The agent that runs our weekly newsletter

We send a weekly newsletter to ~3,500 subscribers. It's part of our content marketing, part of our customer communication, part of our brand voice maintenance.

In the early days, drafting it took me (Bhoomika) 3-5 hours every week. Now an Avery NXR agent does most of it, and I edit for ~30 minutes.

Here's the agent in detail. It's a good example of how custom agents work beyond the pre-loaded templates.

What the newsletter needs to do

Every Friday, our newsletter goes out. The structure:

→ Intro section (what I've been thinking about this week) → 3-5 product updates / customer stories → 3-5 industry news items with our take → 1-2 deep dives (linked to longer blog posts) → A specific CTA → Personalization (greeting, P.S. occasionally tailored)

The content needs to be on-brand, useful, and not feel AI-generated even though significant portions are AI-assisted.

The agent architecture

The agent is named (internally) weekly-newsletter-drafter. It runs every Thursday at 6 PM. Its workflow:

Step 1 — Gather inputs. → Pull industry news from Anna (our news digest agent) for the week → Pull customer stories from Carlos (our pipeline digest agent) → Pull product updates from our GitHub releases + Linear closed issues → Pull recent blog posts we've published

This is a sub-agent chain. The newsletter agent doesn't gather raw data — it asks other agents that already have curated data.

Step 2 — Draft sections. → The intro section: read my Slack messages and Linear comments from the week, summarize what I've been thinking about → Product updates: rank the GitHub releases by user-visible impact, draft 1-2 sentence summaries → News items: filter Anna's news to items relevant to our customers, draft 1-2 sentence takes → Deep dives: pick the 1-2 blog posts most relevant to current themes, draft promotional intros

Step 3 — Compose. → Stitch all the sections together → Add transitions → Match the newsletter template structure → Include subject line options

Step 4 — Deliver. → Save as draft in Substack → Send a Slack message to me with the draft + reasoning notes → Log to audit ledger

The full agent takes about 4 minutes to run.

What's in the prompt

The agent's main composition prompt includes:

→ Voice and tone guidance (5 paragraphs of brand voice direction) → Examples of past newsletters we considered "on-brand" → Examples of past newsletters we considered "off-brand" (with what was wrong) → Structural template → Word count guidance per section → Specific phrases to avoid (LLM cliches like "in today's fast-paced world") → Instructions to flag uncertainty (don't make up customer names, don't fabricate data points)

The prompt is about 1,200 words. We've refined it over ~12 newsletter cycles. Each cycle, when I edit the draft, I add what I had to fix to the prompt as a future-prevention guideline.

What I do in the 30 minutes of editing

The draft is usually 80-85% of the way there. The 15-20% I add:

→ Voice refinement. The agent's voice is close to mine but not exactly mine. I rewrite 1-2 sentences per section in a way that's more obviously me.

→ Specific anecdotes. The agent draws from my Slack messages, but it can't add color commentary the way I would. I add 1-2 anecdotes or jokes that personalize the intro.

→ Strategic emphasis. The agent ranks items by general impact, but I sometimes know "this specific update matters more this week because of X." I reorder accordingly.

→ Subject line picking. The agent generates 3-5 subject line options. I pick or tweak.

→ Final review. Read through for any factual errors, any unintended-sounding phrasing, any links that need checking.

This 30-minute edit is qualitatively different from drafting from scratch. I'm reviewing and refining, not writing. The cognitive load is much lower.

What this saves

Pre-agent: 3-5 hours/week × 52 weeks = 156-260 hours/year on newsletter drafting.

Post-agent: 30 min/week × 52 weeks = 26 hours/year on newsletter editing.

Time saved: 130-234 hours/year. At my opportunity cost (~$80/hour effective), that's $10,400-$18,720/year of value.

The cost: $29/month Avery NXR Pro subscription for the agent + a few connectors. About $348/year. Plus my time to refine the prompt over ~12 cycles (~6 hours total).

ROI on the newsletter agent specifically: 30x to 50x on time investment.

What the agent does NOT do

I want to be clear about what the agent doesn't handle:

→ Strategic content choices. The agent ranks based on user-visible impact, but if I have a strategic reason to emphasize something specific (a launch we're prepping, a customer story we want to amplify), I tell it explicitly.

→ Original writing. The intro section is built from my actual Slack messages. The agent isn't inventing thoughts I never had — it's organizing thoughts I had into draft form.

→ Customer privacy. I review every customer mention. The agent flags any customer references for explicit approval.

→ Brand evolution. When our voice shifts (we've evolved over time), I update the prompt. The agent doesn't auto-evolve.

→ Crisis response. When something specific is happening (an incident, a launch, a controversy), I write the relevant section manually. The agent is for steady-state weeks.

What I'd tell others building newsletter agents

If you're going to build something similar for your own newsletter:

→ Start with a strong existing voice. The agent will mirror what's in the prompt. If your voice isn't clear yet, the agent will sound generic. Establish voice first, then automate.

→ Use real examples in the prompt. Past newsletters that landed well are better training than abstract voice guidance.

→ Run for 6+ cycles before judging. Early drafts will need substantial editing. The prompt improves with each cycle. By cycle 6, the draft is reliable.

→ Keep the editing step. Don't try to remove the human in the loop. The editing pass is where voice + judgment + strategic emphasis happen. Without it, the newsletter sounds AI-generated.

→ Document what you change in editing. When you have to fix something, add a guideline to the prompt so the agent does it right next time.

→ Be honest with readers. Our newsletter doesn't pretend to be 100% hand-written. We've mentioned the AI assistance. Readers don't seem to mind because the quality is good and the voice is consistent.

The bigger pattern

The newsletter agent is one example of a custom agent built for a specific recurring workflow. We have others:

→ Lead qualification agent (processes inbound contact forms) → Documentation reviewer (flags doc updates needed when product changes) → Release notes drafter (generates release notes from merged PRs) → Customer health digest (flags accounts showing churn risk)

Each one took 2-6 hours to build. Each one saves 5-20 hours/week of recurring work.

The 7 pre-loaded templates (Anna, Sophia, Marcus, Priya, Carlos, Yuki, Liam) are good starting points. Custom agents are where the platform gets most personal.

What this means for you

If you have a recurring content task that takes hours every week — newsletters, status reports, weekly updates, daily summaries — there's almost certainly a custom agent that can absorb most of it.

The way to find out: try building one. Pick the task that frustrates you most. Use Avery NXR's blank agent template. Iterate for a few weeks.

The 30x-50x ROI on my newsletter agent isn't unusual for high-effort recurring tasks. It's actually the typical pattern when you build the right custom agent.

→ avery.software — Free Desktop tier. Build the agent that absorbs your weekly hours.