{ "measurement": "Whether the language a Claude Code skill description is written in changes the token cost of having that skill installed", "claude_code_version": "2.1.247 (Claude Code)", "why": "Does the LANGUAGE a skill description is written in change what it costs to have installed?\n\nWHY THIS ONE. ConardLi/garden-skills reached GitHub Trending at 11,112 stars. Pricing it with\nthis site's published skill model gave a number 2.5x under the measurement, and the pack's\ndescriptions are roughly half Chinese. The site's model was fitted on English prose, so the\nobvious candidate was the tokeniser, which limit 2 of skillpack-cost-2-1-233 named as an\nuntested explanation for that corpus's unexplained model gap.\n\nTHE ARMS. Every cell is ONE skill installed alone, measured against a floor from the SAME round.\n CJK the four garden-skills descriptions that are majority Chinese\n English the one garden-skills description that is not, plus four addyosmani/agent-skills\n descriptions chosen to SPAN the CJK arm's character range\n\nWHY THE ENGLISH ARM IS BUILT THAT WAY. Comparing two packs changes author, house style, length\nand language at once. Keeping garden-skills' own English skill in the arm gives a within-pack,\nwithin-author comparison, and picking the addyosmani cells by character length makes the two\narms' length ranges overlap, so length cannot carry the difference.\n\nCONTROLS\n PAIRED floor re-measured every round; every delta is against that round's floor.\n ARRIVAL the reply must be exactly OK, published per run as reply_exact_ok.\n INTERLEAVED all nine cells are measured inside the same rounds, so machine drift cannot\n land on one arm.\n ISOLATION --setting-sources project asserted onto the command line by the runner.\n FIXTURE skills counted back off disk per run.\n BODIES NEVER INVOKED only the listing is priced; no third-party body is executed.\nESTIMATOR guarded mode, from the shared estimator. This harness carries an intermittent\n component that fires per call; three cells lost one round each to it.\nPOPULATION one skill installed alone, on this machine, at the recorded release.", "method": { "population": "one skill installed alone in an otherwise empty project, priced against an empty floor measured in the same round", "estimator": "guarded mode across 4 rounds (shared estimator.py)", "prompt": "Reply with only: OK", "arms": { "CJK": "description majority Chinese (cjk_fraction > 0.4)", "English": "description contains no CJK" }, "controls": [ "paired floor per round", "arrival control: reply exactly OK", "all cells interleaved within each round", "--setting-sources project asserted onto the command line", "skills counted back off disk per run", "no skill body ever invoked" ] }, "results": { "headline": { "english_tokens_per_description_char": { "n": 5, "min": 0.3629, "max": 0.4819 }, "cjk_tokens_per_description_char": { "n": 4, "min": 0.8867, "max": 1.0479 }, "arms_overlap": false, "closest_approach_ratio": 1.84, "english_tokens_per_description_byte": { "n": 5, "min": 0.3629, "max": 0.4819 }, "cjk_tokens_per_description_byte": { "n": 4, "min": 0.4593, "max": 0.5129 }, "per_byte_arms_overlap": true, "what_that_means": "Per CHARACTER the arms are 2.2x apart and do not overlap. Per BYTE they overlap. So most of the per-character gap is a unit artefact: a Chinese character is about 2.3 UTF-8 bytes here. It still matters, because the description limit Anthropic documents counts CHARACTERS, so a Chinese description reaches that cap having spent roughly twice the tokens. What is NOT supported is that the tokeniser treats Chinese badly per unit of encoded text.", "length_ranges_overlap": true, "english_description_chars_range": [ 198, 485 ], "cjk_description_chars_range": [ 167, 518 ] }, "the_model_check": { "model": "39 tokens + 0.276 per description character", "model_published_valid_range_chars": [ 30, 300 ], "english_predicted": 668, "english_measured": 703, "english_ratio": 1.05, "cjk_predicted": 539, "cjk_measured": 1339, "cjk_ratio": 2.48 }, "cells": { "eng-browser-testing-with-devtools": { "arm": "English", "tokens": 145, "description_chars": 326, "description_bytes": 326, "cjk_fraction": 0.0, "tokens_per_description_char": 0.4448, "tokens_per_description_byte": 0.4448, "rounds": 4, "mode_seen_in_rounds": 4, "all_rounds": [ 145, 145, 145, 145 ], "spread": 0, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": true }, "eng-context-engineering": { "arm": "English", "tokens": 77, "description_chars": 198, "description_bytes": 198, "cjk_fraction": 0.0, "tokens_per_description_char": 0.3889, "tokens_per_description_byte": 0.3889, "rounds": 4, "mode_seen_in_rounds": 4, "all_rounds": [ 77, 77, 77, 77 ], "spread": 0, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": false }, "eng-debugging-and-error-recovery": { "arm": "English", "tokens": 120, "description_chars": 249, "description_bytes": 249, "cjk_fraction": 0.0, "tokens_per_description_char": 0.4819, "tokens_per_description_byte": 0.4819, "rounds": 4, "mode_seen_in_rounds": 3, "all_rounds": [ -777, 120, 120, 120 ], "spread": 897, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": false }, "eng-interview-me": { "arm": "English", "tokens": 176, "description_chars": 485, "description_bytes": 485, "cjk_fraction": 0.0, "tokens_per_description_char": 0.3629, "tokens_per_description_byte": 0.3629, "rounds": 4, "mode_seen_in_rounds": 3, "all_rounds": [ -721, 176, 176, 176 ], "spread": 897, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": true }, "solo-beautiful-article": { "arm": "CJK", "tokens": 368, "description_chars": 415, "description_bytes": 801, "cjk_fraction": 0.531, "tokens_per_description_char": 0.8867, "tokens_per_description_byte": 0.4594, "rounds": 4, "mode_seen_in_rounds": 4, "all_rounds": [ 368, 368, 368, 368 ], "spread": 0, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": true }, "solo-gpt-image-2": { "arm": "CJK", "tokens": 299, "description_chars": 289, "description_bytes": 583, "cjk_fraction": 0.634, "tokens_per_description_char": 1.0346, "tokens_per_description_byte": 0.5129, "rounds": 4, "mode_seen_in_rounds": 4, "all_rounds": [ 299, 299, 299, 299 ], "spread": 0, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": false }, "solo-kb-retriever": { "arm": "CJK", "tokens": 175, "description_chars": 167, "description_bytes": 381, "cjk_fraction": 0.658, "tokens_per_description_char": 1.0479, "tokens_per_description_byte": 0.4593, "rounds": 4, "mode_seen_in_rounds": 4, "all_rounds": [ 175, 175, 175, 175 ], "spread": 0, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": false }, "solo-web-design-engineer": { "arm": "English", "tokens": 185, "description_chars": 455, "description_bytes": 455, "cjk_fraction": 0.0, "tokens_per_description_char": 0.4066, "tokens_per_description_byte": 0.4066, "rounds": 4, "mode_seen_in_rounds": 4, "all_rounds": [ 185, 185, 185, 185 ], "spread": 0, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": true }, "solo-web-video-presentation": { "arm": "CJK", "tokens": 497, "description_chars": 518, "description_bytes": 1074, "cjk_fraction": 0.613, "tokens_per_description_char": 0.9595, "tokens_per_description_byte": 0.4628, "rounds": 4, "mode_seen_in_rounds": 3, "all_rounds": [ -400, 497, 497, 497 ], "spread": 897, "reply_exact_ok_all_rounds": true, "model_applied_outside_its_published_range": true } }, "whole_pack_vs_sum_of_parts": { "garden_skills_installed_together": 1454, "sum_of_the_five_measured_alone": 1524, "difference": -70, "note": "installing the five together costs less than pricing them one at a time, so a per-skill figure multiplied by a skill count overstates a pack. No cause is asserted." }, "release_bridge": { "pack": "addyosmani/agent-skills", "measured_here": 2302, "measured_on_2_1_233": 2302, "identical": true, "why_it_is_here": "the same pack measured on 2.1.233 and again on 2.1.247 returns the identical total, so the garden-skills figures are a fact about the pack rather than release drift." } }, "known_limits": { "CARRY FORWARD listing_cost_only": "This prices having the skill installed. It does not price using it: a body is charged when the skill is invoked and nothing here invokes one. A Chinese-language BODY is likely to carry the same tokeniser penalty at a far larger character count, and that is NOT measured here.", "CARRY FORWARD a_pack_is_a_moving_target": "garden-skills is measured at commit aaf9a82 and addyosmani/agent-skills at 7cb7a20. addyosmani was pushed on the day of this run, so its fixture is not the one the 2.1.233 corpus walked even though the total matched.", "CARRY FORWARD one_machine_one_model": "Opus on Claude Code 2.1.247, project scope only, one machine.", "CARRY FORWARD the_injection_control_detects_only_loud_payloads": "Every reply across every round was exactly OK, published per run. That shows no description changed the literal