Getting it into your agent
This one installs as part of its plugin. Adding the marketplace and installing the plugin brings it with everything else the plugin ships.
/plugin marketplace add vibeic/vibe-ic/plugin install vibe-icWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/vibeic/vibe-ic/benchmark-enhancement-capture)<a href="https://agentmods.dev/skills/vibeic/vibe-ic/benchmark-enhancement-capture"><img src="https://agentmods.dev/badge/skills/vibeic/vibe-ic/benchmark-enhancement-capture/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/vibeic/vibe-ic/benchmark-enhancement-capture"><img src="https://agentmods.dev/badge/skills/vibeic/vibe-ic/benchmark-enhancement-capture.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00359 | $0.09926 |
| Opus 5 | $0.00179 | $0.04963 |
| Sonnet 5 | $0.00072 | $0.01985 |
| Haiku 4.5 | $0.00036 | $0.00993 |
Grade A, and why
benchmark-enhancement-capture scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 610 lines — stays where its author put it; the contents beside it link to each section on GitHub.
benchmark-enhancement-capture — the plugin's complement-and-codify loop
This skill is what makes Vibe-IC's plugin compound over time. Without it, every AI-judgment recovery — at ANY step in the flow — is a one-off that evaporates with the session. With it, every recovery has a path to permanently improve the plugin so the same fix is automatic next time.
⭐ THE BENCHMARK AGENT'S #1 MANDATE — converge → capture → distill-to-program (BINDING)
Owner directive (2026-06-22, emphatic): "請把這整套透過 Benchmark 不斷去收斂，然後 capture 這些 fail case 變成可以解，並將所有可以解的都淬鍊在我們的 program 裡面。這是 Benchmark Agent 最重要的事情了。"
The Benchmark Agent's PRIMARY, never-ending job is a LOOP, not a one-shot score:
- CONVERGE continuously. Run the benchmark; for EVERY fail, drive it to an
OFFICIAL PASS — RCA against the harness oracle in convergence mode (§3.9
oracle-for-RCA) → author → verify with the official scorer (e.g.
benchmark/score_one.py). A fail is NEVER accepted as a "floor" untilspec_coverage_check --failureproves itspec-absent. - CAPTURE every fail into a solvable case. Record
(design, prior-fail, recovered-pass, RCA)so the recovery is reproducible — never a one-off. - DISTILL every generalizable fix into the PROGRAM layer (program-first).
Turn each recovery into a deterministic rule / gate / extractor in
programs/*.py(or the gate it belongs to) so the NEXT BLIND run auto-recovers it WITHOUT an agent. This is the load-bearing step: it is what makes the benchmark number COMPOUND instead of evaporating. The gap between the blind pass@1 and the converged 100% IS the backlog of fixes still to distill.
Proven (CVDP campaign, 2026-06): each program-first fix removed a whole
failure CLASS and lifted the blind floor — PR #42 (gate multi-file split +
flat-file-map recovery), #43 (score_one official single-design scorer), #46
(Phase-1 directional-prose port extraction — fixed the #1 blind cause: empty L1
pin_table → guessed port-name/CASE → cocotb "no child object named X"). Full-302
blind reached 185/302 = 61.3% (~1.8× the published SOTA 34 %); convergence reaches
~100 %. Convergence has truly closed only when a fresh BLIND run produces 0
residual that needs a plugin fix, across TWO consecutive rounds — until then,
every loop tick MUST end by distilling its recoveries into programs via the
bucket ladder below.
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 610 lines · 359 tokens per session scan A a9d62a8d7d33
benchmark-enhancement-capture is a skill published in the GitHub repository vibeic/vibe-ic (23 stars, last pushed yesterday), licensed Apache-2.0. It adds 359 tokens to every session and 9,926 once invoked, about $0.0018 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
analog-verify
Pre-simulation review and Spectre simulation verification for analog circuits. Reviews circuit netlist and testbench, runs simulation, produces margin report. Use after analog-design completes a netlist.
step6-verify
Step 6 · Per-module simulation verification (amnesia audit → human-confirmed plan → autonomous local-fix loop → full re-verify; escalate only when local options are exhausted).
boardrepo
Read and review real PCB projects on BoardRepo. Search published hardware designs, open a board's schematic connectivity, bill of materials and files, run KiCad's DRC and ERC, and check a design against a fabrication house's limits. Use whenever the user names a BoardRepo board or URL, asks to find a published board…
analog-netlist-crawl
Crawl and analyze post-layout parasitic netlists without running SPICE. Answers "what's the effective resistance from node A to node B across this massive R mesh?", "inside the VREFN mesh, which device pins are electrically farthest apart?", "which nets have the worst coupling?", "where does settling bottleneck?" — by…
analog-design
Transistor-level circuit design for one analog sub-block. Produces Spectre netlist with hand-calculation rationale. Use when designing a specific circuit block after architecture is defined.
analog-pipeline
MANDATORY — MUST load this skill when the user mentions: OTA, ADC, PLL, comparator, bandgap, LDO, amplifier, opamp, or any analog/mixed-signal IC design task. Full analog design pipeline: spec -> architecture -> design -> verify -> deliver. Orchestrates analog-decompose, analog-behavioral, analog-design…