[hsliuustc0106]

Required human contributor answers

Source: owner Hongsheng, this task (draft three repo skills; draft PR only). Later room clarification: production GPU identity is H200 / SM90, not a hard-coded L20X / sm_89 label. Each tip must name its RSI phase and the primary versus alternatives, or it is not merge-clean.

1. What outcome are you trying to achieve?
   Repo-level skills so agents measure Public231 C1, gate later optimization on a remeasure with the same protocol and the same recorded hardware, and bring in a new open model in a fixed order.

2. What is the highest-leverage variable for that outcome?
   A written, repo-accurate recipe agents load before those tasks, instead of inventing a hardware label, stacking unmeasured tips, or bundling a model into a serving issue.

3. How will this work move that variable rather than create busywork?
   Three SKILL.md files under the existing .agents/skills/ layout. No runtime code changes.

4. Who is the customer or user, and what do they actually need?
   Maintainers and coding agents on system1-omni. They need the measurement, the remeasure gate, and the new-model sequence in the repo.

Is a PR needed? Yes. Owner Hongsheng approved a draft pull request for these three skills.

Bug-fix evidence: Not a bug fix. Review withhold: do not treat a non-matching SKU as INCONCLUSIVE. The documented production target is H200 / SM90. Evidence records the measured device string and compute capability. Paired runs must share that recorded identity.

Purpose

Add three skills agents can load for measurement and model bring-up:

• .agents/skills/public231-c1/SKILL.md — Public231 C1: warmup 5, concurrency 1, three rounds, SHA, pass rate, p50/p95, the bench.py clock, both no-thinking treatments, metrics-only room posts. Every paste records device and compute capability. Examples name the production target as H200 / SM90. A different nvidia-smi string is recorded as printed and is not INCONCLUSIVE.
• .agents/skills/optimize-remeasure-gate/SKILL.md — paste p50/p95 plus device and compute capability before editing; do not stack the next tip until that tip is remeasured with the same protocol on the same recorded identity; incomplete evidence is INCONCLUSIVE. Each tip states its RSI phase (bring-up/scheduling, parallelism/comms, kernel, or launch/post) and which alternatives were considered versus the primary, or it is not merge-clean. The PR stays draft until a separate owner yes.
• .agents/skills/new-model-integration/SKILL.md — own GitHub issue, draft PR, recipe and operator checklist, then Public231 C1, then continuous optimization on that branch under the same gate. Serving-chain issues stay separate from the new-model RFC.

Placement matches the existing .agents/skills/ tree (add-new-model, self-review, precheck-pr). There is no .cursor/skills/ tree in this repo.

Test Plan

Docs-only skill files. No runtime code.

• Confirm the three paths exist and frontmatter has name and description.
• Check relative links resolve to current benchmarks/, recipe/, CONTRIBUTING.md, and sibling skills.
• Confirm commands match benchmarks/bench.py (--warmup default 3, so C1 passes --warmup 5; clock post_to_body_json_and_schema_validation; three separate measured runs).
• Confirm no skill gate marks a run INCONCLUSIVE for failing to match L20X, sm_89, H200, or SM90. Comparisons require the same protocol and the same recorded device plus compute capability.

Test Result

System1-Omni Version / Commit: c1dd645

Relative links in the three skills resolve. Claims were checked against benchmarks/README.md, benchmarks/bench.py, benchmarks/latency-leaderboard.md, and the open new-model vs serving-chain issue split (#61, #85, #125, #150 versus new model issues such as #114 and #136). No Rust or GPU run: the diff does not change runtime code.

Demo / evidence

N/A. Skill documents only; no inference, serving, or benchmark campaign in this PR.

Self-review

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.

Draft only. Do not merge and do not mark ready without a separate owner yes.

  

[hsliuustc0106]

the skill should not be spefified to a particular hardware