KubeCon Schedule Curator

If the ICS is malformed or missing fields, work with what you have. KubeCon's Sched exports are sometimes messy — speaker names might be in the description body rather than a structured field. Do your best to extract them.

Count the total number of sessions parsed and tell the user so they know the scale.

Step 2: Apply heuristic filters (first pass)

Score each talk using these heuristics. The goal is to separate signal from noise before the user sees anything.

Strong positive signals (boost)

• Deep Dive or Case Study in the title or tags — these are almost always higher quality
• Maintainer Track sessions — actual project contributors discussing internals and roadmap
• Speaker is a known maintainer of a CNCF project (check if the description mentions they're a maintainer, contributor, or committer)
• Speaker from a company running k8s at scale — large SaaS companies, cloud providers, platform engineering teams at scale
• Title suggests a war story — phrases like "lessons learned", "how we", "what went wrong", "in production", "at scale", "post-mortem"
• Specific technical depth — the abstract mentions concrete technologies, architectures, numbers, or trade-offs rather than vague promises

Strong negative signals (deprioritize)

• "Introduction to", "Getting Started", "101", "Beginner" — the user already knows the basics
• Sponsored Session or Sponsor Showcase — almost always marketing
• Two or more speakers from the same vendor discussing their own product — likely a product demo
• Vague buzzword-heavy abstract — "unlock the power of", "revolutionize your", "seamless integration" with no concrete details
• Lightning talks about very narrow vendor-specific tools (not general lightning talks, which can be great)

Neutral — let the user decide

• Panel discussions — hit or miss, depends on the panelists
• BoF (Birds of a Feather) — valuable for networking, less for content
• Keynotes — mixed quality but the user might want to attend for the spectacle

What NOT to filter on

Don't filter by topic area yet. The user hasn't told you what they're interested in — that's what the iterative rounds are for. Apply only quality heuristics in this pass, not topic preferences.

Step 3: Present the first batch

Show the user 10-15 of the highest-scored talks. For each talk, present:

### [Talk Title]
**Speaker(s):** Name (Company)
**Track:** Deep Dive / Case Study / etc.
**When:** Day, Time — Time
**Room:** Location

> [2-3 sentence summary of the abstract — what the talk is actually about, not the marketing fluff]

**Why this made the cut:** [brief reason — e.g., "maintainer of Cilium discussing eBPF internals" or "production war story from a team running 2000 nodes"]

Group the batch by day/time slot if possible so the user can see their schedule taking shape.

If any recommended talks overlap in time, flag it clearly:

> [!warning] Schedule conflict
> This talk overlaps with [Other Talk Title] (same time slot). Both are recommended — pick one to attend live and catch the other on YouTube later.

After presenting, ask the user for feedback. Something like: "Which of these look good? Which aren't interesting? Any topics you want to see more or less of?"

Step 4: Learn and refine

Based on the user's feedback, infer their preferences:

• If they consistently reject networking/service-mesh talks → deprioritize that topic area
• If they love the platform engineering talks → surface more of those
• If they say "too basic" about something you rated as deep → calibrate your depth threshold higher
• If they flag a speaker they know is good → look for other talks by that speaker

Apply the learned preferences and present the next batch. Keep batches to ~10 talks to avoid overwhelming the user.

Repeat until:

• The user says they're happy with their list
• You've covered all the high-quality talks in the schedule
• The user has enough talks to fill their days

Throughout the rounds, maintain a running list of accepted talks so you can:

• Flag new conflicts with already-accepted talks
• Avoid re-recommending rejected talks
• Track how full each day is getting

Step 5: Generate outputs

When the user is satisfied, produce two outputs:

Markdown schedule

Save to Fleeting/ in the vault (or wherever the user specifies) as YYYY-MM-DD_kubecon-schedule.md:

---