This is the bvlogythingy...I hope you like it!

This is the longer version of a post I put on LinkedIn today. My last few entries were about how Agentic_JobSearch4me is working for my own search. This one is about finding out if it could work for other people. I'm calling the public version GoReachAgentic, and today I put up a one page sign-up at goreachagentic.com to see who is interested.

Why it starts with outreach

The tool I use every day does two things. It finds and scores job listings and builds applications, and it tells me who in my network to contact. I'm starting the public version with the second part, because the conversations that went somewhere in my search all started with a person I already knew. When I counted on October 2, five people had referred me for six roles since the rebuild, and none of those came from a cold application.

The outreach side is also the smaller thing to build for someone else. It needs one file from LinkedIn and a resume. The application side needs a lot more access, like email, and that can wait until I know people want the first part.

What it does

It takes the connections file LinkedIn lets you download and ranks every person in it. For each person it gives a reason they are on the list and something you could ask them, and it drafts a short note when you ask for one. You send the note yourself. It doesn't message anyone for you, it doesn't apply to jobs for you, and it never asks for your LinkedIn password.

It looks at the whole network, including people at companies with nothing posted right now. Those people can still tell you who is hiring in their org. That goes back to something I wrote about in July, which was to rank everyone and not hide anyone.

How the page came together

The page took two tries, and the second one happened this morning.

• The first page was a mess. The first version went together on Sunday. It had the form in two places, and it asked how you heard about it, whether you were looking for work and what you would pay. This morning I had Claude throw it out and start over.
• One form, shown once. The new page asks for a first name, a last name and an email. A LinkedIn profile and a box for features you would want are both optional. The page keeps track of which link you came from, so it doesn't need to ask how you heard about it.
• Two clicks is too much to ask. We also tried a version with an intro page that led to the form on a second page. I dropped it, because I didn't want anyone to have to click twice to sign up.
• The first deploy was a 404. The first time the page went up, it published the wrong folder and the address came back "page not found". One config file fixed it.
• Every person on the page is made up. The example list shows Pat Example, Sam Sample and Lee Placeholder. I'm not putting my real connections on a public page.

What happens next

The version for other people is not built yet. The page is up for one week first, and I will read the sign-ups on October 13. The mark I set is 15 sign-ups from people outside my own connections.

If you are in a search, or you know someone who is, you can sign up here. There is a box on the form for features you would want. Please tell me what would make it useful for you! And if you are hiring a technical program manager, please let me know :)

This is the longer version of a post I put on LinkedIn. The post covers the week of September 28 to October 2. Here I wanted to go further into what happened with the board, and to put this week next to the five before it.

Last week I said this week would be about getting the most out of the network lane. It turned into something else. Until now I have run Agentic_JobSearch4me by typing to Claude in a terminal, and the plan from the beginning was for a screen to replace that. This week I tried to make that screen, which I call the board, the only way I use the tool. That took most of the week, and every number in the image below is lower than last week.

What the board is

The board is a window on my Mac that shows what the tool did overnight and the few things that need me. On Monday I opened the version I had and told Claude the whole UI was broken. Before rebuilding it, Claude went back through how I had actually used the tool over the past weeks, and we planned the new screen from that. It has four zones. I press Submit on applications the tool has already filled in. I read and send notes to people I know. I sign in when a site asks for it. And I decide the cases the tool held back because it was not sure. Everything else is supposed to happen without me.

What using only the board turned up

Using only the board showed me how much Claude had been doing by hand in those terminal sessions. I logged 24 problems over the week. These are the ones that taught me the most.

• Folders with no status. After I submitted an application, nothing moved its folder. That had always been a step Claude did by hand in a session, and on Wednesday it was skipped. By Thursday 16 folders had no status. Now a folder follows the status of its application, and that rule lives in one place in the code.
• People I had already written to. On Thursday the board offered me notes to seven people I had already contacted. The history was logged, but the board did not show it, so a follow-up looked like a first note. The follow-ups were also built from a template, and six of them named an application that had already been rejected. The tool no longer writes a follow-up on its own to anyone I have a real conversation with. When we replayed the old drafts against the new rule, it refused all 21 of them.
• Receipts waiting on me. Every "application received" email showed up as a decision for me to accept. Those are recorded in the background now.
• A column one letter wide. One long link in a draft squeezed the column next to it until the text ran one character per line. The sample data we tested with had no long links, so the test passed and the real board did not.

Seventeen of the 24 are fixed. The board still does not show the history next to each note, and six problems from the old board are still on the list. My first plan for the follow-up problem also went through the skeptic committee, which turned it down and sent it back in a different order.

