HomeLearnCoursesWebinars
The AI PM HubReading the latest with you
LearnCoursesServicesWebinars
עב
  • The Map
  • Articles
  • Videos
  • Skills
  • I'm new to AI. Where do I start?
  • I have a product idea. How do I get to an MVP?
  • How do I work with Claude and Claude Code?
  • I'm job hunting or preparing for interviews.
  • AI PM Hub Weekly #1: Agents Write Code, but They Don't Order Pizza Yet
  • Stop Being the Glue: Turn Your Claude Skills Into an Agent That Runs on Its Own
  • Rules, Commands or Skills? A PM's Guide to Picking the Right One
The AI PM Hub

Taught by Ofer Regev — helping you go from chatting with AI to building with it.

Join WhatsApp group
  • LinkedIn
  • YouTube
  • GitHub
  • theaipmhub@gmail.com
Work with me
  • Live courses
  • Team workshops
  • Consulting
Free content
  • Learn: the Map
  • Articles
  • Videos
  • Skills
  • Webinars
  • WhatsApp group
    • For agents
    • llms.txt
    • sitemap.xml
    © 2026 The AI PM Hub · Learn to build with AI
    AboutContactTermsPrivacyAccessibility
    Claude Skills
    Agents
    Automation

    Stop Being the Glue: Turn Your Claude Skills Into an Agent That Runs on Its Own

    Article·7 min read

    Published October 29, 2026 · By Ofer Regev

    Live webinar, Sunday 15 November, 10:00 (Israel time): I'm walking through this whole setup live, with a real run. Join the free webinar

    A few weeks ago I wrote about an 8-step job search you can run with Claude skills. The skills worked. But after a month of using them, I noticed something: every step still needed me.

    I had to remember to check the job-alert emails. Open a chat. Paste my background and rules again. Copy the result into the tracker. Then do it again for the next role. Each step had a skill, and I was still the thing connecting them.

    This piece is about how I stopped being the glue. It's written around a job search, but the pattern works for any process you repeat: lead follow-up, hiring, research monitoring, customer-feedback triage.

    A skill is a recipe. An agent is a cook.

    A skill is a recipe: it makes Claude excellent at one task, the moment you ask for it. An agent is a cook who shows up on schedule, knows where the kitchen is, and leaves a note on the counter.

    Being the glue cost me three things:

    1. Forgetting. Job alerts piled up unread unless I went and looked.
    2. Re-explaining. Every new chat started with pasting my background and rules again.
    3. Scattered state. Results lived in chats, files and my head, so I couldn't answer "where do I stand?" in one look.

    Step 1: give every skill's output an address

    This is the part most people skip, and it's the part that makes automation possible.

    "Here's a nice answer in chat" can't be automated. "A new row in this table" can. Before you schedule anything, make sure each skill writes its result to a predictable place:

    SkillWhere the result lands
    Target company finderA ranked list in the Target Companies tracker
    Fit analyzerAn honest verdict (strong, stretch or long shot) on the application row
    CV tailorA file linked from the application row
    Company researcher and interview prepA prep file for that role

    More articles

    Stop Blasting the Same CV: An 8-Step Job Search System You Can Run With Claude
    AI for PMs
    Claude Skills
    Career

    Stop Blasting the Same CV: An 8-Step Job Search System You Can Run With Claude

    Most job hunts fail on process, not talent: the same CV everywhere, company research the night before, interview prep rebuilt every round. Eight steps, each one a ready-made Claude skill, with you reviewing every output.

    6 min readOct 11, 2026

    The tracker (mine is in Notion) becomes the shared memory. Every skill and every agent reads from it and writes to it, so none of them has to remember anything. That matters, because a scheduled agent remembers nothing.

    Step 2: write the agent as a standalone brief

    I run two scheduled agents. Neither is a new skill. Each one is a prompt that chains the skills' logic, points at the tracker, and runs without me.

    • Daily job-alert triage, every morning around 08:15. It reads the last day's job-alert emails, drops roles at companies already in the tracker, fit-checks the rest, adds the promising ones as rows, and sends me a short phone summary.
    • Weekly pipeline check, Sundays around 08:45. It finds target companies with no contact, flags outreach that has stalled, refreshes the target list when it runs low, checks careers pages for new roles, and sets missing next-step dates.

