EPISODE 03 · AI tools & workflows · 2026.05.17

Tools aren't productivity.
Workflows are.

This is a talk about AI tools.
More accurately — it's a talk about not being a "tool collector."

AGENT 101 SERIES · EP 03 · NO-BULLSHIT EDITION
§ 01 · Sort it first

"Tool" and "workflow" aren't even on the same level

How many AI apps are on your phone? How many subscriptions are you paying for? Now answer one question: last week, which one actually finished a task for you? This is where most people are stuck — mistaking "owning a tool" for "having a capability."

  • CONCEPT · 01 · TOOL

    A tool = a single-point capability

    Claude writes, Perplexity searches, Midjourney paints — each is a knife that does one thing. But a knife doesn't cook dinner. A tool is the raw material of capability, not capability itself. Swapping in a sharper knife won't make the meal appear on its own.

  • CONCEPT · 02 · WORKFLOW

    A workflow = capabilities wired into a chain that runs itself

    Trigger → a few steps → one output → review a week later. That's where the leverage is. The same Claude, dropped into "every day at 8:00, read email → classify → draft replies → push to phone," stops being "a tool that chats" and becomes "a process that works for you daily."

  • CONCEPT · 03 · COLLECTING

    Collecting = thinking "installed" equals "able"

    This is most people's real state: 50 tools bookmarked, each tried once, none in their actual daily work. Collecting produces the feeling of "I'm using AI" — the same thing the "1000-prompt bundle" from EP01 sells: anxiety relief, not results.

§ 02 · Audit your stack

"I found another killer app" is an empty phrase

This sentence shows up ten thousand times a day — on X, in newsletters, in every "AI tools weekly." It sounds like progress, but it fails the three simplest checks. If it fails, you're not getting stronger. You're chasing novelty.

  • TOOL TEST · 01

    When did it last actually finish a task for you?

    Not "I tried it and it was impressive" — "it finished a task that would've taken me an hour." If you can't name a specific moment and a specific output, it's still a demo to you, not a tool.

  • TOOL TEST · 02

    Can you name its exact position in one of your workflows?

    "It's the transcription step — step 3 — in my weekly-report pipeline." Say that, and it has a position. A tool with no position is floating, and you'll forget it. A tool's value isn't how strong it is — it's where it's wired in.

  • TOOL TEST · 03

    If you uninstalled it, would your output get worse?

    The brutal one. If removing it changes nothing, it never entered your production. The subscription you pay buys the comfort of "I'm using AI," not output. A tool that hurts to uninstall is a real tool.

See the difference? The first has no finish line — there's always a "better one" next week. The second has a finish line, an output, and a failure signal. Every tool we cover today matters only when it's "wired into a flow." If it can't be, skip it.

§ 03 · After sorting

Tool collector vs workflow builder — two completely different people

Two people pay for the same pile of AI subscriptions; a year later the gap between them is an order of magnitude. The difference isn't which tool they used — it's whether their head runs "tool thinking" or "workflow thinking." This is the only contrast you need to remember.

TOOL COLLECTOR · most people around you

Chases tools, not output

Tries a new tool every week, bookmarks full, workflows empty.

  • Chases "this week's strongest AI tool"
  • Installs a pile, each used exactly once
  • Loves specs, reviews, and beta waitlists
  • Output a year later is unchanged
WORKFLOW BUILDER · who you're becoming

Chases output, not tools

Wires one pipeline a month, but each runs on real work.

  • Builds only 2–3 new workflows a quarter
  • Each one wired into a real daily task
  • Asks one thing: "which step did it remove?"
  • Output is verifiable, shippable on a screen
§ 04 · What a workflow looks like

Four steps: Trigger → Steps → Output → Review

Every workflow that "runs itself" is just these four steps turning. Remember them and you can tell at a glance whether a task should be a workflow — and why it's failing: usually a missing Trigger (step 1) or a missing Review (step 4).

WORKFLOW LOOP TRIGGER STEPS OUTPUT REVIEW
  • 01

    TriggerTRIGGER

    A fixed starting point: every day at 8:00, every Monday, a new email arriving, a form being submitted. A "workflow" with no trigger isn't a workflow — it's "I use it when I remember to," which dies eventually.

  • 02

    StepsSTEPS

    Break the task into a few repeatable steps, each assigned a tool. The more specific the better — "read → classify → draft" beats "handle this for me" a hundred times over.

  • 03

    OutputOUTPUT

    Something you can put on a screen and hand off: a draft, a list, a ready-to-send email. The output must be concrete and verifiable — never "it feels more efficient."

  • 04

    ReviewREVIEW

    A week later, look at the failure signal: still doing it by hand? Output nobody uses? Some step is wrong — go back and tune. A workflow with no review quietly decays into yet another "installed but unused" tool.

