[jayleaton]

What this changes

Before: contributor agents had no procedure for preparing a pull request, and maintainers kept asking for the same missing pieces. After: docs/pr/SKILL.md (tensorfold-pull-request) is a pre-flight an agent runs before gh pr create. Its §3 and §7 map each step to a line of the revised template. AGENTS.md and README#contributing-with-agents point to it.

docs/pr/SKILL.md   new, 125 lines
  1 Scope, issue and coordination   -> Issue, Stacked on, atc overlap
  2 Base and head                   -> Head commit, Stacked on, checklist line 1
  3 Fill the template               template section -> skill step; delete only what the template says to delete
  4 Kernels, arithmetic, platforms  -> Kernel contract section, Paths and platforms
  5 Tests and checks                -> Lean, Tests run, CI, Not run
  6 Hygiene                         -> checklist hygiene lines
  7 Final self-check                only what the template checklist does not already ask
AGENTS.md, README.md   +1 line each

Each rule comes from CONTRIBUTING.md, the kernel contract or the ATC skill, plus a recent survey of 100 PRs. This change doesn't touch .github/.

• Issue: n/a. This is maintainer-requested repository process work with no issue.
• Stacked on: Issue forms and a sharper pull request template #1596 (the issue forms and revised template, 8ae199036). This PR adds one commit, 542857cfe. The base stays main because Issue forms and a sharper pull request template #1596's branch is on a fork, and GitHub only accepts a base branch in this repository. The diff therefore also shows Issue forms and a sharper pull request template #1596's .github/ files until Issue forms and a sharper pull request template #1596 lands. Review 8ae199036..542857cfe for this change alone.
• atc overlap --json (ATC 0.7.0, 2026-10-29, against main): complete: true.
  • Seven open PRs that share README.md merge cleanly: lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508, lfm2: a cheaper DSpark hold (one concurrent encoder) #1394, minimax_m3: a CUDA backend for MiniMax-M3.1-Flash at TP over GB10s, from the release's own formats #1388, minimax_m3: serve MiniMax-M3.1-Flash (428B-A23B MoE, images) natively on Metal #1235, lfm2_moe CUDA: attention pages from the shared KV pool, and the prompt cache (stacked on #1492) #1503, lfm2_moe: serve LFM2.5-24B-A2B on CUDA (plain serve, shared kernels) #1492, minimax_m3 (CUDA): draft several streams' DSpark blocks in one pass #1426. lfm2: keep a drafter's tapped rows on every forward #1390 is unknown.
  • lfm2: draft with LiquidAI's DSpark drafter (--drafter) #1391, lfm2: the DSpark confidence head schedules the verify #1392 and lfm2: the DSpark drafter loads one matrix at a time and keeps a lane in one buffer #1393 show conflicts on README.md. Each has the same single conflict hunk in the model table against main without this change, so the conflict is inherited and this line adds no new hunk.
  • The only local worktree overlap is Issue forms and a sharper pull request template #1596's branch, and it is clean.

Checks on this head

• Head commit: 542857cfe226228b9768c2e2c8a37bdb32224661
• Lean checker, run on the merge base and on this head with the same checker as CI: lean: 425 problems, 425 on the base, 0 new. The 425 come from the base tree.
• Tests run:
  • Every relative link and anchor in docs/pr/SKILL.md, AGENTS.md and README.md resolves (script check).
  • git diff --check is clean.
  • The skill's kernel commands match docs/kernel-contract.md and the code. CUDA limits come from tf-cuda-test kernel-contract-limits DIR (zig/tests/cuda/main.zig), and hip-kernel-contract-limits is a zig build step (zig/build/hip_contract.zig).
• CI: fmt passes; the other workflows have not started (fork workflows wait for approval).
• Not run: the Zig host suite and zig fmt, because no Zig source changed. No workflow runs a markdown check on docs.

Receipt

n/a: this is documentation and 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)
• 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.