup/podcast-script, asks how a topic is being discussed, or wants a ranked list of items to read. Requires the code-intel-copilot MCP server to be registered (`claude mcp add code-intel-copilot -- npx tsx /absolute/path/to/src/mcp/server.ts`).
allowed-tools	mcp__code-intel-copilot__search_items, mcp__code-intel-copilot__semantic_search_items, mcp__code-intel-copilot__get_item, mcp__code-intel-copilot__aggregate_items, mcp__code-intel-copilot__mirror_status, Read, Write

Digest Insights

You are answering insight-style questions over a local Postgres mirror of a personal news/research/podcast ingestion pipeline (the "code intel digest"). The mirror contains tens of thousands of items across 9 categories (ai_news, ai_dev, tech_articles, research, newsletters, podcasts, community, product_news, marketing), with LLM relevance scores and semantic embeddings for most items.

Always do first

1. Call mirror_status exactly once and check dbMode:
   • dbMode: direct — the server is reading the live production database (read-only role). Data is real-time; mirror freshness fields are N/A. Note "Data: live production (real-time)" at the top of the answer and continue.
   • dbMode: mirror — the server is reading the hourly local mirror. If staleMinutes > 90, warn the user that data may be stale and ask whether to proceed. If staleMinutes <= 90, note the freshness in one sentence at the top ("Data through ") and continue.

Do not skip this step. Users need to know how fresh the answer is.

Tools (in priority order)

Tool	When to use
semantic_search_items(query, limit)	Conceptual / theme queries: "how are people thinking about X", "context engineering", "what's changing about Y". Prefer over keyword search when the query is abstract or uses shifting terminology.
search_items(query?, category?, since?, until?, limit?)	Named-entity queries ("anthropic earnings"), date-bounded lookups, or category-restricted lists. Also how you get "most recent items" — omit query.
aggregate_items(group_by, since?, until?, category?, limit?)	"Top sources publishing on X", "most active authors this month", "category distribution last week".
get_item(id)	Only after a search — when you need the cached full_text to quote accurately or deeply analyse a specific item.
mirror_status	First call of every session.

Core principles

• Cite {id, url} for every item you reference. The user may want to click through. If you mention "a recent article about X", include the URL.
• Don't overclaim. If you surface 5 items about a topic, don't present that as "the whole industry is doing X". Say "I found 5 items discussing X".
• Respect the score. item_scores.final_score (surfaced via aggregate_items.avgFinalScore) is the project's own LLM-based useful-to-Stephanie signal. Higher scores are better; weight toward them when ranking.
• Dates are tricky. Items have published_at (source publish time) and created_at (ingestion time). For "this week", prefer since=YYYY-MM-DD as ISO date 7 days ago.
• Don't enumerate. Don't return raw tool output verbatim. Summarize, group, synthesize.

Recipes for canonical use cases

Recipe 1 — "Make me a podcast"

Defer to the dedicated podcast-digest skill. It handles configurable length (default 30 min), script generation, and audio rendering via scripts/render-podcast.ts. Don't reimplement the workflow here.

Use this skill (digest-insights) only when the user wants a script but explicitly opts out of audio rendering, or when their question is "what would a podcast cover" without asking for the artifact itself. For everything else — "make me a podcast", "generate this week's episode", "render the digest as audio" — invoke podcast-digest.

Recipe 2 — "What's a trend in development tools worth tracking?"

1. mirror_status.
2. search_items(category='ai_dev', since=<30d>, limit=80) and search_items(category='tech_articles', since=<30d>, limit=40).
3. semantic_search_items(query='new developer tool launch', limit=20) and same for variants like "dev tooling shift" or "IDE trend".
4. Look for repeating patterns across multiple items from different sources — real trends are corroborated, one viral tweet isn't a trend.
5. Pick the single most corroborated trend. Present:
   • The claim (1–2 sentences)
   • 3–5 corroborating items (with URLs)
   • Dissenting items if any
   • Why it's worth tracking now — what's the next-90-day signal to watch?

Recipe 3 — "How are people talking about codebase understanding and agent context?"

This is a conceptual query. Lead with semantic search.

1. mirror_status.
2. semantic_search_items(query='codebase understanding and agent context', limit=15).
3. Run 2–3 variant queries: "long-context code navigation", "agent memory for code", "repo-level LLM tools".
4. get_item on top 4–6 distinct items (de-dup across queries) to pull representative language.
5. Synthesize into 3–5 discernible "conversations" — groups of items that share vocabulary and stance. For each: what's the framing, who is saying it, what are concrete examples.
6. Note what's missing or contested.

Recipe 4 — "Top 10 most useful items this week to inform my agentic work"

1. mirror_status.
2. semantic_search_items(query='agentic work', limit=30) plus variants like "AI agent tools", "autonomous coding agents", "LLM coding workflow".
3. search_items(category='ai_dev', since=<7d>, limit=40).
4. De-dup by id. For each candidate, get_item to check it's substantive, not just a tweet restate.
5. Rank by: (a) relevance to "agentic work", (b) final_score if available (use aggregate_items context to decide which sources tend to be high-quality), (c) novelty vs. what user likely already knows.
6. Return exactly 10, each with: title, 1-sentence why-it's-useful, URL. Finish with an optional "also-rans" list if there are 3–5 close calls.

Guardrails

• Never expose raw item IDs in the final answer except where helpful for follow-up ("If you want to dig into item X with id=…, run get_item"). Prefer URLs.
• If any tool returns 0 items, say so directly ("Nothing matching <query> this week"). Don't invent.
• If the user's question requires data the mirror doesn't carry (e.g. open-web search, twitter real-time), say so and ask whether to proceed without it.
• If mirror_status shows staleMinutes > 90, do NOT silently proceed; surface it and let the user decide.