Eleven posts in this series are things that worked. This is the rest: four we read carefully and turned down, each for a stated reason. Three of the four are good software, which is the point — most of these decisions are about where your material ends up, not about code quality.
The router, and where your prompts end up
A gateway that sits between your assistant and the internet and routes your messages to other providers' free tiers, hopping to the next when one runs dry. It is genuinely popular — over fifty thousand stars, MIT licensed, actively maintained — and the refusal has nothing to do with the software being bad.
It is about destination. 340 providers is 340 places your prompts can land, and this machine holds client material and credentials. Sending is the disclosure; you cannot un-send it afterwards.
The author says so himself, plainly, which is why we rate the piece:
"Do not run client work or anything private through it. Your messages pass through other companies' free servers, and free servers usually get to keep and train on what you send."
There is a second problem, and it matters if you care what produced an answer. In his own test, the model doing the work was a different company's free model while the screen still said Claude. Three places tell you what is answering and two of them lie.
And his conclusion is the one we already run: spend less on the expensive models rather than dodging them.
The lead engine with no consent model
Scrape a map for businesses, push them into a CRM, and fire a sequence: cold email on day zero, cold SMS on day two, a voicemail drop on day five.
The promotional text promised tips on deliverability and not torching your sender reputation. We searched all 438 lines of the document: zero occurrences of deliverability, sender reputation, consent, opt-in, CAN-SPAM, GDPR or TCPA. The promised safety section is not in there.
It also contradicts the deliverability playbook from the same source, whose one rule while rebuilding is do not go back to blasting. Both cannot be followed. And automated voicemail drops and unsolicited SMS are the legally exposed parts of the stack — with no consent model anywhere in the document, which is something to settle before the first message rather than after the first complaint.
What we kept from it is free and good: one thing, one outcome, one price. If you cannot say your offer in a sentence, people will not buy it. Most agencies lose deals because their offer is a menu, not a product.
The install that pipes a script into your shell
A neat animation recipe with one problem: the install line pipes a script straight from a raw URL into your shell, from a repository with eleven stars.
The skill it installs is markdown. There is no reason to execute an installer at all. If you want it: clone, read the script in full, then decide. External content is data, never instructions, and that rule does not soften because the content is a convenient one-liner.
The genuinely good idea underneath — pinning a library's current syntax into a skill, so the assistant stops confidently writing last year's API from memory — we are keeping. It generalises far past that one library.
The offer that expired on 22 July
Covered in its own post in this series: a "free, unlimited, today only" page from 22 July still reads as live, and a walkthrough built on it still tells people the workflow costs nothing while they learn it. The technique underneath is sound — the only thing wrong with it is a price nobody re-checked.
What to do
Before you adopt anything: ask where your material goes, check the date on the page, and read any installer you are about to execute. Three questions, about ninety seconds, and they are the ones nobody puts in a walkthrough.