writing-style

name	writing-style
description	Writing standards for code comments, documentation, specs, PRDs, and team communication. Applies RFC 2119 requirement language and token-efficient writing.

Writing Style

Apply these standards to all written output: code comments, docs, specs, PRDs, wiki pages, commit messages, and team communication.

RFC 2119 Requirement Language

Spec: https://www.rfc-editor.org/rfc/rfc2119

Every directive MUST use an RFC 2119 keyword in ALL CAPS. This applies to wikis, skills, rules, specs, PRDs, issue acceptance criteria, and any content that tells a reader or agent what to do. Agents weight capitalized keywords more reliably than prose; soft language is silently downgraded or ignored.

• MUST / REQUIRED / SHALL: absolute requirement, no exceptions
• MUST NOT / SHALL NOT: absolute prohibition
• SHOULD / RECOMMENDED: strong recommendation, exceptions need justification
• SHOULD NOT / NOT RECOMMENDED: discouraged, acceptable with careful reasoning
• MAY / OPTIONAL: truly optional, implementer's choice

Decision framework: Does the system break without it? → MUST. Degrades? → SHOULD. No impact? → MAY.

Conversion table: soft language MUST be rewritten

Soft phrase	Rewrite as
always / never	MUST / MUST NOT
be sure to / make sure	MUST
don't / do not	MUST NOT
don't forget to	MUST
needs to / has to	MUST
try to / aim to	SHOULD
should probably / ideally	SHOULD
can / feel free to	MAY
avoid	SHOULD NOT
remember to	MUST

Pre-save checklist

Before saving any directive sentence, confirm it contains one of: MUST, MUST NOT, SHOULD, SHOULD NOT, MAY. If it doesn't, rewrite it or demote it to descriptive prose (no directive verb).

When lowercase is correct

Use lowercase "should/must/may" only for descriptive prose, not directives: "the build should finish in under a minute" (observation) vs "the build SHOULD finish in under a minute" (requirement). If the sentence tells someone what to do, it's a directive. Capitalize.

In CLAUDE.md and skill instructions, use "YOU MUST" or "IMPORTANT" as emphasis markers to improve adherence to critical rules.

Reading and complying with RFC 2119

Keywords are binding on the reader, not suggestions to weigh against convenience. When you encounter a keyword while executing work:

• MUST / MUST NOT / SHALL / SHALL NOT: absolute. You MUST NOT downgrade a MUST to a SHOULD because it would take longer, because an example in the session shows otherwise, because the codebase seems to do it differently, or because you judge it unnecessary. The writer chose MUST deliberately. If the requirement seems wrong or impossible, STOP and ask the user. Do not silently relax it.
• SHOULD / SHOULD NOT: strong default. Deviation requires a stated justification based on facts of the current situation, not convenience or time pressure. "It would take too long" is never a valid justification.
• MAY: truly optional. Your choice.

The failure mode this prevents: an agent reads "tests MUST cover the edge case", finds writing the test tedious, and silently demotes it to "tests should cover the edge case where practical". That is the agent overriding the writer. MUST means MUST. If you can't comply, stop and report. Do not proceed with a weakened version.

Prior examples in the current session do NOT override a MUST. If session context shows code that violates a MUST from the wiki/skill/spec, the session code is wrong. Flag it, do not use it as permission to violate the MUST.

Prohibitions go before the procedure

A MUST NOT placed after the steps it governs is read too late to prevent anything. Aerospace documentation standards put warnings and cautions before the step for exactly this reason.

• MUST place prohibitions, approval gates, and safety conditions above the procedure they constrain.
• MUST NOT bury an approval gate in a closing paragraph. The versioning rule is the model: "MUST NOT deploy without explicit user approval" precedes the deploy steps.

Controlled Vocabulary

RFC 2119 is controlled language applied to one word class. It fixes the sense of modal verbs so they cannot be read two ways. ASD-STE100 (Simplified Technical English, the aerospace maintenance standard from ASD's STE Maintenance Group) applies the same move to the whole document under one constraint: one word, one meaning, one part of speech. Its approved dictionary is aerospace vocabulary and MUST NOT be adopted here. Two of its rules earn their keep in agent-facing writing.

One term per concept

Pick one word per concept and use only that word. A synonym reads as a new concept to an agent, which then hunts for a distinction that does not exist.

• MUST use the same term for the same thing across a document set. In this repo a rule is always-loaded markdown in rules/, a skill is on-demand markdown in skills/, and a command is a skill invoked by slash. MUST NOT substitute "guardrail", "convention", or "standard" for any of them.
• MUST NOT vary wording for style. Repetition is correct here. Elegant variation is a defect.
• SHOULD define a term once, where the reader first meets it, then reuse it verbatim.

Name the noun, do not point at it

"This", "it", and "that" pointing back at a whole clause is the most common ambiguity in technical prose, and the one agents most often resolve wrongly.

• MUST NOT open a sentence with "This" or "It" referring to the preceding sentence as a whole. Repeat the noun.
• ❌ "The validator compares same-named files only. This means repo-only rules are safe."
• ✅ "The validator compares same-named files only. The same-name comparison means repo-only rules are safe."

Repeating a noun costs a few tokens. Resolving the wrong referent costs a wrong edit.

Token Efficiency

Every token costs time, money, and cognitive load. Be concise without losing clarity.

Why this matters: brevity is front-loaded effort. A short text costs the author more and every reader less. The work doesn't vanish, it moves from the author once to every reader never (the same ledger as "code is a liability": ongoing cost, one-time value). As Pascal wrote in 1657, "I have made this longer than usual because I have not had time to make it shorter." The line is commonly misattributed to Mark Twain.

