| Monday · May 18, 2026 |
Issue № 005 |
15 min read |
|
|
The FindSkill Weekly Brief
The Skill
Just for FindSkill Pro members. The AI news that actually matters for your work.
|
|
|
|
|
Mia
AI Learning Editor · FindSkill.ai
|
Welcome to Issue 005.
I'm Mia. Every Monday I land in your inbox and tell you what the week's AI news actually means for the way you work — without the jargon, the hype, or the "10 things you MUST know" energy. Pro members only.
Three Anthropic launches stacked on top of each other this week, and they only look like three stories until you put them in a row. Multiagent Orchestration on May 6. Agent View inside Claude Code on May 12. Claude for Small Business on May 13. Read those three sentences again — the direction is the same one. AI is moving from "one thing you talk to in a tab" to "many things you run from a dashboard." The unit just changed.
A quick reply note. Last week I asked what you found inside ChatGPT's memory audit. Twenty-something of you wrote back. The most common surprise — "I had no idea it was storing X" — was almost always something the model had inferred from a casual aside in March. Memory drift is real and it goes silent. I'm sitting on the best examples for Issue 007 (Pro guide drop). Keep them coming.
— Mia
|
01 |
This Week in AI
Three stories worth your attention
|
Tool Launch
Claude Code shipped a dashboard for running many agents at once. The role just changed.
On May 11–12, Anthropic released Claude Code v2.1.139 with Agent View and the `/goal` command. Agent View is a single dashboard inside the CLI that lets a developer run, monitor, and reply to many sessions in parallel — eight, fifteen, twenty if your wallet can handle it. The `/goal` command pairs with it: you state an objective, the agent runs hands-free toward it and pings you when it needs approval. The release also formalized `--bg` and `/bg` for background tasks. None of these are features in the usual sense. They're the cockpit. Twelve months ago you ran one Claude session in one terminal and watched it work. As of last Tuesday you run a dozen and a dashboard tells you which one is stuck, which is done, and which just asked you a question.
|
What it means for you
Even if you don't write code, this matters. The pattern shipping in Claude Code this week is the pattern shipping in everything else next quarter. Pick one tool you actually use — your AI inside Notion, the assistant inside your CRM, the agent in your help desk — and look for the same shape: multiple agents running in parallel, a dashboard for oversight. The job stopped being "type a good prompt." It started being "set the right objectives and watch the cockpit."
|
Tool Launch
Anthropic shipped Claude for Small Business — 15 prebuilt workflows that already plug into QuickBooks, Stripe, HubSpot, Square, and Docusign.
On May 13, Anthropic launched Claude for Small Business as a Cowork plugin — 15 ready-to-run agentic workflows plus 15 skills, covering month-end close, AR chasing, invoice reconciliation, payroll prep, campaign management, contract redlines, and onboarding flows. Native connectors to QuickBooks, PayPal, HubSpot, Canva, Docusign, Stripe, Square, Webflow, Microsoft 365, and Google Workspace. Forbes covered it. TechCrunch covered it. Linas Beliūnas posted breathlessly on LinkedIn. What nobody yet has done is rank the 15 workflows by hours saved versus hours pretended. We tested all of them — the killers vs. the theater post shipped Friday and the seven that actually save time aren't the ones the launch reel led with.
|
What it means for you
If you run anything resembling a small business — a five-person agency, a solo practice, a 30-person company — the Cowork plugin is worth fifteen minutes this week. Not because every workflow is great (about half are). Because connecting it once tells you which parts of your stack the AI now reads natively, and that changes what you can ask of it. Skip the marketing demo. Open it, connect one tool, run one workflow, judge from the output.
|
Trend Spotlight
Three Anthropic launches, one direction — orchestration is the new layer.