Week over week

Here is every number from the first image, week by week since the rebuild. Each week runs Monday to Friday and is labeled by its Monday, so 9/28 is this week. The highlighted bar is this week, and the line under each number is the change from the week before.

The trends so far

• TrendWeeks spent building are slow weeks. The two weeks with the fewest applications are September 14 (13) and this one (62), and they are also the two weeks with the fewest packages built (50 and 49). Both went to working on the tool.
• TrendReplies fell for the first time. They went 14, 21, 24, 27 and 32 over the first five weeks, then 18 this week. Messages sent went 42, 253, 110, 72, 152 and 75 over the same six weeks.
• TrendReferrals come from conversations. Five people have referred me for six roles since the rebuild, and none of those came from a cold application. This was the first week since the first one with no new referral.
• TrendFewer listings are coming in. Listings scanned went from 817 in the first week to 413 this week, and 126 roles scoring 70 or higher is the lowest week so far.

I do have an interview next week, which came out of the screening and take-home from September. Next week is about getting the volume back up, this time from the board. If you are hiring a technical program manager, or want to see how the board works, please let me know :)

This is the longer version of a post I put on LinkedIn. It started with two videos from Mo Bitar on his channel Less Bitter, "Opus 5.5 is disturbing" and "I talked to 50 people building with AI. It's a mess." His advice to solo builders is to keep things simple. Fewer rules files, a fresh conversation for each task, work on main, and no big multi-agent setups. It got me thinking about how I use Claude, so I had Claude compare my setup against his list.

What I was doing well

• Running Opus at the highest effort setting, which matches his "no lower than Opus" advice.
• Adding most of my rules after something actually went wrong, each one dated to the incident.
• Starting each session from live data (calendar, Gmail, BigQuery) instead of notes an earlier session wrote.

What I was doing poorly

• One build session ran from September 3 to September 26 and was compacted 57 times.
• A feature branch had quietly become my real main, 108 commits ahead of it.
• A long handoff paragraph at the bottom of my rules file loaded into every session and kept going stale.

So I merged everything back to main and made working on main a rule. I cut my rules file from 165 lines to 110 and replaced the handoff paragraph with a single line. I also built a /wrap command that saves what matters at the end of a session, which makes starting a fresh conversation for each task easy.

The one thing I kept on purpose is my skeptic committee, and it is exactly the kind of multi-agent setup he says to drop. I kept it because LLMs lean positive. They tend to agree with you and tell you the plan is solid. This comparison is a good example. Claude's first answer said my setup mostly fit Mo's advice and I only needed to try two things. When I ran that answer through the committee, it found I was doing about 7 of the 10 things he says to stop. Mo says in the first video that "skepticism should be the default stance," and that is what the committee is for.

How it works

Five reviewers each read the plan or analysis on their own, at the same time, without seeing each other's work. Each one has a single lens. For a build plan that is correctness, meaning, data, test quality, and whether the work is worth doing at all. For an analysis it is inference, data integrity, method, premise, and value. Each reviewer is told that a manager will check their findings, so a made-up problem counts against them.

The manager is not part of the panel. It reads the original first and forms its own view before it looks at anything the reviewers said. Then it rules on every serious finding as upheld, overstated or refuted, and names the reviewer behind any false alarm. It looks for anything all five missed, writes out the exact changes to make, and makes one go or no-go call.

Why it is set up this way

A manager that just summarizes the panel ends up inheriting its mistakes. Reviewers have a bias of their own too, since they want to look useful, and sometimes one will raise a problem that isn't really there. A manager whose job is to challenge the panel catches both. The structure now lives in code, and the workflow will not run without five distinct seats, because Claude got it wrong twice in one day when the rules were only written down. A run costs around 600,000 tokens, so I save it for decisions that are hard to undo. Day to day, a single reviewer agent does the checking.

The code is public on GitHub if you want to look at it or try it yourself. If you are hiring a technical program manager, or want to talk through how the committee works, please let me know :)

This is the longer version of a post I put on LinkedIn on Saturday. The post covers the week of September 21 to 25. Here I wanted to put that week next to the four before it, so the week over week changes and the trends are easier to see.

Last week I said this week would be back to volume, and it was. Agentic_JobSearch4me now builds a resume and cover letter package overnight for the roles that score 70 or higher, so each morning starts with a stack of roles ready for me to review. It built 278 packages this week and I submitted 218 applications, the most in any week since I rebuilt the tool in late August.

Volume showed me a limit I had not planned for. Some employers cap how many applications one person can send in a window, and a batch of twelve to one company ran straight into that. The tool already tracked a few of these caps. I am now adding a per-employer cap, so when a batch has several roles at the same company it picks the strongest one or two instead of all of them.

