feat(leads): open on something true about the reader, and guess an inbox last - #146
Merged
Merged
Conversation
…box last Modelled on a cold email that actually earned a reply. What made it work was the first sentence: a specific, checkable observation about the recipient, not a compliment and not a description of the sender. The custom-pitch prompt forbade exactly that. It said to state nothing about the recipient's business as fact, which was the right instinct with the wrong remedy — without research "I loved your work" is a fabrication, so the rule prevented hallucination by banning the strongest opening available. The remedy is research rather than silence. A prospect's own homepage title and description are fetched and quoted into the prompt, and the model may refer to that and nothing else. Only what the site says about itself is used: no inference, no summarising of prose, nothing the reader could not verify by looking at their own homepage. The prompt names the real failure mode, which is not inventing a company but extending one true sentence into a paragraph of assumed detail about their customers, traffic and roadmap. Praise is still refused. An observation is "you focus on X"; a compliment is "impressive product", and praise from a stranger reads as a form letter because it is one. The draft now has a shape — observation, shared problem, what is offered, one small ask — rather than a list of prohibitions and a word limit. One thing from the example is deliberately not copied. Its subject line promised a quick question and asked none; the reply that earns is owed to the trick rather than the offer, so the prompt refuses subjects that promise what the body does not deliver. Guessed shared inboxes are a last resort after crawling the site and searching for it have both failed. They are marked as guesses, rank below any address actually published, and can never be one of the addresses the never-contact list exists to protect — guessing careers@ would manufacture precisely the sends that list prevents. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
vu1nz Security Review0 finding(s) in PR #? No security issues found. |
Contributor
Author
|
in the install.sh curl installer, what I do is i install mise and activate it for this session then install whatever dependencies I need (ie: node v26 LTS). don't rely on me having shit setup correct (most devs have no experience with command line). I've seen an engineer at IBM who install everything into $HOME and me and another guy were cracking up...."how does he manage this mess??" |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Modelled on a cold email that actually earned a reply. What made it work was the first sentence: a specific, checkable observation about the recipient — not a compliment, not a description of the sender.
The prompt forbade exactly that
customDraftSystemsaid:Right instinct, wrong remedy. Without research, "I loved your work" is a fabrication — so the rule prevented hallucination by banning the strongest opening available.
Fixed with research, not silence. The prospect's own homepage title/description is fetched and quoted into the prompt; the model may refer to that and nothing else. Only what the site says about itself — no inference, nothing the reader couldn't verify on their own homepage.
The prompt names the real failure mode, which isn't inventing a company but extending one true sentence into assumed detail about their customers, traffic, and roadmap.
Praise is still refused: "you focus on X" is an observation; "impressive product" is a form letter.
The draft now has a shape
Observation → shared problem → what's offered, concretely → one small ask. Previously it was a list of prohibitions and a word limit.
One thing deliberately not copied
The example's subject line promised "quick question" and asked none. The reply that earns is owed to the trick rather than the offer, so the prompt refuses subjects that promise what the body doesn't deliver.
Guessed inboxes, last
hello@/info@/contact@/press@— only after crawling and searching have failed. (Fittingly,press@is how the example reached us.)source: "guess"so a human can tellcareers@would manufacture precisely the sends that list preventsChecks
tsc --noEmitcleanTwo prior tests asserted the rule this changes; both updated to the new contract rather than deleted.
🤖 Generated with Claude Code