§ 05 · The tool map

Sorted by job-to-be-done, not by hype

Every tool ranking out there sorts by "newest and hottest" — useless to you. This map sorts by "the job you need done" into five categories. Each names a few tools that hold up, plus one minimal runnable workflow. Tools change, jobs don't; remember the job, and any tool is swappable.

  • CATEGORY · 01 · THINK & WRITE blank page → a first draft

    Scratch paper for your brain: think the unclear things clear

    Contrarian take: this layer isn't short on tools — it's short on treating it as a "thinking partner," not an "answer machine." The real use isn't "write this for me," it's "help me question the fuzzy thing until it's clear" — you judge, it expands.

    WORKFLOWscattered thoughts → one piece
    01DUMPpour every scattered point into a Claude Project
    02CLARIFYhave it ask you the 5 sharpest questions · you answer
    03DRAFTfrom the Q&A · produce a structured first draft
    04CUTyou only subtract · trim to what you actually mean
    SETUP10 min to build the Project · ROIblank-page anxiety gone · a piece 2h → 40min
    Claude ChatGPT NotebookLM Claude Projects
  • CATEGORY · 02 · SEARCH & RESEARCH an afternoon → 30 minutes

    Cited research, not "ask one, get one"

    Contrarian take: using Perplexity as "a smarter search box" wastes it. Its value is "deep research" mode — give it a question, it runs 30 pages itself, reads, compares, outputs a cited report. Your job is to ask the right question, not to search keywords.

    WORKFLOWone question → one cited report
    01FRAMEturn a fuzzy question into a bounded research brief
    02DEEPPerplexity Deep Research · runs 30+ pages
    03VERIFYpick 3 key claims · open the citations, check sources
    04BRIEFdrop into Claude · 3 paragraphs for the boss
    SETUP2 min · ROIone research task: an afternoon → 30min
    Perplexity Claude + web NotebookLM
  • CATEGORY · 03 · ORCHESTRATE run by hand → runs itself

    The real leverage layer: wire tools into a self-running line

    Contrarian take: 90% of people only swap things in the "model layer" and never touch this one — yet all the leverage is here. The orchestration layer welds "trigger + a few tools + one output" into a line. Once it's built, the workflow actually "runs itself."

    WORKFLOWform in → auto-process → push to channel
    01TRIGGERn8n / Zapier · listen for new form / new email
    02PROCESScall Claude · classify / summarize / draft
    03ROUTEbranch on result · write DB / create task / notify
    04PUSHresult to Slack / Teams / email
    SETUPhalf a day for the first one · ROIfrom "I remember to run it" → "it runs itself"
    n8n Zapier Make Claude Code
  • CATEGORY · 04 · THE LAUNCHER three tab-switches → one hotkey

    Put AI within arm's reach

    Contrarian take: when a tool doesn't stick, it's often not the tool — it's that "opening it" is too much friction. Bind your common AI actions to one hotkey / one desktop entry, and friction drops from "open browser, log in, paste" to "one key." Whether a tool sticks comes down to exactly that friction.

    WORKFLOWselect any text → one-key action
    01SELECThighlight text in any app
    02HOTKEYRaycast / quick-action · one keystroke
    03ACTrewrite / translate / summarize / reply · preset
    04PASTEresult drops back in · never leave the window
    SETUP15 min to wire the hotkey · ROIsaves countless tab-switches a day
    Raycast Claude Desktop Connectors
  • CATEGORY · 05 · CAPTURE forgotten by EOD → notes in 30s

    Stop using your brain to record meetings — use it to decide

    Contrarian take: taking meeting notes spends your most expensive resource (attention) on the cheapest work. Let a tool listen the whole time; 30 seconds after the meeting it produces structured notes — decisions, action items, owners, deadlines. You just think in the room.

    WORKFLOWmeeting → notes → action items
    01RECORDGranola / Fireflies · hit record before
    02SUMMARYauto transcript + summary + action items
    03FOLLOWdrop into Claude · write the follow-up email
    04TRACKpush action items to Todoist / Notion / Linear
    SETUP5 min, once · ROI30 min saved per meeting
    Granola Fireflies Claude Notion / Linear
§ 06 · Run it tonight

Three workflows — pick one, run it end-to-end tonight

Everything above is theory until one flow actually runs. The three below are zero-code, same-day effect, installable before you sleep — pick one, finish it, and by 8 AM tomorrow you'll see the first output. Don't do all three. Do one. Finish it.

DEMO · 01 15 min setup · 0 min daily use

Wire "read email daily" into a pipeline

In the time it takes to drink a coffee, let a workflow read the last 24 hours of email and output "what you have to do today." Opening the inbox stops being anxious — you read the list, not the emails.

