lab-newsletter

name

lab-newsletter

description

Produce the AICell Lab's recurring AI-generated newsletter — a nightly deep-research digest of news relevant to the lab (AI-for-biology, bioimaging, the lab's own software/papers, collaborators, funding, events). Collects sources locally, writes an on-brand digest post in the `newsletter` category, builds, commits to main, and posts to Slack #general. Use when running the nightly newsletter pipeline (svamp workflow `nightly-newsletter`) or when asked to generate/test a lab newsletter issue.

Lab Newsletter Pipeline

This skill is the reusable procedure the nightly nightly-newsletter svamp workflow runs. It turns "what's new and relevant to the lab" into a short, cited, on-brand post on aicell.io and a Slack notice.

The website operating conventions (front matter, build, deploy) are in the sibling aicell-website skill — read it for the exact post front matter and the build/commit/deploy flow.

Modes

• prod (default, nightly): full deep research, ~3–8 sources, one digest post.
• test: a fast smoke run that exercises the whole pipeline end-to-end with a minimal real post (1–2 quick sources). Use it to verify the workflow plumbing (scheduler → spawn → produce → commit) without a long research run.

Pipeline (run top to bottom)

1. Research

The digest is grounded in what the lab does, then extends toward where the field (and the lab's strategy) is heading. Cover two buckets and mix them:

Bucket A — the horizon (the future the lab is building toward):

• The virtual cell and AI models of cells; automated / autonomous scientific discovery ("AI scientists", self-driving labs); AI agents for science and automation.
• Advanced AI tools & development — notable new models, frameworks, agent tooling, and capabilities (not just bio-specific) that the lab could build on.
• State-of-the-art cell biology & omics — single-cell / spatial / multi-omics, imaging, foundation models for biology.
• The bigger picture: what AI + automation could mean for science and humanity — problems we could potentially solve with AI agents and automation.
• Treat this as a strategy radar: when something could reshape how the lab works or what it should pursue, say so in one line.

Bucket B — the lab's own focus (keep it concrete and current):

• Derive the focus list from the site: read content/project/*/index.md titles+summaries and the content/authors/aicell-lab interests — the live priorities (today: ImJoy, BioImage Model Zoo, BioEngine, Hypha, Agent-Lens, REEF, the self-driving microscope, human/whole-cell modeling, Safe Colab, the Research Navigator).
• Track news touching those, the lab's own releases/papers, collaborators, and SciLifeLab / KTH / DDLS / WASP / Alpha Cell announcements, funding, and events.

Re-deriving Bucket B from content/project/* each run keeps the digest following the lab's focus as projects evolve.

Where to look: gather from the curated source set in sources.md (the lab's "information channels" — preprint servers, journals, AI-lab blogs, virtual-cell/automated-discovery initiatives, the lab's own GitHub/sites, funders, and events). Keep that file pruned and current as strategy shifts.

Sweep X (Twitter) via the lab-x CLI — a primary signal source:

• scripts/lab-x.py monitor --since-hours 24 --min-likes 30 → recent high-signal posts from the watch-list (x-accounts.md: AI labs, AI-for-bio researchers, bioimaging, aggregators). Fold the genuinely relevant items into the digest, with their links.
• scripts/lab-x.py search "<topic>" --max 20 for topic sweeps (e.g. "virtual cell").
• Living watch-list: also run scripts/lab-x.py discover and add any clearly relevant new accounts to x-accounts.md (verify with lab-x.py info <h> first). Every run, ask: "is there a new person/channel worth following?"

Use WebSearch/WebFetch (or the deep-research skill in prod). Prefer primary sources (papers, official blogs, release notes). Never fabricate facts, numbers, or quotes. Every claim must trace to a source you actually fetched.

Tag relevant lab members. When an item strongly matches someone's work, add a short friendly @-mention so they see it — using the Slack IDs + keyword map in .claude/skills/lab-newsletter/lab-members.local.md (gitignored). Tag sparingly (strong matches only, ≤1–2 people per item; skip for routine lab-own news). The tag is for the Slack announcement (<@SLACK_ID>); on the website post, name the person in plain text instead of a Slack mention.

2. Collect sources locally

Save the raw source material into a dated folder so the digest is reproducible:

data/newsletter/<YYYY-MM-DD>/
  sources.md        # list: title — url — 1-line why-relevant — fetched-at
  *.txt / *.pdf     # optional saved copies of key sources

data/newsletter/ is git-ignored for the bulky raw captures (see .gitignore); keep sources.md small and commit it with the post if useful for provenance.

3. Filter & dedup

• Keep only items genuinely relevant to the lab. Quality over quantity — one well-curated digest beats many thin items.
• Drop anything already covered: check existing content/post/ slugs and recent newsletter posts.

