I Automated the Pipeline
It filled the funnel with cold leads and ate my calendar
The line I got wrong
Last spring I published a post mortem on eleven months of job searching, run like a product retrospective. It had four chapters. The first one was about resumes and the machines that read them, and I closed it with a tick: "Resume reads clean now. This one's done."
Then I did what a product manager does after a retrospective: I specced Resume Factory, an AI pipeline that automates my job search from posting to finished application package. Claude, running in Cowork, built a tracker, a Gmail mirror that reads recruiter and job-alert mail into it, hard gates for roles I will not take, and a script that tailors a resume and cover letter to each posting. I had been tailoring every application since March, pasting a prototype resume and the posting into a chat under rules I wrote to stop the AI making things up (strict as F!@&!K). The pipeline turned one prototype into six, added a story bank I wrote myself, and ran the same rules on every package. I directed, reviewed and applied. Between June 1 and November 5 Resume Factory logged 871 job descriptions, packaged 123 (customized and laid out), and I submitted 109 applications, twice the monthly pace of Cohort 1.
The number of cold screens a month did not move. I am re-litigating chapter one.
By the Numbers
Applications: 109
Recruiter screens: 15
Hiring-manager rounds: 9
Cold screens: 3
§ 01 — Cold vs. warm
A cold application means I found the posting and applied with no prior relationship at the company. A cold screen is a recruiter's call in reply to one of those applications, whether the recruiter works for the company or for an agency. A warm screen started with a person: a recruiter who contacted me first, or a referral. Cohort 1 was counted the same way.
The Challenges
CH1: Twice the effort, the same cold yield
Per month, I sent twice as many applications and got no more cold screens. The two cohorts are compared on monthly pace because their windows differ.
The cold screen rate went from 6.7% (7 of 105) to 3.3% (3 of 91). That looks like a halving, but on only seven screens and then three, the two rates fall in essentially the same bracket (p=0.34). Nearly twice the cold volume bought no extra cold screens. What grew, grew because I submitted more applications, period.
CH2: Two doors
There are two ways into a job, and I wrote about the difference when a recruiter invite turned out to be an application invite in a costume. Through the human door, someone has already chosen me before I do anything. Through the jobs portal door, I choose them and wait (and wait, and wait, pretty much forever).
Ten of twelve against three of ninety-one. Warm channels together held level with Cohort 1, 67% against 69%, and referrals dipped from 3 of 4 to 2 of 6, too few to mean anything. The jobs portal number is the one that has not changed its mind across two cohorts and two tracking systems.
CH3: What the pipeline changed, and what it did not
The pipeline changed how much reached me, not what came back. Before June 27 everything was manual: Evernote tags, hand sourcing, whatever landed in my inbox. The tracker and the Gmail mirror went live that day; the gates followed over the next weeks.
So Resume Factory filled the top of the funnel, doubled the middle, and left the cold door exactly where it was. Applications did not follow intake because the gates and my hours cap them, which is the point of gates.
What it does show is where the limit sits. Filling the funnel did not widen me (yay Ozempic!). Every job description the mail mirror pulled in still needed a human to read it, and in August it pulled in 425. Applications held at about twenty a month because that is what my hours allow, not because the pipeline ran dry. So from late August the work went into rules rather than resumes.
By October I trusted what reached me. Most of the rule work was making sure Cowork was not flushing a dream job down the toilet. The pipeline's output was never the bottleneck. I was.
Even in October, two of every three job descriptions that got past the rules died before I applied. Across all five months, 183 closed at the employer's end before I applied, more than a third within five days of capture. The pipeline made applying cheaper. It did not make the window any wider.
Both cohorts sent parse-clean, role-tailored resumes, and the second sent deeper ones, twice as often. The cold screen rate did not rise.
CH4: The human door, briefly
Every role I have held since 2019 came through recruiter outreach. The earlier post said that channel works, and Cohort 2's data backs that up, and then some. I am in an interview process through one of those recruiters right now, so this post names no companies.
What the human door gives me that a jobs portal never has is an answer. A recruiter who carried my name to a hiring manager came back with a specific no. A jobs portal comes back with a form letter or with nothing at all, and after sixty days of silence I log it as ghosted.
