40-authoring

A writing and review guide for benchmark instructions and proposals, including length, file naming, required outputs, and alignment with grading rules.

In plain words
What is it for?
Use it to outline, critique, and red-team instruction.md files and proposal documents before they are written or submitted.
Why use it?
It helps authors avoid vague, overly prescriptive, or incomplete task instructions that make benchmark results less useful.

Cursor rule for Cursor

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add rules/xclaw-bot/benchmark-task-authoring/40-authoring
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Cursor.

Per session 1,028 This file is loaded in full into every session.
When invoked 1,028 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.01028 $0.01028
Opus 5 $0.00514 $0.00514
Sonnet 5 $0.00206 $0.00206
Haiku 4.5 $0.00103 $0.00103

Measured 2d ago against content hash 5921c7e491c4, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

40-authoring scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.cursor/rules/40-authoring.mdc · 89 lines

How it starts

The opening of the file, as written. The whole thing — 89 lines — stays where its author put it; the contents beside it link to each section on GitHub.


description: instruction.md style rules, the 4-section proposal format, and rubric alignment. Auto-attaches to instruction and proposal files. globs: /instruction.md,/proposal-*.md alwaysApply: false

Authoring the prose artifacts

These are human-written (see 00-mission.md). Your role is outline, critique and red-team — never composition. What follows is the standard to critique against.

instruction.md

It is a prompt, not a document. Written as a domain expert briefing a skilled colleague.

  • ≤1500 tokens. No title, no section headers, no preamble, no roleplay.
  • Absolute paths in backticks`/app/data.csv`, never data.csv.
  • Name every expected output file and its exact format.
  • Describe the what, never the how. No step-by-step procedure, no mandated tools or libraries ("use scipy.optimize" is a defect). Overspecification is the single most common way a task stops being hard.
  • Do not list available tools, libraries or environment details that are not direct inputs — the agent should discover them.
  • Do not spell out prerequisite knowledge that experts agree on unanimously. Field-standard definitions and formulae may be assumed.
  • Structured output ⇒ the exact schema is normative and stated. Examples alone are not sufficient.
  • Disclose only what is genuinely non-standard, dataset-specific or arbitrary and which the agent must match: an output schema, a tolerance, a rounding rule, a pick among equally standard conventions.
  • LLM-flavoured prose is a rubric red flag in itself.

The 4-section proposal

House format — match an existing proposal-*.md. Header line Category: X Sub-Category: Y, then:

  1. Why this task is genuinely difficult. Lead paragraphs describing the problem, then italic-led paragraphs: The professional and why it is valuable. (name who is actually paid to do this), The data. (synthetic or real, provenance, whether realistically challenging), The pitfalls. (the concrete traps, and why difficulty is reasoning rather than tedium).
  2. Intended solution approach. The key insight first, then numbered steps from priors to answer, then expert effort in hours. Another domain expert should be able to reimplement from this alone.
  3. How the solution will be verified. What is checked and how; why exactness or each tolerance is calibrated as it is; anti-cheat; what discriminates a correct solver from the plausible wrong one; cross-validation against an independent implementation.
  4. Category and sub-category. Justify the labels against what the task actually asks the agent to do and produce.

Read the full file on GitHub · 89 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 89 lines · 1,028 tokens per session scan A 5921c7e491c4

Subscribe to this mod's changes

40-authoring is a cursor rule published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 18d ago), licensed MIT. It adds 1,028 tokens to every session, about $0.0051 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.