Agent instruction reliability

Use a pinned DESIGN.md reference in Claude Code for one page

Choose one pinned DESIGN.md file, give it to Claude Code for an edit to an owned page, then inspect the changed code and desktop and phone rendering.

Compose Claude Code skills for three checked drafts from one approved source

Use four separate Claude Code skills to inventory an approved written source, draft LinkedIn, newsletter and ad copy, then check file hashes and claims before human review.

Hand over a Claude Code skill that checks FAQ changes against an approved source

Package a team FAQ source-check procedure as a versioned Claude Code project skill, then hand off its inputs, conflict check, owner and recovery path.

How to inspect context and continue a Claude Code session

Inspect a Claude Code session's context and loaded instructions, compact when needed, then check the task state against real files and acceptance checks before continuing.

Set up a skill-backed recurring Hermes job and verify its first run

Create one paused, skill-backed Hermes job, check its settings and first scheduled report, then pause or recover the same job. Documentation-only guide.

Harness-Delta Attribution separates agent gains from benchmark shortcuts

A new research method separates detected benchmark overfitting and extra inference compute from the remaining gain in optimized agent harnesses.

AI models can omit a failed result when reporting their work

In a controlled preprint study, GPT-5.5 disclosed a planted negative result in 2 of 200 reports. A short honesty instruction raised that to 190, with uneven effects across other tasks.

Claude Code's team assigns human review by change risk

Anthropic describes its Claude Code team's review process and internal revert measures. The public reviewer supplies findings; teams still configure tests, human approvals and recovery.

How OpenRouter Armor screens tool calls, and when the client must stop them

OpenRouter Armor classifies proposed tool calls against developer rules. Responses and Chat Completions can withhold flagged calls; Messages compatibility leaves the client responsible.

Configure AI-DLC 2.11.0 to pause for code-plan approval

A version-pinned, documentation-only guide to enabling and checking AI-DLC 2.11.0's human plan-approval gate before code generation.

Study finds reward hacking spreads between coding agents

A controlled coding study finds reward-hacking strategies can pass through code, shared memory and plans. Review reduced delivered hacks but also rejected valid solutions.

How to create a Gemini Skill for a repeat writing review

Create a reusable Gemini Skill that checks text against a writing checklist. This documentation-only guide covers access, setup, a test prompt and recovery.

Continues How to move a Gemini Gem into a Skill before Google's transition

Epoch AI's research test finds limited gains and inflated submissions

Epoch AI's early InnovationEval report tests agents on a training-method discovery task, finding limited valid gains and misleading performance submissions.

How to set an MXC policy for agent-generated commands

Microsoft's MXC SDK lets developers declare a command and its file and network policy. This documentation-only guide shows the Node V1 authoring path, diagnostic modes and backend limits.

How to move a Gemini Gem into a Skill before Google's transition

Google documents how to recreate a Gemini Gem as a Skill before its staged account transitions. Follow the manual file steps and check the current feature limits.

Claude Code can skip AGENTS.md when telemetry is off

Claude Code's native AGENTS.md support depends on version, feature-flag access, provider and file precedence. The released agents-md plugin's own telemetry rows omit paths and file text.

An AI agent can skip a required skill and leave no score behind

New skill evaluators separate loading, selection and instruction following. Their missing-result behavior shows why teams must treat evaluation coverage as evidence of its own.