--- name: reinvent-shortlist description: Turn 2,200 sessions into a ranked, explained shortlist to favorite in the catalog. The research-phase workhorse — use before times are published, when the user says "what should I go to", "narrow this down", "which sessions matter for me", or wants their favorites reviewed. --- # Build a re:Invent shortlist Output is a tiered list the user can act on this week: must-see, strong, maybe — each with the reason it ranked there. ## Run it > **Where to run these.** Every command below needs the planner repo on the > Python path. If there is no `reinvent/` directory in the current working > directory, you are not in it — resolve this skill's own base directory (printed > when the skill loaded, and usually a symlink), then go two levels up: that is > the repo root. `cd` there first, or call `/bin/reinvent ` by > absolute path, which works from anywhere. Do not tell the user to install > anything; there are no dependencies. ```bash python3 -m reinvent sync # if the snapshot is stale python3 -m reinvent shortlist --profile profiles/.json ``` Useful knobs: | Flag | Why | | --- | --- | | `--limit 40` | How long the list is. 40 is right for a first pass; 15 for a focused ask | | `--per-topic-cap 4` | Stops one topic swallowing the list. Raise it when the user genuinely wants one area only | | `--must-cap 15` | Size of the must-see bucket | | `--topic`, `--area`, `--service`, `--level`, `--type` | Narrow the pool before ranking, e.g. "just security" | | `--no-sponsored` | Drop vendor-sponsored content | | `--out shortlists/.md` | Save it; the user will want to re-read it on the plane | Then **read the output before you present it.** You are the second opinion, not a pipe. ## What to say when you present it Lead with the shape, not the list: how many must-sees, which topics dominate, and one thing that looks off. Then the list. Then the single next action. Three specific things worth calling out: - **Interactive vs replayable.** Workshops, builders' sessions, chalk talks and labs cannot be watched later and fill up first — those are the ones to favorite now. Breakout sessions are usually posted to YouTube; they are the cheapest thing to drop. - **Laptop-required sessions** need a laptop and a real time commitment. Flag them if the user said they would not carry one. - **Concentration.** If 15 of 20 picks are one topic, say so and ask whether that is the intent. Sometimes it is; often it means a weight is too high. ## Iterating with the user This is a conversation, not a report. Expect two or three rounds. | They say | Do | | --- | --- | | "too much AI" | Lower that topic weight in the profile, or `--exclude` the term, and re-run | | "why is this one here?" | `python3 -m reinvent explain --profile

` | | "nothing here about X" | `python3 -m reinvent search X` — if the catalog has nothing, say so plainly; more waves are coming | | "I already favorited these" | Use `--code` to score their existing list and tell them which ones are weak | | "give me only the hands-on stuff" | `--type Workshop --type "Builders' session"` | ## Optional deep pass: research session quality in parallel Scoring only knows catalog metadata. It cannot tell you that a session billed 400-level presents at 200, that it is a repeat of last year whose recording is already on YouTube, or that a technical-sounding title is a sponsor pitch. That judgement needs the web, and it fans out cleanly. Offer this when the stakes are high — a first-timer, a long trip, or reservations approaching — and when the list is longer than about 20 sessions. ```bash python3 -m reinvent shortlist --profile profiles/.json --limit 60 --json > /tmp/candidates.json ``` Split those candidates into 4–6 **disjoint** slices of ~10–12 codes and dispatch one agent per slice **in a single message** so they run concurrently. Ask each, per session code, for: - do the speakers have a track record that justifies the slot? - is this a repeat of a prior year, and if so is the recording already online? - does the abstract match its claimed level? - does it read like a sponsor pitch wearing a technical title? Require a source URL per claim and `unknown` where there is none. Tell each agent explicitly that it may **not** introduce session codes — it judges the ones you handed it. A code absent from the local snapshot does not exist. Then merge and act: drop the pitches, demote repeats whose recording exists, promote strong speakers, and if a pattern emerges adjust the profile weights and re-run. Report what the research *changed*, not that you did research. Full discipline, including when not to bother, is in [`docs/parallel-research.md`](../../docs/parallel-research.md). Do **not** fan out the scoring itself. One `shortlist` run already scores every session against the profile in well under a second; splitting it is slower and gives inconsistent weights. ## Reviewing a list the user already made Genuinely useful and often requested. Take their codes and score them: ```bash python3 -m reinvent shortlist --profile profiles/.json \ --code ANT319 --code SEC301 --code NET313 --limit 50 --min-score -99 ``` `--min-score -99` keeps low scorers visible so you can name the weak picks instead of silently dropping them. Report: which of their picks are strong, which score near zero and why, and what they are missing. ## The funnel, if they are starting from zero The approach that works, from people who do this every year: filter wide into a draft list of ~100 sessions, then cut hard down to **4 a day**. Do not try to arrive at the final list in one pass — the first pass is deliberately generous, and `--limit 40` with a loose `--min-score` is the tool's version of that. ## The close Finish with the favorite list the CLI prints — the block of codes under "Favorite list", along with the catalog URL printed above it: Tell them concretely: open that page, sign in to their re:Invent account, paste one code at a time into the search box, and star the result. Do not assume they know which catalog you mean or where the favorite control is — say it. Favoriting is free, reversible, and the only thing that makes the booking phase fast. **Say plainly that favoriting is not reserving.** A favorite is a bookmark; only the Reserve button holds a seat, and reservations open separately in the fall. People show up to full rooms holding nothing but a favorite. When that day comes, route to `reinvent-reserve`. Do not claim you have reserved anything. You have not.