Rules:

1. Start with action verbs: fix, add, update, remove
2. Use file:line references: auth.ts:45 not "authentication file line 45"
3. Eliminate filler: actually, basically, in order to, it's worth noting
4. Choose specific over generic: "Redis cache" not "caching solution"
5. Bottom Line Up Front: lead with the answer, then context

By audience:

• Client-facing (Basecamp): Plain language, no jargon, explain impact not implementation. The reader is non-technical, so write for clarity over brevity.
• Developer-facing (GitHub issues, PRs, wiki): Technical and concise. Convey context for developers and AI agents. Include file references, reproduction steps, and acceptance criteria.
• Internal (code comments, commits): Minimal. Reader has full codebase context.

By format:

• PRDs: Bullet points, not paragraphs
• Specs: Lead with requirements (MUST/SHOULD/MAY)
• Tasks: Action verb + target
• Code comments: Only where logic isn't self-evident
• Reviews: What changed, not why

When to expand:

• User stories, onboarding docs, error messages: reader has no prior context
• Architecture decisions: future developers need the "why"
• External-facing docs: audience can't ask follow-up questions

When to cut:

• Internal specs, tasks, code comments: reader has shared context
• Commit messages, PR titles: scanning speed matters
• Anything repeating what the code already says

Default to brief. Expand only when the reader cannot infer meaning from context. Two sentences that answer the question beat two pages that fill the context window.

Name Sources to Activate Knowledge

A canonical phrase paired with the author and work names a whole discipline's concept cluster in a handful of tokens. For an agent, this shifts the reasoning pool: "improving daily work" alone pulls from a generic improvement cluster; "Gene Kim's The Phoenix Project, the DevOps Second Way" activates feedback loops, toil reduction, blameless post-mortems, value streams, and the rest of the adjacent vocabulary. Cheap tokens, richer reasoning.

When it fits:

• The named discipline genuinely shapes the work. DevOps in work-habits. Goldratt's commission/omission errors in /sanity-check. RFC 2119 in writing-style. Each one names a body of thought the agent is expected to apply.
• The attribution is compact: canonical phrase, author, source, and any shared vocabulary needed to anchor the cluster.

Anti-patterns:

• Name-dropping without vocabulary. "Gene Kim said improvement matters" activates almost nothing; you need the phrase, the work, and the domain terms together to hit the cluster.
• Citing disciplines that do not shape the work. Invoking Agile Manifesto in a Bash-simplicity skill pulls in adjacent but misaligned concepts, and the cost is confused reasoning.
• Citing one source to decorate every skill. If every skill attributes someone, attribution becomes ornamental and the activation effect dilutes. Keep it rare and deliberate.

This is a technique, not a rule. Most skills should not invoke an external discipline at all. Use it when a skill's guidance genuinely derives from a named body of thought, and the agent will reason better for being told which one.

Punctuation: Avoid the Em-Dash

The em-dash (—) is one of the strongest tells of AI-generated text. LLMs over-produce it dramatically, far beyond how humans punctuate. Readers (especially developers and clients reading client-facing posts) increasingly read em-dashes as "this was written by AI."

SHOULD NOT use em-dashes. Almost every em-dash can be replaced with simpler, more natural punctuation:

Em-dash usage	Replace with
Parenthetical aside	Comma, or parentheses if truly aside
Sentence break for emphasis	Period (start a new sentence)
Introducing a list or example	Colon
Connecting two related clauses	Period or semicolon
Range (5—10)	En-dash (–) or "to"

Examples:

❌ The wiki is the spec — the repo is the implementation.
✅ The wiki is the spec. The repo is the implementation.

❌ Cross-linking turns files into a knowledge graph — each link is an attention signal.
✅ Cross-linking turns files into a knowledge graph. Each link is an attention signal.

❌ Three audiences read the wiki — junior devs, senior devs, and AI agents.
✅ Three audiences read the wiki: junior devs, senior devs, and AI agents.

❌ Memory is unauditable — the user cannot easily review what's saved.
✅ Memory is unauditable. The user cannot easily review what's saved.

MAY use em-dashes only when:

• Quoting source material that contains them (preserve the original)
• A genuine parenthetical aside where commas would create ambiguity AND parentheses would be too formal. Even then, prefer restructuring.

When editing existing content, replace em-dashes opportunistically. When writing new content, default to periods, commas, and colons. If you find yourself reaching for —, that is a signal to break the sentence into two.

List Ordering

Lists with no inherent order (deny rules, config arrays, feature lists, bullet-point explanations) MUST be sorted alphabetically. Alphabetical order makes items easier to find and minimizes git diffs. New entries slot into a predictable position instead of appending at the end.

Applies to JSON arrays, markdown bullet lists, table rows, and any unordered collection.

Code Style

Use descriptive variable and function names. Abbreviations save keystrokes but cost readability. The human reviewing your output MUST be able to understand the code without deciphering names. Short names also collide with terminology in the surrounding conversation and confuse the reader (e.g. using p as a path variable while the discussion is about a payment path; the reader will read p as "payment", not "path").

• MUST use full words: $groupQuantity not $gq, cartItem not ci, filePath not p, response not r
• MUST NOT use single-letter variable names except loop counters (i, j, k)
• Well-known abbreviations (id, url, db) are fine
• Function names SHOULD describe what they do: calculateShippingCost not calcShip

Claude Code Keywords

Reasoning depth is set by effort level, not by prose. /effort accepts low, medium, high, xhigh, and max (available levels depend on the model), plus ultracode, which is a Claude Code setting rather than a model level. A skill MAY pin its own level with the effort frontmatter field.

ultrathink is the one recognised keyword: include it anywhere in a prompt or in skill content to request deeper reasoning for that turn without changing the session effort. think, think hard, and think more are NOT keywords. They are passed through as ordinary prompt text, so MUST NOT be written as if they control anything.