.st162{--c-ink:var(--st-color-text-default);--c-muted:var(--st-color-text-muted);--c-grid:var(--st-color-border-default);--c-rule:var(--st-color-border-strong,var(--st-color-border-default));--c-surface:var(--st-color-bg-surface);--c-surface2:var(--st-color-bg-surface-strong,var(--st-color-bg-subtle));--c-a:var(--st-color-text-muted);--c-b:var(--st-color-accent-primary);--c-brand:var(--st-color-text-brand);--c-good:var(--st-posture-strong,var(--st-color-text-default))} Per month: twice the effort, the same cold yield Applications doubled. Cold-only screens per month did not move. The extra screens were warm. Cohort 1 (Jul 2025 to May 2026, 111 applications) Cohort 2 (Jun 1 to Nov 5, 2026, 109 applications) 0 5 10 15 20 25 10.1 21.0 Applications x2.1 9.5 17.5 Cold applications x1.8 0.64 0.58 Cold-only screens x0.9 1.0 2.3 Warm screens x2.3 1.6 2.9 All screens x1.8 4.2 10.8 Rejections x2.6 Per month = total divided by 11 (Cohort 1) or 5.2 (Cohort 2). Cold-only screens: 7 then 3, too few to call a change. Source: findings.json and the published monthly table.
.st162{--c-ink:var(--st-color-text-default);--c-muted:var(--st-color-text-muted);--c-grid:var(--st-color-border-default);--c-rule:var(--st-color-border-strong,var(--st-color-border-default));--c-surface:var(--st-color-bg-surface);--c-surface2:var(--st-color-bg-surface-strong,var(--st-color-bg-subtle));--c-a:var(--st-color-text-muted);--c-b:var(--st-color-accent-primary);--c-brand:var(--st-color-text-brand);--c-good:var(--st-posture-strong,var(--st-color-text-default))} Two doors: share of applications that reached a screen Through a person, most applications reach a screen. Through a jobs portal, almost none do, in either cohort. Cohort 1 (Jul 2025 to May 2026, 111 applications) Cohort 2 (Jun 1 to Nov 5, 2026, 109 applications) A recruiter wrote to me first 67% (8 of 12) 83% (10 of 12) Someone referred me 75% (3 of 4) 33% (2 of 6) I found the posting and applied cold 6.7% (7 of 105) 3.3% (3 of 91) Screen = recruiter screen, including a staffing recruiter's screen about a client seat. Cohort 2 counts applications only; 1 recruiter call with no application is excluded. Source: findings.json at the 2026-11-05 cut; Cohort 1 from the published charts.
.st162{--c-ink:var(--st-color-text-default);--c-muted:var(--st-color-text-muted);--c-grid:var(--st-color-border-default);--c-rule:var(--st-color-border-strong,var(--st-color-border-default));--c-surface:var(--st-color-bg-surface);--c-surface2:var(--st-color-bg-surface-strong,var(--st-color-bg-subtle));--c-a:var(--st-color-text-muted);--c-b:var(--st-color-accent-primary);--c-brand:var(--st-color-text-brand);--c-good:var(--st-posture-strong,var(--st-color-text-default))} The pipeline, Jun 1 to Nov 5, 2026 Bar width = share of the 871 job descriptions the system captured. The pipeline widened the top; the gates and my hours set the middle; the doors set the rest. Job descriptions captured 871 (100.0%) Packaged (tailored resume and letter) 123 (14.1%) Applied 109 (12.5%) Reached a recruiter screen 15 (1.7%) Hiring-manager round or later 9 (1.0%) Offer 0 Captured and packaged are automated counts; applied, screens and rounds are hand-verified rows. 1 recruiter call with no application is not shown. Source: findings.json at the 2026-11-05 cut.
