Skip to content

TESTING.md

Latest commit

 

History

History
81 lines (61 loc) · 5.13 KB

TESTING.md

File metadata and controls

81 lines (61 loc) · 5.13 KB

testing the skills

Run each prompt in a chat with the noticed MCP connected. Prompts are ordered easy → hard: the first is the happy path, the rest exercise the branches that matter (ambiguity, batches, edge cases). Substitute real names from your own network where it helps.

remember-person

  1. add https://linkedin.com/in/<someone> — met at the fintech dinner, working on btc payments (clean new contact from a URL + context; should preview, then save on confirm, then read back what landed)
  2. add a few from tonight: <name> wants to advise, <name> from <company> on payments, and some sarah doing consumer ai i should follow up with (batch with mixed cases; watch that it asks about the ambiguous "sarah" in a single "need from you" zone and saves the rest)
  3. remember sarah from the ai dinner (bare name, no URL — should refuse to guess and ask for a linkedin or last name rather than silently creating a record)

event-prep

  1. who should i talk to at <Luma event URL>? (should ask your goal for the event first, then triage the roster against your network before enriching anyone)
  2. i'm going to a founders dinner tonight, here's the guest list: [paste names]. i'm raising a seed — who matters? (goal stated up front; check it tiers against "investors / people who can intro to investors" and flags in-network people prominently)
  3. After a shortlist: save the tier 1 and 2 people, tag them nytw-2026 (should hand the batch to remember-person — one preview, one confirm — and not log any "met" interaction)

event-debrief

  1. just had coffee with <name>. they're building <thing>, want to intro me to <person> at <company>, and i should send them our deck. (single-person debrief; check it logs the meeting, captures the follow-up, and treats the mentioned company as a fact, not a new person record)
  2. debrief: [paste a messy voice-note transcript from a group dinner with 3-4 people] (multi-person; watch company-name canonicalization, per-person resolution, and the grouped preview — people / follow-ups / ideas)
  3. here are my notes from the myosin dinner: [paste] (check the event tag, the event-level summary since 2+ attended, and that the readback recaps every write)

research-person

  1. research <name of someone in your network> (dossier from noticed + web, with web findings attributed inline; should proactively offer to save at the end)
  2. prep me for my meeting with <name> then reply save (check the saved note is a tight 3-5 line summary with research lines tagged unverified, and that it reads existing notes before appending)
  3. tell me about <name of someone NOT in your network> (should proceed web-only, mark them not-in-noticed, and ask whether to add them when you say save)

follow-up

  1. follow up with <name you recently met> (should read recent context, suggest an angle from it rather than asking blankly, and draft in your voice with a concrete next move)
  2. follow up with <new name>, connect with them on linkedin (LinkedIn connection note path — check the 300-character cap is enforced and the draft carries something actionable, not "let's stay connected")
  3. After a draft: sent it (should log the touchpoint — and, if it was a brand-new contact, add them to noticed at this point, then read back what was logged)

search-network

  1. do i know anyone at <company>? (specific query, runs directly, returns a table)
  2. any ai engineers? (vague — one broad dimension; should ask one clarifying question like "where, or any company in mind?" before running)
  3. how many investors are in my network? then, after any table, more on #2 (network_summary count for the first; drill-down by number into a dossier for the second)

interview-me

  1. interview me (should pre-fill role/location from your profile and present them as a confirm, then walk the questionnaire one question at a time)
  2. what does noticed know about me? (same flow; check it skips questions it already has strong signal for)
  3. Run it once, then run interview me again (idempotent — should re-ask with fresh pre-fills and append a new dated block without duplicating memory entries)

what to watch for

The behaviors the NYTW revisions were built to get right — confirm they hold across the tests above:

  • Nothing is saved before you confirm. Every write-capable skill previews first.
  • No bare-name guessing. A name with no URL/handle triggers a question, not a silent new record.
  • The readback always appears after a save, and reads like a person talking, not a field dump.
  • Provenance tags never leak into chat — you should never see [from user] or [research, unverified] in a message, only in the stored note.
  • Follow-up drafts always contain a real next step, never generic relationship filler.
  • search-network degrades gracefully if a public-scope query errors (falls back to your own network with a note).

If a skill does something it shouldn't — or misses a moment where it should have acted — log it to the 🔧 NYTW: mcp learnings page so it can be folded into the next revision.