writing

Imported from wustep/maths (.claude/skills/writing/SKILL.md). Install upstream with npx skills add wustep/maths --skill writing. Copyright stays with the author.

Writing

README, PROBLEM.md status, explainers, notes, and GitHub PR titles and bodies. ATTACK.md stays chronological. Walkthroughs stay refs/walkthrough-style.md.

Method from Terence Tao: On writing, Write professionally, Write in your own voice, Use good notation, Don't overoptimise, Organise the paper, Motivate the paper. Copy the method. Do not imitate the manner.

Result first

State the result accurately first. Use the introduction to sell the key points without overclaiming.

Punch line early. Technical details later. Factor into lemmas. A toy or special case before the general statement is fine.

Theorem, then what it generalizes, then the obstruction, then the argument. A motivating near-example before the proof. If the precise statements are slightly technical, say so and do not reproduce them.

Label heuristics (remark, footnote, "roughly speaking"). Keep the formal claim clean. Honest status: what is elementary, what needed a computer, what AI did.

English

Write English, not a string of symbols. Informal remarks are welcome if labeled as such.

Own voice. Paraphrase the literature; cite it. Do not parrot another author's sentences. Past self is a good audience.

Do not over-optimise or over-compress. Brevity that strips examples, remarks, and English makes the note harder, not simpler. Do not be cleverer than the result. Strike out the sentence you are most proud of if it is just style.

Notation

Define a symbol the first time it appears. Global notation early, local notation where used. Introduce a name only if it is used about three times or is central. Reinforce as "the vector space V", not just "V". Stay compatible with the paper you cite. No cute names.

Record

Credit, provenance, precedence. Compare with the published record. A new bound is a verified finite improvement of a published record. Write the inequality. An incomplete search (holes, SAT UNKNOWN, timeout) is not a lower bound. Write the leftover fact. Do not invent jargon. Use the language of the paper.

README and the explainers never say dent or residue; those are agent terms. ATTACK.md, WALKTHROUGH.md, PROBLEM.md status lines, and AGENTS.md may use them.

Before / after

README covering paragraph, same numbers.

Before: The November 2025 table (Davydov–Marcugini–Pambianco, arXiv:2511.02542, Table 5.1) had $\ell_2(10,2)\le 51$. This is a construction, an upper bound, not a conjecture. Sphere covering still only gives $\ge 45$, and an $n=49$ search left 7 holes, so 50 is not shown optimal.

After: The November 2025 table (Davydov–Marcugini–Pambianco, arXiv:2511.02542, Table 5.1) had $\ell_2(10,2)\le 51$. Status: construction; sphere covering $\ge 45$; $n=49$ still 7 holes.

Pull requests

The title is the change. The body says what changed and how to replay or check. No "this PR aims to." No chatbot close.