[jayleaton]

What this changes

Before: a contributor agent had no procedure for preparing a pull request, so descriptions reached review without the issue link, the overlap check, the lean delta or platform coverage. After: docs/pr/SKILL.md (tensorfold-pull-request) is a seven-step pre-flight an agent runs before gh pr create and again after each revision, and AGENTS.md and README.md#contributing-with-agents each gain one line pointing to it.

Why: reviewers kept asking for the same missing pieces; every rule comes from CONTRIBUTING.md, the kernel contract, the ATC contributor skill or a survey of recent pull requests.

• Review pass: step 4 asks for a planted control for each new kernel that fails as planted, with the backend's SUMMARY line showing it red, and for proof and an opt-in before a change to output tokens on a default path; step 3 gates decode and prompt processing at one stream and the widest admitted, with and without --drafter; and the lines that restated CONTRIBUTING now link its exactness and precision, lean code and review and public identity and privacy sections.
• Issue: n/a. Maintainer-requested repository process work; no issue tracks it.
• Stacked on: Issue forms and a sharper pull request template #1596; merge Issue forms and a sharper pull request template #1596 first. This branch adds two commits, 3e520607 (the skill and its two pointer lines) and f15b510a (the review wording), on top of Issue forms and a sharper pull request template #1596's 462f4a60, and it was rebased onto that head so the stack stays in order. docs/pr/SKILL.md, AGENTS.md and README.md are the whole of it. The base stays main because Issue forms and a sharper pull request template #1596's head lives on a fork and GitHub only accepts a base branch in this repository, so until Issue forms and a sharper pull request template #1596 lands the GitHub diff also lists Issue forms and a sharper pull request template #1596's six .github/ files. Review 462f4a60..f15b510a for this change alone.
• atc overlap --json (ATC 0.7.0, against main): complete: false with pullFilesLimit, so the result is partial. It lists 11 open pull requests sharing README.md (minimax_m3: serve MiniMax-M3.1-Flash (428B-A23B MoE, images) natively on Metal #1235, minimax_m3: a CUDA backend for MiniMax-M3.1-Flash at TP over GB10s, from the release's own formats #1388, lfm2: draft with LiquidAI's DSpark drafter (--drafter) #1391, lfm2: the DSpark confidence head schedules the verify #1392, lfm2: the DSpark drafter loads one matrix at a time and keeps a lane in one buffer #1393, lfm2: a cheaper DSpark hold (one concurrent encoder) #1394, minimax_m3 (CUDA): draft several streams' DSpark blocks in one pass #1426, lfm2_moe: serve LFM2.5-24B-A2B on CUDA (plain serve, shared kernels) #1492, lfm2_moe CUDA: attention pages from the shared KV pool, and the prompt cache (stacked on #1492) #1503, lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508, lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634), each unknown because ATC could not fetch their files; no conflict is known, and none is claimed here.

Checks on this head

• Head commit: f15b510
• Lean checker, python3 tools/zig/lean_check.py, on the merge base (465e48b62) and on this head: 425 and 425, 0 new (tools/zig/lean_new.py). Both runs exit 1 on the 425 findings already on main.
• Tests run, with the command and pass, fail and skip counts:
  • relative links and anchors in docs/pr/SKILL.md, AGENTS.md and README.md: 88 checked, 88 resolve, 0 broken (external links excluded).
  • git diff --check: clean.
  • the skill's kernel commands exist and match docs/kernel-contract.md: zig build kernel-contract, kernel-contract-fatbins (zig/build/kernel_contract.zig), hip-kernel-contract and hip-kernel-contract-limits (zig/build/hip_contract.zig), tf-cuda-test kernel-contract-limits (zig/tests/cuda/main.zig).
• CI: 9/9 checks pass on this head (fmt, zig, zig-windows, zig-macos-dev-tools, appimage-linux, desktop, desktop-linux, mobile, release-paths); no failure.
• Not run locally, and why: the Zig host suite, test-golden and zig fmt, because no Zig source changed and no workflow checks markdown; CI's zig job runs on this head.

Receipt

n/a: documentation only; it cannot reach inference.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (none added; the link check above is the check)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks for this. The forms and the pre-flight read well, and the commands they cite match main. Could you add a short description, and say how this relates to #1596? The six .github files here are identical to #1596. If this is stacked on it, please say so in the description; otherwise we'll keep just one of the two.

[jayleaton]

Added a short description; this PR stacks on #1596 and adds only the pull request skill plus its two pointer lines, so merge #1596 first and review 8ae199036..542857cfe.

[ashhart]

Thanks @jayleaton. The pre-flight is careful: one behavior a PR, Refs against Fixes, every caller of the changed code, no tolerance moved after the results, and unrun not counted as passed. Those are points we make by hand in reviews. To match what we ask on kernel and speed PRs, could step 4 add that each new kernel has a planted control that fails as planted (with the SUMMARY showing it red), and that a change to output tokens on a default path needs proof and an opt-in, beside line 72's line about lossy conversions? And step 3's speed lines could add "no slower than main at one stream and the widest admitted; with --drafter, no slower than without it, greedy and sampled". One optional thought: where the skill restates CONTRIBUTING (lines 68-71 and 106-113), a link to the section would stop the two drifting apart. We'll land it right after #1596.

[jayleaton]

All three done on f15b510a, rebased onto #1596's new head 462f4a60 so the stack stays in order.

1. Step 4: each new kernel has a planted control that fails as planted, with the backend's SUMMARY line showing it red. Beside the line about lossy conversions, a change to output tokens on a default path now needs proof and an opt-in.
2. Step 3: decode and prompt processing are no slower than main at one stream and the widest admitted; with --drafter, no slower than without it, greedy and sampled.
3. Steps 4 and 6 now link CONTRIBUTING's exactness and precision, lean code and review, and public identity and privacy sections instead of repeating them; the rules CONTRIBUTING does not cover stayed.

Checked on this head: 88 relative links and anchors resolve (0 broken), git diff --check is clean, and the lean checker is 425 with 0 new. CI is rerunning on the rebased head.