Taste
Taste is a list of refusals, enforced in code.
A model writing one email cannot see the ninety-nine before it. So the decisions that need that view are made before it writes.
The dossier
Research in. Empty data out.
One record per person, built before a word is drafted.
research.person$0.050
the person record
live profile
when LinkedIn is connected
enrich.company$0.005
cached 30 days
their product
two of their pages, then research
- currentRole
- title · company · since
- organizations
- up to 8, current first
- company
- domain · industry · stage
- status
- complete | partial
4,000 characters, never trimming the current role. Cached 90 days, shared across workspaces.
location: "United States"
Refused: a place, not a person
{"status":"failed"}
Refused: a failed lookup
"… is a role based email address"
Refused: a shared inbox
"Nothing Here, Yet"
Refused: an empty profile page
The prompt
Built in a fixed order.
Dashed blocks appear only when there is something true to put in them.
The first touch knows your argument but not your product. The reply knows your product but not your argument.
The angle
Picked before the model writes.
Left to choose, the model reached for the most writable angle.
Before, on a live ledger
Now, a separate call at temperature zero
- angle A
- angle B
- angle Cfits
- angle D
- Judge fit from the PROSPECT block only.
- When none clearly fits, choose the most general angle, never a specific one that would be false for this reader.
- Ignore how well-written an angle is. Writability is not fit.
“A fabricated citation is worse than no evidence at all.”angle-synthesis.md, checked in code
- dossierkept
- replies:2kept
- github:livekept
- https://… not in the evidencedropped
The linter
Sounds like a template? Held.
The product’s own linter, run on two letters from the demo’s fictional cast.
Northwind + background job traces
Hi Priya, I came across Northwind's engineering blog and wanted to reach out — the post on queue backpressure was a great read. Tracepoint helps teams leverage per-job tracing, so incidents get faster, cheaper, and clearer. I'd love to connect if you are open to it. Best, Mira
- em-dash
- banned-opener:I-came-across
- banned-cta:love-to-chat
- ai-vocab
- rule-of-three
The line in your postmortem about turning off tracing
Hey Grace, Your postmortem says you disabled tracing on the worker fleet because per-host pricing made it untenable, and then the next incident took six hours to reconstruct. Those two sentences are the whole pitch. Tracepoint prices per job, not per host. Worth a look? Mira
CLEAR291
sends of "worth a 10-min back and forth"
120 of 312
second touches shipped a meeting ask dressed as a small one
55%
of 411 follow-ups opened "still curious"
Each count is quoted from the comment beside the rule it caused.
What this does not do
Where the rules stop.
Nothing checks the finished draft line by line
Citations are checked on the angle, not on every sentence of the email.
Your edits do not train the next email
Learning from edits exists for LinkedIn replies only.
Nothing drafts from their posts yet
Posts can be captured. No prompt reads them.
No reply-rate lift is measured
Each rule is measured against the failure it fixed, not against replies.
Every rule is a file in the repo. Fork them and argue.
