What's new in tl v0.7.15: ask for a topic, get a ranked, validated channel shortlist

A new release of the tl CLI is out today: v0.7.15. The headline is a serious upgrade to how your AI agent answers one of the most common research questions you ever ask it: “find me channels that cover X.” As always, Claude Code, Gemini, Codex, and OpenCode pick it all up automatically the next time you run tl update.

From “how often does this word appear?” to “which channels actually cover this topic?”

Up until today, the tl-keyword-research skill — the one your agent reaches for whenever you ask about a topic that isn’t a pre-curated tag — was tuned for keyword counting. You’d ask about “cold plunge”, and it would come back with a tidy table of how often “cold plunge”, “ice bath”, “cold therapy”, and friends show up across video titles, summaries, and transcripts. Useful for market-sizing — but only the first half of the question you actually wanted answered.

The question you actually wanted answered is: which channels should I be talking to about this topic?

That’s now the default. When you (or your agent) ask the keyword-research skill about a topic, the headline result is no longer a keyword distribution — it’s a ranked, context-validated shortlist of channels, with each survivor flagged for sponsorability so you can move straight to outreach.

Concretely, the new flow looks like this:

1. Your agent widens the topic into a handful of related keywords — synonyms, sub-areas, common phrasings.
2. It searches the index for channels using those keywords, weighting where the keyword shows up (title beats summary beats transcript).
3. For each candidate, it spot-checks the context the keyword appears in — does “VPN” mean the privacy product, or a video-game power-up? Does “trading” mean stocks, or trading card games? — and quietly drops channels that only use the word in an unrelated sense.
4. Each survivor comes back flagged for sponsorability, so you can see at a glance which ones are already ready to pitch.

What changed under the hood is small in code but big in feel: your agent is no longer handing you raw keyword counts and leaving you to translate that into a shortlist. It’s doing the translation itself, and showing its work.

If you actually want the old keyword-count distribution (still useful for market-sizing or trend-spotting), it’s now an opt-in mode — just ask: “give me the keyword distribution for cold plunge, not channels.” Your agent will switch into counting mode on the spot.

The agent refines its query over multiple rounds — and now stops to check in

The other half of this release is about how the agent gets to that shortlist.

Topic research isn’t a one-shot search. The first set of keywords you’d reach for is almost always too wide (“investing” matches finance channels — but also sermons about “investing in your relationship”, and how-to-bet-on-sports videos that use “investing” tongue-in-cheek). The good version of this work is iterative: try a query, look at what it pulled in, narrow with AND clauses, exclude false friends with NOT, then try again.

In v0.7.15 the keyword-research skill does exactly that, automatically. It runs at least three rounds of refinement — composing and recomposing a boolean keyword expression, dropping the senses that don’t fit your topic, and tightening toward the channels you’d actually want to talk to. At the end of round three, it checks in with you before finalising — “here’s where I’ve landed, want me to tighten anything?” — so you can steer it without having to babysit the early rounds.

If you’d rather it just run end-to-end without pausing, tell it to “run autonomously” and it will. Either way, the agent makes a decision about its query a few times over, instead of trusting the first guess.

In agent-speak, conversations like these now go through cleanly on the first ask:

• “Find me channels that actually cover investing — finance investing, not sermons or sports-betting. Refine until you’re confident.”
• “Pull a shortlist of channels that really talk about cold plunges and ice baths as a wellness practice. Skip videos that only mention them in passing or as a joke. I want sponsorability flagged.”
• “I’m researching the AI-cooking niche — I need channels where both halves show up, not just generic cooking channels that name-drop ChatGPT once. Take a few rounds to narrow in. Check in with me before you finalise.”
• “Topic: TikTok Shop. Run autonomously, validate context, give me the ranked channel list when you’re done.”

You don’t ever need to write the boolean expression yourself — the agent composes and refines it for you. But it’s there if you want it, because…

The final keyword expression is now a reusable artifact

Here’s the part account managers will quietly love.

When the agent finishes refining, the boolean expression it landed on — the one that selected your final shortlist — is now saved as part of the result. It’s written in a clean, reusable form (an AND of OR-clauses, plus explicit NOT exclusions) and it can be re-run verbatim any time. Concretely that means:

• A query you spent fifteen minutes refining last Tuesday for “investing channels” can be re-run today to pick up any new channels that have entered the index since — no need to rebuild the keyword set from scratch.
• You can share the expression with a colleague and they’ll get exactly the same shortlist on their end.
• When the agent saves your work as a TL report, the expression goes with it — so the report has both the channels and the recipe that produced them.

You can think of it as: the agent isn’t just doing the research, it’s leaving you a repeatable recipe. Next quarter’s “who covers cold plunges?” question doesn’t need to start from zero.

Updating

Run tl update to pick all of this up — or just keep working, and the auto-updater will catch you up on your next command. tl changelog shows the running log of what’s landed.

Happy hunting.