Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Internal — invoked headlessly by the peon eval server; humans should run `peon eval <pack>` instead. Execute one PeonPing reroll job — invoked headlessly by the peon eval server as: "Use the peon-ping-remix skill to execute the reroll job at <path>". Reads the job JSON (scope, category, index, caption), rewrites the affected sound prompt(s) to honor the caption, re-renders via scripts/pack-render.py, and logs the change to eval-log.json.
.claude/skills/peonping-peon-ping-remix/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-16 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 31% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 15% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -11% | 0% |
You are executing ONE reroll job for a draft PeonPing pack. Work only inside the job's draft_dir. Do not touch any other directory, do not install anything, do not commit anything.
This skill is internal: it is only ever invoked headlessly by the peon eval server with a job-file path. If you were invoked WITHOUT one (a human ran this directly), do nothing except tell them this skill runs automatically from the eval server and that they should run peon eval <pack> instead. Do not touch any files. Exit.
scope ("sound" or"pack"), category, index, caption, draft_dir.
<draft_dir>/openpeon.json and <draft_dir>/prompts.json. Ifprompts.json is missing, fail with a clear message — you cannot reroll without the current prompts.
categories[category].sounds[index].file; scope "pack" → every file in every category.
context). The caption is a correction, not a reset: "too harsh, want softer" means adjust intensity, keep the instrument palette unless the caption says otherwise.
prompt. tts entries: rewrite text (keepvoice_id — never change a pack's voice in a reroll).
<draft_dir>/jobs/ (e.g. <draft_dir>/jobs/render-job-<category>_<index>.json — approve prunes everything under jobs/, so a render input left at the draft root would ship as junk in the approved pack) and run python3 <peon-ping>/scripts/pack-render.py --job <file> with out set to the target WAV path (overwrite in place). The renderer exits nonzero on silent/failed renders — if it fails, STOP and exit nonzero yourself with its stderr; do not half-update the log. (Resolve <peon-ping> as the scripts directory next to the running peon install, in order: $PEON_DIR/scripts when set, else ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/hooks/peon-ping/scripts, else the repo checkout you were invoked from.)
prompts.json with the newinputs, and append one entry per rendered sound to <draft_dir>/eval-log.json (create as [] if missing): {"ts": "<ISO8601>", "scope", "category", "index", "caption", "old_prompt", "new_prompt", "file"}.
draft_dir.into a same-directory temp file and only replaces the out path (atomic os.replace) on a successful, non-silent render; do not delete a WAV before rendering.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 5,367 | 7,696 | +43% | 1 | 1 | 0% | 276 | 1,139 | +313% | 0 | 0 | — |
case-02 | fail→fail | 5,699 | 6,959 | +22% | 1 | 1 | 0% | 390 | 1,536 | +294% | 0 | 0 | — |
case-16 | fail→pass | 8,783 | 3,195 | -64% | 1 | 1 | 0% | 1,301 | 1,319 | +1% | 0 | 0 | — |
case-03 | fail→pass | 5,404 | 1,739 | -68% | 1 | 1 | 0% | 892 | 1,168 | +31% | 0 | 0 | — |
case-04 | fail→pass | 17,671 | 3,088 | -83% | 1 | 1 | 0% | 2,727 | 1,453 | -47% | 0 | 0 | — |
case-05 | pass→fail | 18,411 | 2,761 | -85% | 1 | 1 | 0% | 3,754 | 1,396 | -63% | 0 | 0 | — |
case-06 | pass→fail | 11,309 | 2,246 | -80% | 1 | 1 | 0% | 2,098 | 1,281 | -39% | 0 | 0 | — |
case-07 | fail→pass | 9,491 | 5,413 | -43% | 1 | 1 | 0% | 1,597 | 1,832 | +15% | 0 | 0 | — |
case-08 | fail→pass | 9,751 | 3,717 | -62% | 1 | 1 | 0% | 1,721 | 1,526 | -11% | 0 | 0 | — |
case-09 | fail→pass | 10,340 | 2,528 | -76% | 1 | 1 | 0% | 1,735 | 1,206 | -30% | 0 | 0 | — |
case-10 | fail→fail | 11,170 | 4,791 | -57% | 1 | 1 | 0% | 1,715 | 1,662 | -3% | 0 | 0 | — |
case-11 | fail→pass | 10,217 | 1,748 | -83% | 1 | 1 | 0% | 1,466 | 1,201 | -18% | 0 | 0 | — |
case-17 | pass→pass | 4,438 | 1,790 | -60% | 1 | 1 | 0% | 611 | 1,114 | +82% | 0 | 0 | — |
case-12 | fail→pass | 11,033 | 13,593 | +23% | 1 | 1 | 0% | 2,138 | 3,473 | +62% | 0 | 0 | — |
case-13 | fail→pass | 13,990 | 4,741 | -66% | 1 | 1 | 0% | 2,399 | 1,535 | -36% | 0 | 0 | — |
case-14 | fail→pass | 12,683 | 2,529 | -80% | 1 | 1 | 0% | 2,350 | 1,360 | -42% | 0 | 0 | — |
case-15 | fail→fail | 9,750 | 4,251 | -56% | 1 | 1 | 0% | 1,869 | 1,760 | -6% | 0 | 0 | — |
case-18 | pass→pass | 5,935 | 10,853 | +83% | 1 | 1 | 0% | 864 | 2,965 | +243% | 0 | 0 | — |
case-19 | pass→fail | 12,744 | 7,252 | -43% | 1 | 1 | 0% | 1,817 | 2,241 | +23% | 0 | 0 | — |
case-20 | pass→fail | 12,954 | 1,972 | -85% | 1 | 1 | 0% | 2,156 | 1,158 | -46% | 0 | 0 | — |
case-21 | fail→fail | 7,530 | 4,751 | -37% | 1 | 1 | 0% | 1,429 | 1,824 | +28% | 0 | 0 | — |
case-22 | fail→fail | 14,438 | 4,812 | -67% | 1 | 1 | 0% | 2,464 | 1,814 | -26% | 0 | 0 | — |
case-23 | pass→pass | 8,812 | 2,919 | -67% | 1 | 1 | 0% | 1,268 | 1,325 | +4% | 0 | 0 | — |
case-24 | fail→fail | 15,364 | 3,430 | -78% | 1 | 1 | 0% | 2,472 | 1,475 | -40% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted, and 22 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +25 percentage points is the difference between those two pass rates over the 22 comparable cases. 4 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.