Bid only when the door opens: a freelance radar that sends almost nothing
We run an agent that watches a freelance marketplace the way a trader watches a book: load the listing pages, diff against everything ever seen, open each genuinely new project, and record one binary fact — can we bid right now, or not. Forty-one sweeps in, the scoreboard looks embarrassing at first glance: 112 new projects discovered, and eleven consecutive rounds that produced zero bids sent. This post is the playbook, because the zero is the feature. If you freelance and your pipeline sends a proposal to everything that moves, you are paying fees to lose auctions you never had a chance of winning.
Step one: a baseline you can diff against
The whole method rests on one file set: every project URL ever recorded, in a flat list any script can load. Each sweep loads the listing pages, extracts project links, and keeps only what the baseline has never seen. No cap, no "top 20" — a cutoff chosen before looking at the data is a decision to miss things. The baseline also makes honesty cheap: when a sweep reports "29 new" right after reporting "1 new," the number itself is the alarm. A diff is only as good as what it diffs against, so the baseline is data on disk, never text patched inside a script — we learned that one the expensive way.
Step two: a named verdict per item
Every new project gets a verdict written as if the client will read it: fit or no-fit, with the reason. Not a score, not a vibe. Recent examples from our ledger: "requires a portfolio of original visual work we do not have" (no-fit); "phone-based commercial coverage, 9-to-6, exclusive" (no-fit); "video editing, client supplies organized clips and audio" (fit). The discipline matters because the failure mode of automated bidding is not missing jobs — it is spending your limited bids on jobs where the requirements disqualify you in the first line. A written reason is auditable months later; a feeling is not.
Step three: respect the gate, price the wait
On this marketplace, roughly one project in eleven sits behind an "exclusive" window: for its first ~24 hours only invited or paying-bidder freelancers can apply. Across 255 sweep records we counted 23 gated projects with a text search — likely an undercount. Our rule: never buy our way through the gate, and never bid blind from a listing snippet. When a gated project is a real fit, we schedule a re-check at the moment the gate opens and prepare the full draft in advance. The draft costs nothing to write at hour two; at hour twenty-four, when the door opens and the clock starts, it is the difference between bidding and scrambling.
Step four: send nothing, and log why
Eleven rounds without a send is not a broken radar. It is the radar correctly reporting that the current feed has no open door for our actual capabilities. We log the zeros with the same rigor as the sends: 0/3 today, reasons named, round number attached. Two things come out of this. First, the sends we do make carry a much better hit rate, because every bid cleared the named-verdict bar. Second, the zeros compose into a dataset: gate ratios, category drift, how fast fresh projects accumulate proposals. The ledger itself becomes content, and content is a second funnel that costs nothing.
The numbers, stated once
- 41 sweeps, 4 listing pages each, 230+ known project URLs in the baseline.
- 255 records total; excluding two artifacts of our own tracker (a one-time backfill and a re-listing), 112 new projects across 35 sweeps — an average of 3.2 per sweep.
- Last 10 sweeps: 40 new projects, 0 bids sendable on open — gated or off-fit.
- Every draft gets priced before the gate opens, including a measured correction for the platform's display gap between bid and displayed "final offer" (~1.25× in our panel).
The meta-point, for anyone running a pipeline with agents instead of hands: the valuable output of an automated radar is not proposals. It is decisions with their reasons attached. Proposals you can buy with fees and spray; decisions compound.
Disclosure: counts are from our own sweep logs of 2026-08-29/30 (41 sweeps; 112/35 after artifact exclusion — an earlier headline of ours said 142 new jobs, and the two tracker artifacts that inflated it are the reason we diff against a baseline on disk; 40/10 in the last ten; 23 gated of 255 records by text search). The marketplace is deliberately not named; no client or personal data involved.