Borrowing it
Nothing to install: this file belongs to benchflow-ai/benchflow. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/benchflow-ai/benchflow/main/.agents/skills/benchflow-experiment-review/SKILL.mdgit clone --depth 1 https://github.com/benchflow-ai/benchflowWrote 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/benchflow-ai/benchflow/benchflow-experiment-review)<a href="https://agentmods.dev/skills/benchflow-ai/benchflow/benchflow-experiment-review"><img src="https://agentmods.dev/badge/skills/benchflow-ai/benchflow/benchflow-experiment-review/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/benchflow-ai/benchflow/benchflow-experiment-review"><img src="https://agentmods.dev/badge/skills/benchflow-ai/benchflow/benchflow-experiment-review.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 1 finding, up to medium
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- medium Agent Snooping · line 19 Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.Fix: Remove all code or instructions that list or read other skills' files or directories. Skills should operate independently; cross-skill access is a privilege escalation.
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 | $0.00099 | $0.03436 |
| Opus 5.5 | $0.00040 | $0.01374 |
| Sonnet 5.5 | $0.00020 | $0.00687 |
| Haiku 4.5 | $0.00010 | $0.00344 |
Grade A, and why
benchflow-experiment-review 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 1mo 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 — 303 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Benchflow Experiment Review
Use this skill to decide whether a Benchflow run trial is clean enough to publish or whether a Benchflow code change is safe to use for new experiments. The standard is: a clean sandbox with only task-needed resources, every agent behavior logged, no verifier leakage, and a final score or healthy failure.
Portability
This skill is intentionally harness-portable. Install or copy the entire
benchflow-experiment-review/ directory into the active harness's skill root,
keeping SKILL.md, scripts/, references/, evals/, and optional
agents/ metadata together. Use the same review procedure regardless of
whether the harness loads skills from .claude/skills, .codex/skills,
OpenHands/Gemini/pi-agent skill roots, or another compatible SKILL.md
directory.
Operating Rule
Do not accept aggregate counts alone. Enumerate the intended matrix by
task_id, harness, model, skill mode, trial id, sandbox type, run root, and
source/ref. Mark each slot as healthy, missing, duplicate, stale, or unhealthy.
Healthy run outcomes are:
pass: agent completed and verifier produced a valid score.fail: agent completed incorrectly and verifier produced a valid score.normal_timeout: agent genuinely ran, timed out, and still produced a complete ACP trajectory, complete LLM trajectory, and reward/scoring metadata.
Infrastructure failures are not healthy failures. A stalled Docker daemon,
Daytona transport failure, missing trajectory/acp_trajectory.jsonl, missing
trajectory/llm_trajectory.jsonl, malformed or empty trajectory files, missing
reward, missing timing, missing token usage for new data, or verifier crash is
unhealthy until rerun or explicitly quarantined.
Hard Trajectory Gate
For every current BenchFlow model trial, both trajectory files plus the trainer-facing rollout row are mandatory:
trajectory/acp_trajectory.jsonl: ACP/tool trace with agent-side events.trajectory/llm_trajectory.jsonl: provider LLM request/response trace with token usage evidence.results.jsonl: Verifiers / Prime-RL-shaped rollout row derived from the healthy LLM trajectory.
What ships with it
26 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.
- agents/openai.yaml 265 B
- evals/evals.json 5.6 KB
- evals/files/clean-pass/result.json 336 B
- evals/files/clean-pass/run_config.json 257 B
- evals/files/clean-pass/trajectory/acp_trajectory.jsonl 196 B
- evals/files/clean-pass/trajectory/llm_trajectory.jsonl 656 B
- evals/files/infra-timeout-missing-usage/result.json 333 B
- evals/files/infra-timeout-missing-usage/run_config.json 302 B
- evals/files/infra-timeout-missing-usage/trajectory/acp_trajectory.jsonl 290 B
- evals/files/infra-timeout-missing-usage/trajectory/llm_trajectory.jsonl 449 B
- evals/files/missing-llm-trajectory/result.json 321 B
- evals/files/missing-llm-trajectory/run_config.json 258 B
- evals/files/missing-llm-trajectory/trajectory/acp_trajectory.jsonl 196 B
- evals/files/no-skill-leak/result.json 336 B
- evals/files/no-skill-leak/run_config.json 252 B
- evals/files/no-skill-leak/trajectory/acp_trajectory.jsonl 254 B
- evals/files/no-skill-leak/trajectory/llm_trajectory.jsonl 510 B
- evals/files/reward-hack/result.json 335 B
- evals/files/reward-hack/run_config.json 249 B
- evals/files/reward-hack/trajectory/acp_trajectory.jsonl 402 B
- evals/files/reward-hack/trajectory/llm_trajectory.jsonl 528 B
- references/harness-skill-catalog-sop.md 13 KB
- references/reward-hacking-patterns.md 6.2 KB
- references/verifier-hardening-checklist.md 5.8 KB
- scripts/extract_harness_skills.py 23 KB runs code
- scripts/validate_run_artifacts.py 36 KB runs code
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.
- 1mo ago First seen · 303 lines · 99 tokens per session scan A fb5fcc4b0728
benchflow-experiment-review is a skill published in the GitHub repository benchflow-ai/benchflow (355 stars, last pushed 7d ago), licensed Apache-2.0. It adds 99 tokens to every session and 3,436 once invoked, about $0.0004 per session on Opus 5.5. 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
qa-review
QA review for code changes — test coverage analysis, edge case identification, test plan generation, regression detection, test health tracking over time.
lastlight-evals
Scaffold, configure and run a Last Light EVALS workspace — the harness that runs Last Light's real workflows against a mocked GitHub and grades them deterministically. Use when the user wants to "set up / scaffold Last Light Evals", "create an evals workspace or instance", "run evals", "compare models", or author new…
commit
Prepare and create git commits for dynobox. Use this skill whenever the user asks to commit, create a commit, suggest a commit message, stage changes, or prepare changes for review. Also use it when the user asks whether the changelog should be updated before committing.
workflow-evidence-audit
Use when auditing validation workflow evidence before a maintainer review or diagnostic handoff.
stet
Use Stet to measure whether an AI coding change is safe to ship. Trigger on model comparisons, AGENTS.md or CLAUDE.md effectiveness, shared instruction, policy, skill, harness, tool-policy, reasoning, or runtime rollouts, repo eval setup, dataset building, regression detection, benchmarking, promote/rollback…
fable-judge
Adversarial verification of finished work: re-runs the claimed verifications, diffs what changed, detects false "done" claims, delivers an evidence-based verdict (VERIFIED / VERIFIED WITH CAVEATS / REFUTED). Use after any agent or model claims work is complete — "/fable-judge", "judge this work", "verify what it did".…