output. It does NOT show none influenced the model, and these are third-party descriptions in two languages.", "CARRY FORWARD model_applied_outside_its_published_range": "The model is declared valid for 30 to 300 characters. Six of the nine cells exceed 300 and are flagged per cell. For those the model comparison is indicative only. This matters MORE here, not less: the 2.48x CJK miss is not attributable to the tokeniser alone while part of the arm sits outside the fitted range.", "PARTIALLY RETIRES the_model_gap_is_unexplained": "That limit listed 'real descriptions tokenising differently from the prose fixture the model was fitted on' as an untested candidate for the model's unexplained miss. This corpus tests it and finds it real for Chinese: the model lands within 5% on the English arm and 2.48x under on the CJK arm. It is PARTIAL because the 2.1.233 corpus also missed on packs that are entirely English, which this does not explain.", "PARTIALLY ADDRESSES per_skill_cost_does_not_order_by_description_length": "That limit recorded per-skill cost failing to order by description length across packs with no single-variable explanation. Language is now one measured cause: the 455-character English description costs 185 tokens and the 167-character Chinese one costs 175. It is not offered as the only cause.", "RETIRED single_skill_packs_are_inside_the_noise": "That limit said a single skill cannot be resolved against the floor. It does not hold at these description lengths: ALL NINE single-skill cells returned a spread of zero across the rounds that fired (six of nine were zero-spread including the misfires), the cells are 77 to 497 tokens, and the floor read 22,900 in every round. It IS still true for short descriptions; these are long.", "the_english_arm_mixes_two_authors": "Four of the five English cells come from addyosmani/agent-skills and one from garden-skills itself. The within-pack comparison (garden's own English skill against its four Chinese ones) is the cleaner one and points the same way; the cross-pack cells widen the arm and make the length ranges overlap.", "cjk_fraction_is_a_crude_instrument": "The arm split is a character-class count over the description, thresholded at 0.4. Every cell here is either 0% or 53-66%, so nothing sits near the threshold, but the measure does not distinguish Chinese from Japanese kanji and no cell tests a mixed script near 40%.", "one_intermittent_component_cost_three_cells_a_round": "Three of nine cells lost exactly one round each to this harness's known intermittent component, which reads as a large negative delta. The guarded mode carried each on 3 of 4 rounds and every retained value is published in all_rounds. The five garden-skills cells additionally reproduced identically in an earlier independent run.", "the_intermittent_component_is_exactly_897_tokens_every_time": "The three misfires are not noise of varying size. Each is exactly 897 tokens below the cell's mode (120 to -777, 176 to -721, 497 to -400). That is the same 897 the skills curve corpus recorded on the same day at n=300, on a different fixture. Two harnesses, one component, identical magnitude. No cause is asserted; it is recorded because a fixed-size component is a different thing from noise and should be measured, not averaged away.", "THE UNIT CARRIES MOST OF THE GAP, AND THE HEADLINE IS PER CHARACTER": "Per character the arms are 2.2x apart and do not overlap. Per BYTE of UTF-8 they OVERLAP: English 0.363 to 0.482, Chinese 0.459 to 0.513. A Chinese character is about 2.3 bytes here, and that accounts for most of the per-character difference. The per-character figure is still the one that matters in practice, because the description cap Anthropic documents counts characters and this site prices skills per character, so a Chinese description reaches the cap having spent roughly twice the tokens. But it does NOT support the claim that the tokeniser handles Chinese badly per unit of encoded text, and any wording implying Chinese costs 2.4x for the same amount of TEXT is wrong. This site prices skills per character and CLAUDE.md per byte, so the same content can look cheaper or dearer depending only on which page is read.", "no_cause_is_asserted_below_the_tokeniser": "The measurement is that Chinese descriptions cost about 2.2 to 2.9x per character. Why the tokeniser behaves that way, and whether it is byte-level fallback on CJK codepoints, is not measured here and is not claimed." }, "runs": [ { "pack": "floor", "round": 0, "status": "ok", "skills_on_disk": 0, "context": 22900, "delta": 0, "reply_exact_ok": true }, { "pack": "eng-browser-testing-with-devtools", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 326, "mean": 326, "max": 326, "total_chars": 326 }, "context": 23045, "delta": 145, "per_skill": 145.