On May 7, Anthropic shipped Multiagent Orchestration inside Managed Agents — a lead agent that decomposes your task and delegates to specialist sub-agents. Harvey (legal AI, ~$200M ARR) publicly reported a 6× completion-rate improvement on multi-step legal workflows after adopting it. Wisedocs cut review time 50%. Netflix is now parallelizing build-log analysis on this stack. Five days later, Agent View shipped — the developer's cockpit for orchestration. The day after, Claude for Small Business shipped — orchestration productized for the people who'd never write the lead agent themselves. The pattern is the same every time a new layer arrives. The dev surface shows up first (orchestration in code), the dashboard arrives next (Agent View), then someone wraps it for non-devs (Small Business). Three weeks. One direction.
|
What it means for you
The 2026 winner isn't the person with the best prompt. It's the person who internalized this layer first. The cognitive shift isn't technical — it's role-based. You're not the operator anymore. You're the editor-in-chief of a small AI newsroom. Different muscle. Easier to learn now while everyone's still typing single prompts than later when you're behind.
---
|
|
02 |
Term of the Week
The one concept to understand this week
|
| |
|
Term №005
Multiagent orchestration
< multiagent orchestration >
A pattern where one "lead" AI agent takes your task, splits it into specialist sub-tasks, hands each to a focused sub-agent, then synthesizes the results into a single output. The lead never does the specialist work — it delegates and edits.
|
Think of it like this →
A general contractor on a kitchen remodel. You hire one person. They don't swing a hammer. They hire a plumber for the sink, an electrician for the lights, a carpenter for the cabinets, a tiler for the floor, and a painter at the end. Each specialist does one thing well. The GC's job is sequencing, judgment, and quality control — not labor. Multiagent orchestration is software-flavored general contracting.
|
|
⚠ Common misconception
"Multiagent just means running multiple Claudes in parallel." No. Parallel copies don't talk to each other or hand work off — they just race. Orchestration has a lead that decomposes the task (this is the planning step that single-prompt AI is bad at), routes work to specialists (each with a narrower job and a tighter prompt), and synthesizes (which is where most one-shot AI loses the plot). The orchestration is the value. Without it, multiagent is just a more expensive single-agent.
|
Where you'll hear it: Anthropic's Managed Agents (shipped May 7 — the canonical commercial implementation). LangGraph (open-source, where the dev community builds these stacks today). CrewAI (the friendlier role-based framework — "you are the Researcher, you are the Writer"). AutoGen (Microsoft's contender). Magentic-One (Microsoft's research preview). Anthropic Cookbook has the reference recipe — half a page of code, surprisingly readable. And every Claude for Small Business workflow you ran in Story 2 is an orchestrated stack underneath the cute card UI.
Claude for Small Business — orchestration in practice →
|
|
03 |
Deep Insight
What Harvey did to multiply their AI 6× — and what to steal from it
|
Anthropic published a number this month that's been bouncing around AI Twitter without context. "Harvey saw a 6× completion-rate improvement." The number is real. The reason it happened is more interesting than the number — and the lesson translates down to anyone running one project, not just $200M-ARR companies.
Harvey is a legal AI startup founded by a former DLA Piper attorney, valued at $5B, sold into roughly 250 of the world's biggest law firms. They were the canonical "AI for law" story in 2024. Then a quiet thing happened in early 2026: their multi-step legal research workflows kept hitting a ceiling. The model wasn't the problem. The architecture was. Here's what they did about it — and the part that scales down to your work.
| | SUBJECT Harvey reported a 6× improvement in completion rate on multi-step legal workflows after migrating from single-agent Claude sessions to Anthropic's Multiagent Orchestration in early May 2026 — published in Anthropic's customer case studies and covered by Forbes on May 11. For context, "completion rate" here means the percentage of multi-step research tasks (e.g., "find every precedent supporting X across these 14 jurisdictions, then draft a memo") that the system finishes correctly without a human picking up the loose ends. The number went from roughly 22% to roughly 78%. |
What's worth stealing isn't the result — it's the four specific decisions that produced it. Each one inverts an instinct most people start with.
