Math Unicode

math-unicode

When emitting mathematical notation in a terminal coding agent (Claude Code, Codex CLI, or similar), always use Unicode glyphs inline — never wrap math in $…$, \(...\), or $$...$$. These terminals do not render LaTeX; raw delimiters appear as plain dollar signs and reduce readability.

When this skill applies

Two conditions, both required: the response carries mathematical notation, and the surface it lands on does not render LaTeX.

Surfaces that do not render LaTeX (apply the skill):

• Claude Code, Codex CLI, and other terminal or TUI coding agents
• Anything read through SSH, tmux, a pager, or a CI log

Surfaces that render math natively (do not apply the skill):

• ChatGPT and Codex on desktop and web, where math already renders
• Notebooks, and any target consumed by KaTeX or MathJax

No host exposes a per-surface predicate to a skill today, so this boundary is a judgement the model makes from its own context. If you cannot tell, and the session is a terminal coding agent, apply the skill.

Triggers (use Unicode math):

• Equations, formulas, derivations
• Filter conditions, set-builder notation
• Statistics: probabilities, expectations, distributions
• Calculus, linear algebra, logic
• Counts, ratios, fractions, drops where precision matters

Skip (do not transform):

• The user explicitly asks for LaTeX or a .tex file
• Math inside fenced code blocks (preserve source syntax)
• Strings being passed to a system that consumes LaTeX (KaTeX MCP, etc.)

Glyph cheatsheet

Greek

lowercase   α β γ δ ε ζ η θ ι κ λ μ ν ξ ο π ρ σ τ υ φ χ ψ ω
uppercase   Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
variants    ϵ ϑ ϕ ϖ ϱ ς

Operators

arithmetic     + − × ÷ ± ∓ · ∗ ⋅ ∘ ⊕ ⊖ ⊗ ⊘ ⊙
big            ∑ ∏ ∐ ∫ ∬ ∭
roots          √ ∛ ∜
calculus       ∂ ∇ Δ ∆
constants      ∞ ∅

Relations

equality       =  ≠  ≈  ≅  ≡  ≜  ≝  ≐  ∝  ∼  ≃  ≢
order          <  >  ≤  ≥  ⊴  ⊵
set            ∈ ∉ ∋ ∌  ⊂ ⊃ ⊆ ⊇ ⊊ ⊋  ⊏ ⊐ ⊑ ⊒
set ops        ∪ ∩ ⊎ ⊔ ⊓        (set difference: A \ B)

Logic & arrows

logic          ∧ ∨ ¬ ⊕   ⊢ ⊥ ⊤
quantifiers    ∀ ∃ ∄ ∴ ∵
arrows         → ← ↔ ⇒ ⇐ ⇔ ↦ ↪ ↩ ↑ ↓ ⇑ ⇓ ⟶ ⟵ ⟷ ⊸

Number sets & brackets

sets           ℕ ℤ ℚ ℝ ℂ ℙ ℍ
brackets       ⟨ ⟩  ⌈ ⌉  ⌊ ⌋  ‖ ‖

Sub/superscript glyph blocks

superscript    ⁰ ¹ ² ³ ⁴ ⁵ ⁶ ⁷ ⁸ ⁹  ⁺ ⁻ ⁼ ⁽ ⁾  ⁱ ⁿ ᵃ ᵇ ᶜ ᵈ ᵉ ᶠ ᵍ ʰ ʲ ᵏ ˡ ᵐ ᵒ ᵖ ʳ ˢ ᵗ ᵘ ᵛ ʷ ˣ ʸ ᶻ
sup (capital)  ᴬ ᴮ ᴰ ᴱ ᴳ ᴴ ᴵ ᴶ ᴷ ᴸ ᴹ ᴺ ᴼ ᴾ ᴿ ᵀ ᵁ ⱽ ᵂ
sup (Greek)    ᶿ                    (θ only; the rest are opt-in)
subscript      ₀ ₁ ₂ ₃ ₄ ₅ ₆ ₇ ₈ ₉  ₊ ₋ ₌ ₍ ₎  ₐ ₑ ₕ ᵢ ⱼ ₖ ₗ ₘ ₙ ₒ ₚ ᵣ ₛ ₜ ᵤ ᵥ ₓ

Do not invent or approximate a missing glyph. A whitelist on its own is not a membership test; these are the exact gaps that make the bare ^x / _x fallbacks in Rules 3, 4 and 5 fire. Two separate causes, same outcome.