The bigger change is in outreach. The tool used to start with a company I had applied to and then look for someone who works there. As of Friday it starts with the people I already know. Each morning it picks people in my network I have not written to yet and drafts a short note to catch up, with no job link and no ask. I read every one and send them myself.

Week over week

Here is every number from that image, week by week since the rebuild. Each week runs Monday to Friday and is labeled by its Monday, so 9/21 is this week. The highlighted bar is this week, and the line under each number is the change from the week before.

The trends so far

• TrendReplies have gone up every week. 14, 21, 24, 26, then 31, whether I sent more messages that week or fewer. Messages sent went 42, 253, 110, 71, then 152 over the same five weeks.
• TrendApplications follow the build. The week of September 14 went to building the tool, and it had the fewest packages (50) and the fewest applications (13). This week the overnight build made 278 packages and I sent 218 applications, the most of both.
• TrendReferrals come from conversations. Five people have referred me for six roles since the rebuild, at least one in every week after the first, and none of them came from a cold application.
• TrendFewer listings are coming in. Listings scanned went down in three of the last four weeks, from 817 in the first week to 480 this week. Roles scoring 70 or higher held between 169 and 202 for four weeks and dropped to 139 this week.

Next week is about getting the most out of the network lane. If you are hiring a technical program manager, or want to see how the tool works, please let me know :)

This is the long version of a post I put on LinkedIn the same day. The short version: the first job search tool I built was a smarter alert filter, it did not work, and the rebuild finally did.

The first version pulled the job alert emails out of Gmail, scored each listing, and showed me the ones that cleared a bar. It felt productive. When I finally sat down with the data in late July, two things were wrong with it. The intake was narrow, a couple of alert sources and a threshold, so I was seeing a thin slice of the market and applying to a thinner one. And the outreach side did not exist as a system at all. I had a spreadsheet of people I had messaged and no rule for who to contact next. When I audited a year of those messages in August, LinkedIn DMs got replies about 30 percent of the time and had produced zero meetings and zero referrals. Every meeting I had actually landed came from a direct email asking for a conversation. I was busy in the channel that did not work.

So I rebuilt it, working with Claude, into the three stages in the graphic below. Click it to see it full size.

Intake runs on its own every hour. It reads alerts from every source I could route into one inbox, LinkedIn, Indeed, Jobright, Dice, Wellfound and the rest, parses each listing out of the digest, and scores it 0 to 100 against a preferences file that says what titles, keywords, and dealbreakers I care about. Anything at 75 or above lands in the dashboard, and I get a digest email and a phone push only when something new cleared the bar. A staleness sweep re-fetches any listing older than ten days and drops it if the posting is closed or no longer scores.

Apply is one role at a time. I pick it, the tool drafts a resume and cover letter tailored to that listing from my master resume, and a second pass checks the output against a set of applicant tracking system rules: no placeholders, no invented facts, dates and titles that match the source. Then I read it. When I mark it applied, it logs the application to BigQuery and a sheet and the role leaves the dashboard on its own. Getting that last part right took a real fix: for a while the logger was failing silently, so a reconcile job now checks both the ledger and my Gmail for confirmation emails before it hides anything.

Outreach is the part that was missing. For every application, the tool looks at who I already know at that company across my connections and a contacts graph I have been building since spring. Each person has a warmth score: a base for how we know each other, plus every interaction we have had, weighted and decayed. Good signals fade on a 120 day half-life, bad ones on 365, so a burned bridge outlasts a good conversation. An eligibility gate runs before anything else, per ask rather than per company, because a referral request can be off the table while a coffee chat is fine. Unknown data fails open so I see it rather than lose it. The warmest people rank first, at most two per company, and a model writes a one line hook for each. The rest of the note is a fixed template that refuses to render with an empty hook, a dollar figure, or a semicolon. Every note lands in my Gmail drafts. Nothing sends until I have read it.

The whole thing is about nineteen thousand lines of Python across sixty modules with 273 tests, built in 480 commits since late May. Streamlit for the dashboard, the Gmail API, BigQuery and Sheets for the ledgers, Gemini for the drafting, launchd for the schedule, Playwright for the couple of places a logged in browser is unavoidable.

Here is what changed. Same person, same market, same hours.

The touches are now mostly email, phone, and text rather than cold LinkedIn messages, which is the channel shift the audit said to make.

I do not have the new role yet. What I have is a search that runs at a volume I could not sustain by hand, with my attention spent on the people rather than the portals. I am considering opening up a cleaned-up version of the tool if there is interest.