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-context-engineering", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 198, "mean": 198, "max": 198, "total_chars": 198 }, "context": 22977, "delta": 77, "per_skill": 77.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-debugging-and-error-recovery", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 249, "mean": 249, "max": 249, "total_chars": 249 }, "context": 23020, "delta": 120, "per_skill": 120.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-interview-me", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 485, "mean": 485, "max": 485, "total_chars": 485 }, "context": 23076, "delta": 176, "per_skill": 176.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-beautiful-article", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 415, "mean": 415, "max": 415, "total_chars": 415 }, "context": 23268, "delta": 368, "per_skill": 368.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-gpt-image-2", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 289, "mean": 289, "max": 289, "total_chars": 289 }, "context": 23199, "delta": 299, "per_skill": 299.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-kb-retriever", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 167, "mean": 167, "max": 167, "total_chars": 167 }, "context": 23075, "delta": 175, "per_skill": 175.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-design-engineer", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 455, "mean": 455, "max": 455, "total_chars": 455 }, "context": 23085, "delta": 185, "per_skill": 185.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-video-presentation", "round": 0, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 518, "mean": 518, "max": 518, "total_chars": 518 }, "context": 23397, "delta": 497, "per_skill": 497.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "floor", "round": 1, "status": "ok", "skills_on_disk": 0, "context": 22900, "delta": 0, "reply_exact_ok": true }, { "pack": "eng-browser-testing-with-devtools", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 326, "mean": 326, "max": 326, "total_chars": 326 }, "context": 23045, "delta": 145, "per_skill": 145.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-context-engineering", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 198, "mean": 198, "max": 198, "total_chars": 198 }, "context": 22977, "delta": 77, "per_skill": 77.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-debugging-and-error-recovery", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 249, "mean": 249, "max": 249, "total_chars": 249 }, "context": 22123, "delta": -777, "per_skill": -777.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-interview-me", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 485, "mean": 485, "max": 485, "total_chars": 485 }, "context": 23076, "delta": 176, "per_skill": 176.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-beautiful-article", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 415, "mean": 415, "max": 415, "total_chars": 415 }, "context": 23268, "delta": 368, "per_skill": 368.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-gpt-image-2", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 289, "mean": 289, "max": 289, "total_chars": 289 }, "context": 23199, "delta": 299, "per_skill": 299.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-kb-retriever", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 167, "mean": 167, "max": 167, "total_chars": 167 }, "context": 23075, "delta": 175, "per_skill": 175.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-design-engineer", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 455, "mean": 455, "max": 455, "total_chars": 455 }, "context": 23085, "delta": 185, "per_skill": 185.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-video-presentation", "round": 1, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 518, "mean": 518, "max": 518, "total_chars": 518 }, "context": 23397, "delta": 497, "per_skill": 497.