Recommendations
A recommendation is only worth more than a list if it uses something the
recommender knows about this person. Memory is what makes it personal; search
is what keeps it current.
Procedure
- Recall taste before searching.
recall(query="<category> preferences", limit=20) and again for dislikes.
Pull both what they liked and, more importantly, what they bounced off — a
dislike is a sharper signal than a like.
- If memory is empty, ask two questions, not ten: one thing in this
category they loved, and one they gave up on. Then continue.
- Find current candidates with
search_web(query). Recommend from what
exists now, not from a stale training-set memory of "recent" releases. Use
fetch_page(url) on a promising list or review to get real detail.
- Rank by predicted fit, not popularity. For each of 3–5 picks, give:
- the pick, linked to the page you found it on
(
[Title](https://...)), so the user can go straight to it,
- one sentence on why it fits them — naming the specific prior taste it
connects to,
- one honest caveat ("slow first hour", "the sequel is weaker").
- Include one deliberate stretch pick and label it as such. A list that only
confirms known taste teaches the user nothing.
- Record the outcome. When the user reacts, store it:
remember(fact="<category>: liked/disliked <title> — <reason>", category="preference").
This is the step that makes the next run better; skipping it makes the skill
a search wrapper.
Rules
- Never recommend something you cannot name a concrete reason for.
- If the user rejects a pick, do not re-suggest it later — that is what step 6
prevents.
- Say when you are unsure. "I think you'll like this, but it's a stretch from
what I know" is more useful than false confidence.
Fork this
Change the category and the taste dimensions in step 4 — for restaurants,
cuisine, noise level, and price band; for gear, budget, use case, and brand
loyalty. The recall-search-rank-record loop is unchanged.
1---2name: recommendations3description: Recommend films, books, games, restaurants, or gear using what the user has already said they like and dislike, plus a web search for current options. Use when the user asks what to watch, read, play, buy, or try next.4license: MIT5version: 1.1.06metadata:7 gaia:8 security_tier: community9 permissions:10 - network:read11 tools_required:12 - recall13 - search_web14 - fetch_page15 - remember16 provenance:17 source: https://skillmd.com/skills/amd/recommendations18 upstream: starter-pack19---2021# Recommendations2223A recommendation is only worth more than a list if it uses something the24recommender knows about *this* person. Memory is what makes it personal; search25is what keeps it current.2627## Procedure28291. **Recall taste before searching.**30 `recall(query="<category> preferences", limit=20)` and again for dislikes.31 Pull both what they liked and, more importantly, what they bounced off — a32 dislike is a sharper signal than a like.332. **If memory is empty, ask two questions**, not ten: one thing in this34 category they loved, and one they gave up on. Then continue.353. **Find current candidates** with `search_web(query)`. Recommend from what36 exists now, not from a stale training-set memory of "recent" releases. Use37 `fetch_page(url)` on a promising list or review to get real detail.384. **Rank by predicted fit**, not popularity. For each of 3–5 picks, give:39 - the pick, linked to the page you found it on40 (`[Title](https://...)`), so the user can go straight to it,41 - **one sentence on why it fits *them*** — naming the specific prior taste it42 connects to,43 - one honest caveat ("slow first hour", "the sequel is weaker").445. **Include one deliberate stretch pick** and label it as such. A list that only45 confirms known taste teaches the user nothing.466. **Record the outcome.** When the user reacts, store it:47 `remember(fact="<category>: liked/disliked <title> — <reason>", category="preference")`.48 This is the step that makes the next run better; skipping it makes the skill49 a search wrapper.5051## Rules5253- Never recommend something you cannot name a concrete reason for.54- If the user rejects a pick, do not re-suggest it later — that is what step 655 prevents.56- Say when you are unsure. "I think you'll like this, but it's a stretch from57 what I know" is more useful than false confidence.5859## Fork this6061Change the category and the taste dimensions in step 4 — for restaurants,62cuisine, noise level, and price band; for gear, budget, use case, and brand63loyalty. The recall-search-rank-record loop is unchanged.