Context usage by target

[richardwhatever]

It'd be great to know how much context i'm using with my current set of skills on each target.

[runkids]

Thanks for the suggestion! We've been exploring this idea and built a prototype, but ran into some design questions we'd love your input on.

What we found:

The tricky part is that most AI tools load skills dynamically — not all skills are in context at once. For example, Claude Code only loads a skill's full content when its description matches the current task. So a simple "total size per target" number can be misleading, since you'd rarely hit that ceiling in practice.

What might actually be useful:

• Showing the largest individual skills so you can identify which ones to trim
• Counting only SKILL.md content (what actually gets loaded) vs. the entire skill directory
• Per-skill token estimates to help with budgeting

To help us design this well, could you share more about your situation?

1. What problem are you actually hitting? (e.g., context window errors, slow responses, skills getting ignored)
2. How many skills do you have installed, roughly?
3. Do you use different skill sets per target, or is it mostly the same across all targets?
4. Would a "top N largest skills" list be more actionable than a per-target total?

Want to make sure we build something that actually solves your problem rather than just showing numbers. 🙂

[richardwhatever]

Hi, Thanks for the response. My understanding was that all the active skill's YAML Frontmatter description field (and maybe name) are loaded into context for every request (to enable the ai to decide if a skill is relevant). If that is correct then it'd be good to know how big that sum of descriptions is. I've about 70 skills in total, maybe a max of 50 on any given target. I use different skillsets on different targets - but with some insight into context use i'd probably trim this more aggressively.

…

On Wed, 25 Mar 2026 at 11:12, Willie ***@***.***> wrote: Thanks for the suggestion! We've been exploring this idea and built a prototype, but ran into some design questions we'd love your input on. *What we found:* The tricky part is that most AI tools load skills *dynamically* — not all skills are in context at once. For example, Claude Code only loads a skill's full content when its description matches the current task. So a simple "total size per target" number can be misleading, since you'd rarely hit that ceiling in practice. *What might actually be useful:* - Showing the *largest individual skills* so you can identify which ones to trim - Counting only *SKILL.md content* (what actually gets loaded) vs. the entire skill directory - Per-skill token estimates to help with budgeting *To help us design this well, could you share more about your situation?* 1. What problem are you actually hitting? (e.g., context window errors, slow responses, skills getting ignored) 2. How many skills do you have installed, roughly? 3. Do you use different skill sets per target, or is it mostly the same across all targets? 4. Would a "top N largest skills" list be more actionable than a per-target total? Want to make sure we build something that actually solves your problem rather than just showing numbers. 🙂 — Reply to this email directly, view it on GitHub <#99?email_source=notifications&email_token=AAKSCNRNL7KTDWOQLX7D6YT4SOWHNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRTGA3TKNRVUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-16307565>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAKSCNW3M4ST2SSKZRMW2CD4SOWHNAVCNFSM6AAAAACW6AABOWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMZQG42TMNI> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[runkids]

Thanks, that's a really good point. We'll design around that. Appreciate the input!

[runkids]

Shipped in v0.18.0! 🎉

skillshare analyze is now available — it shows exactly how much context window each target consumes.

skillshare analyze               # interactive TUI (default)
skillshare analyze claude        # single target details
skillshare analyze --json        # machine-readable output

It breaks down two layers per skill:

• Always loaded — name + description (loaded every request for skill matching)
• On-demand — skill body (loaded only when triggered)

The TUI lets you switch between targets, sort by token cost, and filter by name. It also runs 7 built-in lint rules to flag quality issues (missing fields, short descriptions, no trigger phrases).

The Web UI (skillshare ui) also has a new Analyze page with charts and per-skill token breakdown.

Full docs: https://skillshare.runkids.cc/docs/reference/commands/analyze

Thanks for the suggestion — feel free to share any feedback or ideas for improvement!

[richardwhatever]

fantasic, will check it out

[richardwhatever]

hey, just an idea, but as soon as i used it (great UI by the way) i realised i had > 8k tokens always loaded, so i started searching for groups to remove (e.g. my coreyhaines-marketing skills are not relevant to certain builder agents). I'd love to be able to sum the tokens by the search result set. e.g. in the All skills table as i search it dynamically updates the results - it'd be great if it also showed the dynamic sum of tokens.

[richardwhatever]

how are you counting tokens? i belive there are different algorithms for claude vs codex for example.

[runkids]

Great idea! Dynamic token sum in search results would definitely be useful for managing context budget.

On token counting — we use a simple character-based estimation (chars ÷ 4), based on OpenAI's rule of thumb that one token generally corresponds to ~4 characters. It's not model-specific (Claude and Codex tokenizers differ), but gives a reasonable ballpark for comparing relative skill sizes.

Adding the dynamic sum to the search/filter view is on the list — thanks for the suggestion!

[runkids]

Shipped in v0.18.1! 🎉

--filter flag added — and the TUI + Web UI now show dynamic token subtotals when filtering. Should make it easy to evaluate which groups to trim.

Thanks for the idea!