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "floor", "round": 2, "status": "ok", "skills_on_disk": 0, "context": 22900, "delta": 0, "reply_exact_ok": true }, { "pack": "eng-browser-testing-with-devtools", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 326, "mean": 326, "max": 326, "total_chars": 326 }, "context": 23045, "delta": 145, "per_skill": 145.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-context-engineering", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 198, "mean": 198, "max": 198, "total_chars": 198 }, "context": 22977, "delta": 77, "per_skill": 77.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-debugging-and-error-recovery", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 249, "mean": 249, "max": 249, "total_chars": 249 }, "context": 23020, "delta": 120, "per_skill": 120.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-interview-me", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 485, "mean": 485, "max": 485, "total_chars": 485 }, "context": 23076, "delta": 176, "per_skill": 176.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-beautiful-article", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 415, "mean": 415, "max": 415, "total_chars": 415 }, "context": 23268, "delta": 368, "per_skill": 368.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-gpt-image-2", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 289, "mean": 289, "max": 289, "total_chars": 289 }, "context": 23199, "delta": 299, "per_skill": 299.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-kb-retriever", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 167, "mean": 167, "max": 167, "total_chars": 167 }, "context": 23075, "delta": 175, "per_skill": 175.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-design-engineer", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 455, "mean": 455, "max": 455, "total_chars": 455 }, "context": 23085, "delta": 185, "per_skill": 185.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-video-presentation", "round": 2, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 518, "mean": 518, "max": 518, "total_chars": 518 }, "context": 22500, "delta": -400, "per_skill": -400.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "floor", "round": 3, "status": "ok", "skills_on_disk": 0, "context": 22900, "delta": 0, "reply_exact_ok": true }, { "pack": "eng-browser-testing-with-devtools", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 326, "mean": 326, "max": 326, "total_chars": 326 }, "context": 23045, "delta": 145, "per_skill": 145.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-context-engineering", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 198, "mean": 198, "max": 198, "total_chars": 198 }, "context": 22977, "delta": 77, "per_skill": 77.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-debugging-and-error-recovery", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 249, "mean": 249, "max": 249, "total_chars": 249 }, "context": 23020, "delta": 120, "per_skill": 120.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "eng-interview-me", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 485, "mean": 485, "max": 485, "total_chars": 485 }, "context": 22179, "delta": -721, "per_skill": -721.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-beautiful-article", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 415, "mean": 415, "max": 415, "total_chars": 415 }, "context": 23268, "delta": 368, "per_skill": 368.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-gpt-image-2", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 289, "mean": 289, "max": 289, "total_chars": 289 }, "context": 23199, "delta": 299, "per_skill": 299.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-kb-retriever", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 167, "mean": 167, "max": 167, "total_chars": 167 }, "context": 23075, "delta": 175, "per_skill": 175.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-design-engineer", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 455, "mean": 455, "max": 455, "total_chars": 455 }, "context": 23085, "delta": 185, "per_skill": 185.0, "reply_exact_ok": true, "reply": "OK" }, { "pack": "solo-web-video-presentation", "round": 3, "status": "ok", "skills_on_disk": 1, "description_stats": { "n": 1, "median": 518, "mean": 518, "max": 518, "total_chars": 518 }, "context": 23397, "delta": 497, "per_skill": 497.0, "reply_exact_ok": true, "reply": "OK" } ] }