  • Get Claude Pro, enable the Gmail Connector in settings
  • Create a Project, name it "Daily Email Assistant"
  • Paste the prompt template into Project Instructions (handed out live)
  • Set an 8 AM phone alarm to remind you to run it
  • After a week, come back and tune the prompt — the more you use it, the better it knows you
claude · daily email assistant
YOUrun today's email
CLAUDEreading the last 24h…
GMAILread 47 messages
▌ Must do today (by priority)
· Reply to David — contract terms (waiting 18h)
· Confirm Thursday 2pm TI team meeting (needs reply)
· Approve Sarah's expense report $842
▌ Auto-archived 38 notifications / subscriptions
▌ Drafted 6 replies, ready for you to send
DEMO · 02 20 min setup · 1h saved each time

Turn "repeat writing" into a Project

Weekly reports, standard replies, product updates — anything you write 3+ times a week with a similar structure, make it a Claude Project, and put your voice and template into the instructions. Never start from scratch again.

  • Pick one writing task you do 3+ times a week with a stable structure
  • Create a Project; put the fixed structure + samples of your voice into the instructions
  • Drop in your last few finished pieces as a "style corpus"
  • From now on, give it only this round's raw material and let it draft
  • You only subtract and review — no more writing from zero
claude project · weekly report
YOUraw material: 3 project updates + 2 risks
CLAUDEorganizing in your report template…
▌ This week (already in your voice)
· Project A: shipped performance baseline report
· Project B: North America pricing approved
▌ Risks & decisions needed
· Adjust Europe pricing in parallel? (your call)
✓ Draft ready · edit 2 things and send
DEMO · 03 half-day setup · then runs itself

Weld two tools together with n8n / Zapier

This is the step where you graduate from "AI user" to "workflow builder": make one tool's output automatically become another tool's input. E.g. new form in → Claude classifies → push to Slack. Build it once, it runs forever.

  • Sign up for n8n (self-hostable) or Zapier
  • Pick a trigger: new form / new email / schedule
  • Add a Claude node in the middle: classify / summarize / draft
  • Add an exit: write to Notion / create a task / send to Slack
  • Run it for a week, see which step needs tuning — that's "Review"
n8n · auto form-handling flow
01TRIGGERnew customer form submitted
02CLAUDEinfer intent + urgency · draft a reply
03ROUTEhigh priority → create Linear task and @me
04SLACKpush to #inbound · attach the draft
✓ Handled automatically · no need for you to be there
§ 07 · The next 90 days

From tool collector to workflow builder

This 90-day path doesn't promise you'll "use every tool." It gives you four specific actions, each with an explicit failure signal. If a week's failure signal lights up, go back to the previous week. Don't keep moving forward.

  • WEEK 1 – 2 · Stop collecting

    Uninstall half — keep only the two or three you actually use

    Clean the bookmarks: cancel / delete every AI tool you haven't truly used in the last three months, keep 2–3. Failure signal: by the end of week 2 you've installed something new again. What to do: kill every "AI tools weekly" notification — cut the source.

  • WEEK 3 – 4 · Wire the first flow

    Pick a task you do 3+ times a week and wire it into a workflow

    Use the four steps from §04 (Trigger→Steps→Output→Review) to turn a high-frequency task into a Claude Project or a pipeline. Failure signal: by end of week 4 you're still writing the prompt from scratch every time. What to do: the task is too messy to SOP — pick a more specific one.

  • WEEK 5 – 8 · Connect the tools

    Turn on connectors, wire the first automation

    Enable the official Gmail / Calendar / Drive / Notion connectors, then weld one simplest automation with n8n / Zapier. This is the moment you graduate from "AI user" to "workflow builder." Failure signal: connectors are on, but "let it run itself" never shows up in your daily work. What to do: go back to weeks 3–4 — your workflow didn't lock in.

  • WEEK 9 – 12 · Maintain the system

    Hold 2–3 workflows steady; you only do judgment

    Train 2–3 core workflows to "run themselves daily"; you only do high-value judgment and final review. Failure signal: by week 12 you're still doing manual work all day with no flow running on its own. What to do: the problem isn't the tool — it's that you never separated "doing the work" from "building the system."

After 90 days you should be able to answer: "In the past week, which workflow ran end-to-end on its own, where I only looked at the result?" — if you can answer, with screenshots of the output, the path worked. If you can't, you spent 90 days "swapping tools," not "building a flow."

Real Agent Use Cases

Tools go out of date. Workflows don't.

Half the tools you bookmarked will be gone next year. But the workflows you build keep running — just swap the tool.
"Execution is always undervalued." If you don't wire one flow tonight, you won't tomorrow either. That's not a pep talk. It's an observation.

← EP 02 · Multi-agent & Skills · EP 03 (you are here) · EP 04 (coming)