4. Write the digest post

Create content/post/newsletter-<YYYY-MM-DD>/index.md:

---
title: "Lab Newsletter — <Month D, YYYY>: <Short Descriptive Headline>"
summary: "Today in AI for life science: <one-line teaser>."
date: '<NOW>'        # the ACTUAL run time in UTC, e.g. `date -u +%Y-%m-%dT%H:%M:%SZ`
lastmod: '<NOW>'
draft: false
featured: false
authors:
  - Happy Agent
tags:
  - newsletter
categories:
  - newsletter          # the dedicated category for these AI-generated digests
---

Descriptive title: keep the recognizable Lab Newsletter — <Month D, YYYY> prefix, then append : <Short Descriptive Headline> that captures the day's lead items (e.g. "Thinking Microscopes & Self-Imagining Cells"). This makes the post list and search results read like real stories instead of bare dates. Keep the headline punchy (≤ ~7 words). ⚠️ Don't future-date the post. Use the real current UTC time for date (date -u +%Y-%m-%dT%H:%M:%SZ), not a fixed hour like 06:00Z. Hugo (and CI, which builds without --buildFuture) silently skips future-dated content, so a post stamped 06:00Z won't render or deploy if the nightly fires earlier (the 0 5 * * * cron is 05:00 local, ≈ 03:00 UTC). Body: a short intro line, then 3–6 bulleted items. Each item = a bold headline, 2–4 sentences of plain-language digest, and an inline link to the source. House voice: warm, first-person-plural, lightly linked. End with a one-line "Why it matters for the lab." Keep it skimmable. No multi-MB images.

Give it a cover. Generate an on-brand featured image with the lab-image CLI (consistent AICell Lab signature style):

scripts/lab-image.py generate --prompt "<one line capturing today's digest>" \
  --out content/post/newsletter-<date>/index.jpg   # or featured.jpg in the folder

Hugo auto-processes featured.jpg. This makes each digest rich and recognizable.

5. Build & commit

Get the correct env first. Automations run as fresh agents/shells without the build toolchain on PATH, and in Claude Code each Bash call is a fresh shell (env does not persist between calls) — so source scripts/setup-build-env.sh in the SAME command as the build. That script puts Hugo extended 0.101.0 + Go at stable paths (~/.cache/aicell-build, not ephemeral /tmp), sets the cache vars, and loads secrets from ~/.svamp/aicell-newsletter.env.

# one shell: source env, then build with the resolved $HUGO_BIN (NOT bare `hugo`)
. scripts/setup-build-env.sh && "$HUGO_BIN" --gc --minify -b https://aicell.io   # must exit 0
git add content/post/newsletter-<date> data/newsletter/<date>/sources.md
git commit -m "Newsletter: <Month D, YYYY> digest"
git push origin main                          # nightly is pre-authorized routine content

Do not run hugo mod get (it upgrades the theme to v5.9.0 and breaks Hugo 0.101.0; the theme is pinned in go.mod).

6. Announce to Slack #ai-digest (daily) — weekly highlights go to #general

Announce the published digest to the lab using the lab-slack CLI, which formats a card linking to the post:

scripts/lab-slack.py announce --channel ai-digest --post content/post/newsletter-<date>/index.md \
  --note "<@U06DE2TPG1W> the BioEngine item might interest you · <@U082CS9FMNE> lab-automation news below"   # → #ai-digest

# Build the --note from lab-members.local.md: @-mention members whose keywords strongly
# match items in today's digest (sparingly; skip if nothing is a strong match → omit --note).

It loads the bot token from ~/.svamp/lab-slack.env automatically. If Slack isn't configured on this machine the command errors — that's non-fatal for the run: still publish the website post and note Slack was skipped. Preview without posting via --dry-run.

Slack setup

Slack is handled by the lab-slack skill (scripts/lab-slack.py), which posts as the lab bot using tokens in ~/.svamp/lab-slack.env (SLACK_BOT_TOKEN, SLACK_USER_TOKEN) — never committed. Use scripts/lab-slack.py announce --post <path> to announce a post, or scripts/lab-slack.py dm --to <email> --text "…" to DM someone. (The legacy scripts/post-to-slack.sh webhook helper still works as a fallback.)

Scheduling (how the nightly run is wired)

The nightly digest runs as a svamp workflow on a cron schedule (workflows are the single automation surface as of svamp-cli 0.2.176+):

svamp workflow add nightly-newsletter --on schedule --cron "0 5 * * *" \
  --run "svamp session send <owner-session-id> '<the pipeline prompt>'"

The generated yaml (.svamp/workflows/nightly-newsletter.yaml) is GitHub-Actions style (on: schedule: - cron, jobs: run: steps: - run). The daemon's ~20s cron tick fires it. Inspect with svamp workflow list / svamp workflow run <name> (manual).

History & gotchas (verified 2026-06-23/24):

• Workflows now fire on the daemon (cron scheduler shipped in 0.2.176). The old svamp trigger/routine system was removed entirely (#0048) — those commands now error. If you find references to triggers, they're stale; use workflows.
• Never use svamp session send . in an automation: . resolves via SVAMP_SESSION_ID, absent in the daemon/workflow shell → No session found matching: .. Target a real session id, or spawn a fresh agent.
• The step enqueues to the owner session's message queue (drained when that session is idle), so <owner-session-id> must be a live session that goes idle near the schedule time. More robust alternative: have the step spawn a fresh one-shot agent — svamp session spawn claude -d /Users/weio/workspace/aicell.io --message "<pipeline prompt>" — which produces and commits independently of any live session. Use this for on-demand runs and manual tests too.

Guardrails

• Never invent facts or quotes; always link the real source you fetched.
• Stay on-topic for the lab's mission; respect confidentiality (no unpublished internal work).
• Small scoped commit; reuse the newsletter category and existing tags.
• The site footer already discloses AI-assisted content.