No Unicode code point exists at all:

subscript letters      b c d f g q w y z    → x_b, x_c, x_d
subscript capitals     all 26               → A_B, M_N  (Unicode has no
                                              subscript capital, 0/26)
superscript capitals   X Y Z                → A^X, M^Y, A^Z
superscript ∞          none                 → never ^∞ as a glyph; use ∫[a..∞]
Greek sub/superscript  μ ν σ and most rest  → I_ν, ∂_μ, x^ν (Rules 3, 4, 8)

A code point exists, but no monospace font in the measured set ships it, so treat it as absent:

superscript S          U+A7F1, Unicode 17.0 → A^S
superscript C F Q      U+A7F2..U+A7F4       → A^C, A^F, A^Q
superscript q          U+107A5, outside BMP → x^q
Greek subscripts       ᵦ ᵧ ᵨ ᵩ ᵪ            → x_β, I_ρ, x_γ
Greek superscripts     ᵝ ᵞ ᵟ ᵠ ᵡ            → x^β, x^φ

Two Greek exceptions worth knowing, because the older wording here got them wrong. ᶿ (superscript θ, U+1DBF) renders as widely as the Latin blocks below and is a normal part of the portable set, so write xᶿ, not x^θ. And ρ does have a subscript, ᵨ U+1D68; it is missing from most terminal fonts rather than from Unicode, which is why I_ρ stays the recommendation.

Measured coverage of the portable set: 17/26 lowercase subscripts, 0/26 subscript capitals, and 19/26 superscript capitals with a code point that some monospace font ships. The ∞ gap is why big-operator bounds use a bracketed range rather than stacked scripts: ∫₀^∞ would render one bound as a glyph and the other as a caret in the same operator. When unsure, the bare ^ / _ form is always acceptable; a wrong or missing glyph is not.

If your terminal font does carry the Greek scripts or the rarer capitals, see references/extended-glyphs.md for the opt-in set.

Font coverage

Unicode says a code point exists. Whether a terminal draws it is a property of the font. These counts come from reading the cmap table of twelve monospace fonts: the six most widely installed programming fonts (JetBrains Mono, Fira Code, Cascadia Code, Hack, Source Code Pro, IBM Plex Mono) and six system fonts (DejaVu Sans Mono, Liberation Mono, Ubuntu Mono, Ubuntu Sans Mono, Noto Mono, Cousine). scripts/measure-font-coverage.mjs regenerates them and the test suite asserts every number below against the result.

Read that table as a ranking, not a verdict. A glyph the primary font lacks is usually supplied by fontconfig from another installed font, so it still draws, but at a different advance width, which is what breaks column alignment in aligned derivations and matrices. Only a glyph no installed font carries shows as tofu.

Two consequences worth acting on:

• The operators, Greek letters and digit scripts survive everywhere. Prefer them, and they carry most of the load in ordinary output.
• The letter sub/superscripts are DejaVu-class. JetBrains Mono, Fira Code, Hack and IBM Plex Mono ship none of them, so xᵢ and xᵀ arrive through fallback there. That is the reason Rules 3 and 4 keep x_i and x^T as first-class alternatives rather than last resorts.
• ⱼ and ⱽ are the two thinnest glyphs in the portable blocks, carried by one font each. They sit in the ranges above rather than in Glyphs to avoid because their fallback, x_j and A^V, is already what Rules 3 and 4 say to write when in doubt.

Common LaTeX → Unicode

Style rules

Rule 1 — Inline math: Unicode, no delimiters

Bad: The filter $f(T; m) = \{(s,r) : n_{s,r} \geq m\}$ produces the cohort. Good: The filter f(T; m) = { (s,r) : n_{s,r} ≥ m } produces the cohort.

Rule 2 — Block math: own line(s), still no delimiters

Bad:

$$|Q| / |T| = 5238 / 31075 \approx 16.9\%$$

Good:

|Q| / |T|  =  5 238 / 31 075  ≈  16.9 %

Rule 3 — Subscripts

• Single Unicode-renderable index: prefer the glyph (x₁, x₂, xᵢ, xⱼ, xₙ). ⱼ is the thinnest-supported glyph in this block, so x_j is equally fine.
• Single index with no subscript glyph: use a bare underscore — I_ν, ∂_μ, and x_c / x_b / x_d, which have no subscript form at all. Check the gap list under Sub/superscript glyph blocks before reaching for a glyph.
• Multi-character or grouped subscript: use _{...} syntax — the underscore reads unambiguously as a subscript and stays more legible than bracket-style indexing:
  • n_{s,r} ← (s,r) has no Unicode subscript form
  • x_max, σ_obs ← multi-letter
• Never mix: don't write x_₁ or x_{1} when x₁ works.

Rule 4 — Superscripts (powers)

• Simple / Unicode-mappable exponent: prefer the glyph — x², x³, xⁿ, eˣ, A⁻¹, and the transpose xᵀ / vᵀ.
• Single exponent with no glyph: use a bare caret — x^ν, (z/2)^a, and A^X / A^Y / A^Z, which have no capital superscript code point, plus A^S, which has one (U+A7F1, Unicode 17.0) that no terminal font ships. Aᵀ is fine, and so is xᶿ; see the gap list under Sub/superscript glyph blocks.
• Multi-character or expression exponent: use caret + parentheses, never ^{...} — x^(k+1), x^(i), e^(iπ). A bare x^{T} / x^{(i)} leaks LaTeX source; write xᵀ (single glyph) or x^(i) (parenthesized).