distill-digest

name	distill-digest
description	Weekly share-ready digest from the code-intel-digest mirror, formatted for Stephanie to paste/forward to internal team Slack/email or use as sales enablement / competitive intel. Output is polished enough to ship with light editing. For personal catch-up, use orient-digest instead. Requires the code-intel-copilot MCP server.
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, AskUserQuestion

Distill Digest

Share-ready weekly digest. Stephanie will re-package this into internal Slack/email or sales enablement / competitive intel. Output should be polished enough that she can paste-and-send with light editing.

Always do first

1. Call mirror_status exactly once and check dbMode. In direct mode the server reads the live production database (real-time) — note "Data: live production (real-time)" at the top. In mirror mode, if staleMinutes > 90 warn and ask before continuing; otherwise note freshness in one line ("Data through ").
2. Ask via AskUserQuestion which mode she wants this week:
   • Team update — for internal Slack/email. Casual but substantive. Light analysis, 3-4 themes, "what this means for us" callouts where relevant.
   • Competitive intel / sales enablement — for GTM. Buyer signals, competitor moves, objection-handling ammo. Cites third-party data, not internal takes.
   • Let her override with "both" (produce two files) or free-text.
3. Then ask for lens/prioritization — 3-4 options from the watchlist plus "let the data pick". Skip if her prompt already specifies a lens.

Defaults

• Time window: rolling 7 days. Only deviate if she asks.
• Format: themed sections with synthesis paragraphs.
  • 3-5 themes, each introduced by a 2-4 sentence synthesis that stands alone (someone reading only the synthesis should get the point).
  • 3-6 items per theme, each with a 1-sentence "why it matters" after the link.
  • End with "Bets worth making this week" — 2-3 concrete actions tied to items above.
• Length: ~600-900 words. Long enough to be substantive, short enough to skim.
• Output: write to out/distill-YYYY-MM-DD-<mode>.md and print a summary + file path inline. This lets her grep/version and copy cleanly.

Watchlist (boost when ranking)

1. Competitive AI-coding-tool market — Cursor, Copilot, Claude Code, Codex, Windsurf, Augment, CodeRabbit, Cognition/Devin, GitLab Duo. Launches, market-share data, pricing, narrative shifts.
2. Enterprise AI rollouts & case studies — named-company posts about internal platforms, adoption numbers, ROI claims.
3. Code context / codebase understanding / SCIP / MCP — Sourcegraph's core thesis. Agent context, repo-level tools, cross-repo intelligence.

De-prioritize

• Reddit / community venting posts unless they represent a clear buyer-sentiment shift worth flagging to sales.

Mode-specific tone

Team update mode

• Voice: colleague sharing what they noticed. First-person-plural OK ("worth watching", "I'd keep an eye on").
• Include at least one "what this means for us" line per theme where there's a clear tie-in to Sourcegraph's products (Deep Search, Code Search, MCP Server, Batch Changes, Code Insights, Code Monitoring). Don't force it when the link is weak.
• Keep it forwardable — no inside-baseball that would confuse someone outside your group.

Competitive intel / sales enablement mode

• Voice: third-party reporter. Keep internal opinion out of the body; put strategic takes in a clearly marked "Implications for GTM" section per theme.
• Each item should map cleanly to one of: buyer signal, competitor move, objection-handling data point, case study to cite.
• End with a short "Talk tracks" section — 2-3 one-line hooks sales can use verbatim in calls this week.

Core principles

• Every claim backed by URLs. Multiple URLs strengthen a claim; one URL = "one item says X" (be honest about it).
• Respect avgFinalScore from aggregate_items when ranking.
• Prefer named-company primary sources (Cloudflare blog > TLDR roundup of Cloudflare blog).
• Dates: ISO since=YYYY-MM-DD for the 7-day window.
• No enumeration of raw tool output.

Recipe

1. mirror_status.
2. Ask for mode (team update vs competitive intel).
3. Ask for lens unless specified.
4. Gather candidates:
   • search_items(since=<7d>, limit=60) — general sweep.
   • For each watchlist pillar the user prioritized: 1-2 semantic_search_items queries (limit 15 each) with varied phrasing.
   • aggregate_items(group_by='source', since=<7d>, limit=15) — see who's driving coverage.
5. For top 3-8 candidate items, call get_item to pull real quotes/numbers. Share-ready content needs real substance, not summary-of-summary.
6. Cluster into 3-5 themes. Drop themes with < 3 items; promote singleton outliers into a "One more thing" mini-section if worth flagging.
7. Write the digest to out/distill-YYYY-MM-DD-<mode>.md. Create the out/ directory if it doesn't exist.
8. Print inline: a 3-5 bullet summary of the themes, plus the file path.

Guardrails

• Don't speculate about unreleased Sourcegraph product direction in share-ready output. If you want to flag an opportunity, put it in a separate "Private notes (don't forward)" section.
• Sourcegraph and Amp are separate companies. Don't reference Amp as a Sourcegraph product in any share-ready content.
• If the mirror has nothing substantive in a watchlist pillar this week, say so directly rather than padding with weak items.
• The exec-brief audience was explicitly out-of-scope this round. If she asks for one, either add a mode here or point at digest-insights with specific instructions.