benchmark-task-authoring AGENTS.md

An instruction set for creating difficult Terminal-Bench 2 and Harbor benchmark tasks, which test what coding agents can do in a terminal.

In plain words
What is it for?
Use it to design, build, critique, or debug agent-benchmark tasks and keep related instruction files in sync.
Why use it?
It helps authors avoid tasks that look difficult but can be solved by simply replaying information supplied by the verifier.

Instructions file for CodexOpenCode

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 instructions/xclaw-bot/benchmark-task-authoring/agents-md
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Codex, OpenCode.

Per session 6,621 This file is loaded in full into every session.
When invoked 6,621 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 $0.06621 $0.06621
Opus 5 $0.03311 $0.03311
Sonnet 5 $0.01324 $0.01324
Haiku 4.5 $0.00662 $0.00662

Measured 2d ago against content hash d0eafb62816f, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

benchmark-task-authoring AGENTS.md 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.

AGENTS.md · 400 lines

How it starts

The opening of the file, as written. The whole thing — 400 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Authoring hard agent-benchmark tasks

Which file am I? The AGENTS.md entry point, read automatically by OpenAI Codex, Google Antigravity, and other AGENTS.md-aware agents. Cursor reads the same knowledge from .cursor/rules/; Claude Code from SKILL.md. All of them are generated from one source — edit SKILL.md or references/, then run python scripts/port.py.

Everything below applies whenever you are designing, building, critiquing or debugging a Terminal-Bench 2 / Harbor benchmark task.

This skill packages what was measured across 30+ Terminal-Bench 2 task slots — the ones that cleared a difficulty gate, and the many that died first. It exists because the intuition almost everyone brings to "write a hard task" is wrong in a specific, repeatable way, and the correction is cheap once you know it.

New here? Fifteen minutes, in this order

1 · Read "The one law" and the kill-list below (5 min). Do not skip to the reference files. If you internalise only one thing, make it the pre-build test: what must the agent build that the verifier could not hand it by simulating?

2 · Set up retrieval (1 min). The field manual is ~35k tokens and will not fit in one read call. Index it once and query it instead — this is the difference between a design question costing ~1k tokens and ~35k:

python scripts/dr.py index --no-embed
python scripts/dr.py ask "what trips ava_review verifier_coverage" --fast

3 · Put a board up (1 min). Before you write anything, see where your slots actually stand — and specifically which review stage each one is sitting at:

python scripts/taskdesk.py --org <your-task-org>

4 · Then, when you have a slot to work: references/ci-stages.md before your first push (it is the difference between one three-hour CI run and three of them), and the phase table below for everything else.

The two mistakes that cost the most, stated once so you can avoid both: designing a task where the agent checks a property instead of constructing an object, and treating CI as a test loop instead of a confirmation step.

Read the full file on GitHub · 400 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 · 400 lines · 6,621 tokens per session scan A d0eafb62816f

Subscribe to this mod's changes

benchmark-task-authoring AGENTS.md is an instructions file published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 18d ago), licensed MIT. It adds 6,621 tokens to every session, about $0.0331 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.