| i. | They didn't chase a bigger model. They decomposed the workflow. Harvey's team had been on Opus 4.6 and looking at Opus 4.7. Instead, they kept the same Opus model and broke the work apart. Old setup: one Claude session asked to "research, analyze, draft." New setup: one lead agent, three specialist agents (research, analysis, drafting), one critic agent. The model didn't get smarter. The job got smaller. That's where most of the 6× came from. | | ii. | They separated reading from writing. Legal research and legal drafting use different cognitive modes — one is exhaustive, the other is selective. When the same agent did both, the model bled the modes into each other and produced thorough-sounding sludge. Separating them let each prompt be specific to its mode. Research agent: "find everything, miss nothing." Drafting agent: "use only what's been handed to you, write tight." Different jobs, different prompts, different evaluation criteria. | | iii. | They added a critic before synthesis. The lead agent doesn't just stitch sub-agent outputs together. It first runs a critic that grades the sub-agent work against a rubric (the same rubric Anthropic's Outcomes feature exposes). Anything that fails goes back to the originating sub-agent for revision. Most AI failures aren't "the model can't do this" — they're "the model didn't notice it missed something." The critic notices. | | iv. | They invested in evals more than prompts. This is the line buried at the bottom of Harvey's case study and it's the most important one. The team spent more eng time on measuring whether the orchestrated stack was actually better than on tuning any individual prompt. A working rubric that can grade output objectively is worth more than a clever prompt that can't be evaluated. Build the scorecard first; the prompts get sharper because you can finally see them lose. |
What changed, in numbers Input · Before (single-agent) 22% completion rate on multi-step research workflows. Average researcher reviewed AI output for ~45 minutes per matter and had to re-run portions roughly half the time. "Faster than nothing, slower than a good associate. | | | Output · After (orchestrated) 78% completion rate on the same workflow class. Researcher review dropped to ~12 minutes and rejection rate fell to about 15%. Customer renewals up. Per-matter cost roughly flat (more API calls, fewer human-hours). |
The trap here is reading "Harvey" and thinking "that's not me." The architecture is identical at any scale. The two-agent version of this — a writer and a critic running on your hardest weekly deliverable — is what Section 04 walks through. Do that this week. The skill of running an AI crew is built one critic at a time.
|
04 |
The Workflow
One way to use AI at your job this week
|
| |
Run a writer-and-critic loop on the one weekly thing you secretly hate writing.
Pick the deliverable you produce every week that you usually rewrite three times — a client proposal, a board update, a weekly team email, a roadmap doc, a candidate-rejection note. Open two chats in your AI of choice (ChatGPT, Claude, Gemini — any of them). Chat 1 = Writer. Paste: "You are a draft writer for [deliverable type]. Your job is to produce a confident first draft from the source material I provide. Do not soften. Do not hedge. Do not summarize the brief back to me — write the deliverable." Chat 2 = Critic. Paste a five-line rubric: clarity, specificity, accuracy, tone match, length. Below the rubric: "You are a senior editor. Your job is to score the draft against the rubric, identify the two weakest sentences, and rewrite them. Do not praise. Do not rewrite the whole thing — only the two sentences you flagged." Send your source material to Chat 1, paste the draft into Chat 2, paste the critic's feedback back into Chat 1 with "incorporate this and produce v2." Stop after v2.
Why it works: The fundamental failure of one-shot AI writing is asking the model to draft and judge its own draft. Models are bad at grading themselves — they reliably score their own output higher than it deserves. Splitting the roles forces the second pass to start from criticism, not from approval. Two passes also matters: the first revision usually fixes the obvious; the second tightens. More than two is diminishing returns because the model starts editing for style, not substance. The writer-critic loop is the simplest possible orchestration. It's what Harvey does at industrial scale in Story 3. You can do it before lunch.
|
Do this week
Pick one deliverable. Run the loop tomorrow morning, once. Compare v2 against the version of this deliverable you'd normally send. If v2 is better, you've just built the smallest possible AI crew — and the next deliverable type is your call.
|
|
|
05 |
The Side Play
One way to make money with AI this week
|
| |
| ◆ Income Idea · Play №005 |
Sell a working two-agent crew for one specific role.
Pick a role you've either done or watched closely — recruiters, account managers, financial advisors, junior PMs, real estate agents, freelance copywriters, executive assistants. Build a single Notion page or PDF called something like "Your first AI crew — a writer-and-critic loop for [role], in 30 minutes." Inside, put four things: (1) the two-agent prompt pair, customized for that role's specific deliverables (a recruiter's intake email, an AM's renewal pitch, an EA's calendar-conflict reply); (2) the role-specific rubric the critic uses — between 5 and 8 lines, concrete not generic ("clarity" is generic; "names the deadline" is concrete); (3) a handoff message template — the exact words you paste between agents — because most readers will fumble this part once and quit; (4) five fully worked examples — source brief → writer draft → critic feedback → final output — so a buyer can verify the loop works before customizing. Price at $39–$69 on Gumroad or Lemon Squeezy. Position it plain: "You don't need a smarter AI. You need a second AI that disagrees with the first one."
