Download references/cold-start-decision-card.md from Gingiris/gingiris-user-interview: direct link, hf CLI and curl.
- Browser
- Download file 1.93 kB
-
https://huggingface.co/datasets/Gingiris/gingiris-user-interview/resolve/main/references/cold-start-decision-card.md
- Command line
-
hf download hf://datasets/Gingiris/gingiris-user-interview/references/cold-start-decision-card.md
-
curl -L -o cold-start-decision-card.md https://huggingface.co/datasets/Gingiris/gingiris-user-interview/resolve/main/references/cold-start-decision-card.md
Cold-start decision card / 冷启动目标与成本
Use when turning research into a small acquisition or validation experiment. Set one primary decision, not a list of channels to fill.
- Segment, task, stage and eligibility; why this cohort fits.
- One initial channel, example leads with source/fit (not scraped bulk outreach), message drafts matched to the task, and one useful content example.
- Outcome: following, reply, meeting, activated user, retained use or actual payment. Define each separately, with window and deduplication key.
- Cash, tools and human hours; budget is an upper bound, not an obligation to spend. Zero cash does not mean zero cost.
- Baseline and forecast assumptions; unknown conversion/CAC/LTV stay unknown. Distinguish cash CAC from fully loaded cost and never divide by projected payers as if they were actual customers.
- Owner, time/budget cap, next review, continue/change/stop conditions.
If the original request is “paying customers” but the sensible next step is “interviews”, explain original target → proposed target → evidence/reason → question still unanswered → when to revisit payment. Do not silently replace paid acquisition with follower growth. Research drafts do not send messages without authorization.
Constructed example: third-party posts already mention an open-source tool, but interested users cannot find its author. Test an author entry and a relevant task invitation. Count replies, completed tasks and payments separately; stars, clones, followers and asking about price do not establish paid demand.
For content variants, keep project, format and landing page comparable, name the variable changed, retain IDs and measurement windows. Report counts as well as rates. Natural recommendation traffic is not a randomized experiment; tiny samples can be inconclusive. Only expand when downstream quality and the business cost limit support it.