.st162{--c-ink:var(--st-color-text-default);--c-muted:var(--st-color-text-muted);--c-grid:var(--st-color-border-default);--c-rule:var(--st-color-border-strong,var(--st-color-border-default));--c-surface:var(--st-color-bg-surface);--c-surface2:var(--st-color-bg-surface-strong,var(--st-color-bg-subtle));--c-a:var(--st-color-text-muted);--c-b:var(--st-color-accent-primary);--c-brand:var(--st-color-text-brand);--c-good:var(--st-posture-strong,var(--st-color-text-default))} Rules, not resumes What became of every job description the system captured, by month. In August, 87% of what got past the rules died before I applied. By October, with intake at a quarter of August's, it was 67%. Applied Died before I applied Failed a rule Rule added 0 100 200 300 400 0 1 2 3 4 5 21 8 32 Jun 28% died 24 119 169 Jul 83% died 23 164 237 425 Aug 87% died 12 45 71 128 Sep 79% died 23 49 33 106 Oct 67% died 0 June 27 : Tracker and Gmail mirror go live; the hard gates follow 1 August 21 : A scored triage board, to see what the gates throw out before it is gone 2 September 4 : A duplicate check 3 September 20 : A Date Applied column with a fourteen-day follow-up clock 4 October 2 : Ecommerce and DAM roles re-ranked 5 October 18 : The Director tier cut Died before I applied: the listing closed, or it was never reviewed before it went stale. Failed a rule: a hard gate or the duplicate check. "% died" is the share of what got past the rules. Counts come from the close notes in my tracker, matched by pattern, so read them as about. Two rows still open (August, October) are counted in the totals above the bars only. Months are capture months; June includes 17 job descriptions logged from before the tracker launched on June 27; November (5 days) is not shown. Source: findings.json at the 2026-11-05 cut, and dead_listings.py.
.st162{--c-ink:var(--st-color-text-default);--c-muted:var(--st-color-text-muted);--c-grid:var(--st-color-border-default);--c-rule:var(--st-color-border-strong,var(--st-color-border-default));--c-surface:var(--st-color-bg-surface);--c-surface2:var(--st-color-bg-surface-strong,var(--st-color-bg-subtle));--c-a:var(--st-color-text-muted);--c-b:var(--st-color-accent-primary);--c-brand:var(--st-color-text-brand);--c-good:var(--st-posture-strong,var(--st-color-text-default))} How each screen started How each screen and hiring-manager round started. People went from 61% of screens to 80%, and to 89% of the rounds that came after. Recruiter outreach Referral Cold application Cohort 1 screens (18) 8 3 7 61% people Cohort 2 screens (15) 10 2 3 80% people Cohort 2 hiring-manager rounds (9) 7 1 1 89% people Cohort 1 is the re-count by door (18 screens; the first post counted 19). Cohort 2 counts applications only. Source: findings.json at the 2026-11-05 cut; Cohort 1 from the published charts.
Who opened the door
After five months of the pipeline, the answer to "what works" is the one I had before I built it, only louder. Ninety-one cold applications produced three screens. Twelve came through a person, and eight of the nine hiring-manager rounds started with someone who already knew my name.
Now what? Being found
I spent the summer scrambling to vet a phat pipeline, and more volume was never going to help. So the next project is not a better resume or more applications. It is that old familiar workstream: being found on the internet. SEO, and now GEO, which is getting named when someone asks an AI assistant.
The next post looks at the other door: what is on the public record about jobs portals and this crappy-ass challenging job search.
Reader, I’m not alone.
Notes on the numbers
Every figure about my own search comes from scripts run against my tracker, last on 2026-11-12, window closed at 2026-11-05. Screens and hiring-manager rounds are a hand-verified list, each with a debrief on file. "Closed with no answer" now means sixty days with no reply, so the two cohorts' silence rates are not compared. Claude in Cowork built the tracker, the mail mirror and the analysis script and has run them since June; Claude Code re-cut the numbers and drew the charts for this post; I set the rules, read every row that mattered, and wrote the conclusions. Cohort 1 counted hiring-manager screens and Cohort 2 counts hiring-manager rounds or later: close, not identical. The first post's funnel counted 19 Cohort 1 screens. This re-count, by door, finds 18, and the charts here use 18. Cohort 1's door counts (105 cold, 12 recruiter outreach, 4 referral) add up to 121 against 111 applications, because some applications came through more than one door.