Why it works: Multiagent is the buzzword of May 2026. Every news outlet covered it, no one has wrapped it into a thing a non-developer can use this week. The writer-critic loop is the lowest-friction on-ramp — it's two prompts and a rubric. The customer pays you because you've already done the role-specific work of figuring out what "good" looks like for their job. Realistic revenue: $400–$2,500/month with no audience; $2K–$6K/month if you DM it to your existing LinkedIn network in that role. Not a unicorn. Honest numbers — and the timing window is roughly two months before this is a $7 commodity.
|
Do this week
Pick the role today. Write the two prompts and the rubric this evening. The five worked examples can come after the first sale — list it first, polish it second.
|
|
|
06 |
The Stack
Three tools I'm actually testing this week
|
01 |
Claude Agent View
Multi-agent dashboard inside Claude Code · Included in Claude Code (Pro/Max plans)
The reason this is worth knowing even if you don't code: it's the first consumer-grade cockpit anyone has shipped for parallel AI work. Spend ten minutes watching a developer demo of Agent View — there are good ones from Anthropic's blog post and a couple of Reddit threads — and you'll see the shape every other AI surface is going to copy in the next two quarters. The lesson isn't the UI. It's that the UI exists at all. We wrote the honest tmux-vs-Agent-View comparison on Saturday — three scenarios where Agent View wins, three where tmux still does.
|
02 |
HubSpot AEO Sensor
Free public dashboard tracking AI search citations · Free
Brand new — HubSpot quietly launched it May 14 and it tracks daily volatility, citation, and referral patterns across ChatGPT, Gemini, and Perplexity at industry level. The reason it belongs in this issue's stack: when AI agents start writing your work-related stuff, AI agents start reading your work-related stuff. If anything you publish (a personal blog, a company page, a product page) feeds an industry that's volatile in the AEO Sensor, you should know which way the wind shifted this week. Free, takes five minutes, useful even if you ignore it for a month and then check back. The 10-minute walkthrough is up.
|
03 |
n8n
Open-source visual workflow builder with AI nodes · Free (self-host) or $20/mo cloud
The non-developer's entry into orchestration. n8n has been around for years as a Zapier alternative, but the AI nodes that shipped over the last two quarters changed what it is. You can wire up a five-step crew visually — "incoming email → classify with Claude → draft reply with GPT-5.5 → critique with another Claude call → save to Notion" — without writing code. I built a real customer-onboarding pipeline in it last week and it took about ninety minutes. It's the friendliest version of "I want to orchestrate things" that exists for someone who's not going to learn LangGraph this month. Self-host it if you care about your data going through anyone else's servers; the cloud plan is fine for tests. ---
|
|
07 |
Inside FindSkill
What's new for members this week
|
| New |
Claude for Small Business (English course)
The practitioner companion to Story 2. Connects the Cowork plugin, walks through five of the 15 workflows on real example data, and shows which connectors are worth wiring first. Start the course →
|
|
| New |
Claude for Legal (English course)
The smaller-firm version of what Harvey built. Three-agent stack for a solo or small-firm attorney: research, drafting, critic. Won't replace Harvey for AmLaw 100 work. Will replace half a paralegal's hours for an immigration solo. Start the course →
|
|
| New |
Three European delegation courses shipped
Agentes IA para Autónomos (Spanish) for solo professionals building their first agent stack. KI-Agenten im Mittelstand (German) for the German Mittelstand, with governance scaffolding for the regulated reality. Mistral AI pour PME (French) for French SMBs that want the sovereign-AI option. Three different markets, same lens. See all courses →
|
|
|
|
If you build one two-agent loop this week, which weekly deliverable do you try it on first?
Reply with the deliverable you picked and I'll send you back the role-specific rubric I'd write for it. Worst case, it becomes a section in Issue 007's Pro guide.
↵ Hit reply
|
|
The Skill · by FindSkill.ai
|
|