    Every run follows the same loop:

    1. Wake up. A schedule fires a fresh session with no memory of yesterday.
    2. Read the state. Check the tracker for what already exists, so nothing is duplicated.
    3. Gather new input. The last day's emails, or the careers pages.
    4. Judge. Apply the fit rubric to each new role.
    5. Write. Add rows with the verdict, the evidence, the link and a next-step date.
    6. Report. One short message that fits on a phone screen.

    Because every run starts cold, the prompt has to carry everything. Mine always has six parts:

    • The goal in one sentence. "I never want to waste time on roles I don't fit."
    • Exact data locations. Which tracker, which columns, and an instruction to check the structure first.
    • The rubric, written out. My profile, my known gaps, and what counts as strong, stretch or long shot.
    • Numbered steps, each with a clear point where it stops.
    • The output format. What the final message says and how long it is.
    • The guardrails. Next section.

    Step 3: guardrails that make it safe to leave alone

    Letting something run while you sleep is only a good idea if it can't do damage. My rules:

    • Add, don't edit. The daily agent only adds new rows. It never changes existing ones.
    • No outward actions. It doesn't apply, doesn't message anyone and doesn't touch my inbox.
    • An exclusion list. Companies I've ruled out are named in the prompt and never come back.
    • Honest about gaps. If a careers page can't be read, it says "unreadable", never "no roles".
    • Visible judgment. A role it screens out is still logged as closed, with the reason, so I can overrule it.

    Anything that reaches another person or can't be undone stays my decision. And every new agent starts read-only: I watch three runs before I give it permission to write.

    Step 4: a status page, so you can trust it

    An agent you can't see is an agent you won't trust. The phone message tells me what happened today. A status page tells me where everything stands.

    Mine is a private page that Claude built and published. It reads the tracker live, so there's no copy to keep in sync, and it's pinned one tap away. It shows every application grouped by stage, the next-step dates (so overdue items stand out), the tailored CV on each row, and when each agent last ran. I can change a status right on the page and it writes back to the tracker.

    Four design choices worth copying:

    1. Let the data drive the page. Stages, order and colours come from the tracker itself. When I added "Pending Response" and "Paused" as stages, the page needed no change.
    2. Load fresh data when you open it. Nothing runs in the background just to keep a screen up to date.
    3. One row tells one story. Verdict, evidence, link and next step sit together.
    4. Make the agent's work auditable. Anything an agent added is visible, with its reasoning, and I can reverse it.

    My mornings now look like this: the agent has written new rows overnight, I read the phone summary, open the page, and spend my time on the two or three roles marked strong. When something changes in real life, like a rejection or an interview invite, I update the status that same day, and the next run starts from there.

    What went wrong first

    What went wrongThe rule I follow now
    Early prompts assumed context the scheduled run didn't haveWrite every prompt as a standalone brief
    The tracker filled up with roles I'd never takeFit-check before writing; log long shots as closed, with the reason
    Rows with no next-step date went staleThe agent sets a date on every open row it touches
    I approximated stages with notes when none fitAdd the real stage to the tracker
    Pages it couldn't read were reported as "no roles"Report them as unreadable, never as empty
    Real life moved faster than the trackerUpdate the status the same day

    One honest limit

    The agents save me time on triage and bookkeeping. They don't replace networking, interview practice or deciding what I actually want. And the fit verdicts are only as good as the rubric I wrote.

    Build your own this week

    You don't need to code. The agents are prompts, and the status page is something you ask Claude to build.

    1. Pick one repeating process with a clear definition of done.
    2. Create the tracker your skills will write to.
    3. Write the rubric: what good, stretch and bad look like.
    4. Draft the prompt: goal, data, rubric, steps, output, guardrails.
    5. Run it by hand three times before you schedule it.
    6. Schedule it, add-only, then add a status page.

    The short version: a skill becomes an agent when its output has an address and it runs on a schedule. Keep the state in a tracker, not in chats, keep the agent's permissions narrow, and give yourself a page where you can see what it did.

    Want to see it run? Join the free live webinar on Sunday 15 November at 10:00 Israel time: the real prompts, the status page and a live demo, plus time for your questions. Join the webinar

    ← Previous: AI PM Hub Weekly #1: Agents Write Code, but They Don't Order Pizza YetNext: Rules, Commands or Skills? A PM's Guide to Picking the Right One →

    On this page

    • A skill is a recipe. An agent is a cook.
    • Step 1: give every skill's output an address
    • Step 2: write the agent as a standalone brief
    • Step 3: guardrails that make it safe to leave alone
    • Step 4: a status page, so you can trust it
    • What went wrong first
    • One honest limit
    • Build your own this week