Tasks Are Not Goals
A task tells an agent what to change. A goal should also carry intent, boundaries, stop conditions, and evidence.
ctxt.dev — Context Engineering (EN)
Blog about context engineering, structuring information for LLMs and AI development
Latest posts
Signum Can Now Be Installed in Codex App as a Plugin
What changes when a contract-first agent workflow becomes an installable Codex App plugin, not just a Claude Code command.
Stop Writing CLAUDE.md From Scratch
Every agentic codebase starts with an empty CLAUDE.md and fills it with hallucinations. Here's the six-file harness signum scaffolds for you.
AI Agents Need Permission Boundaries, Not Personalities
Most agent runtimes add more roles. punk starts from a harder premise: trust comes from boundaries, durable state, and proof.
My AI Agent Said 'Done.' It Skipped an Entire Acceptance Criterion.
The hardest bug was not in the code. It was in the trust model between the engineer agent and the orchestrator.
Your AI Agent Can't Tell Which Solution Is Current
AI agents rewrite code but leave the old version behind. jj's predecessor chains make ghost solutions detectable -- git can't.
Your AI Spec Is Already Stale
When agents read project.intent.md as ground truth, stale specs become execution bugs. Here's how I caught real drift in two projects.
Switching AI CLIs Without Losing 32 Skills: Why I Built nex
A Rust CLI that makes your AI agent skills portable across Claude Code, Codex, and Gemini. One command to install, one command to switch.
What a Formal Verification Agent Taught Me About Code Audit
Studying Mistral's Leanstral -- an agent for Lean 4 theorem proving -- led to concrete improvements in Signum, a multi-model code audit pipeline.
One Pass Isn't Enough: How Signum Learned to Fix Its Own Code
AI code verification as a loop, not a gate. Iterative audit, contract self-critique, and shared context across tasks in Signum v4.6.
Environment is context: security auditing for AI agent workstations
We carefully design prompts and tools but rarely audit the environment where the agent actually runs. Sentinel makes that measurable.
Skillpulse: Your AI Skills Are Flying Blind Without Telemetry
A PostToolUse hook that logs every skill activation to local JSONL. No existing tool tracks whether the model actually follows a skill's instructions.
Research Agents Lie. The Fix Is Adversarial Verification.
Most AI research tools optimize for coherent synthesis, not factual accuracy. Delve adds a claim-level adversarial verification stage that changes the trust model entirely.
From Plugin to Product: How Herald Became Sift and Why the Data Model Changed Everything
A local news plugin worked until it didn't. The fix was a different data model, language, and delivery surface.
Spec-Gated Delivery: Why PR Review Is the Wrong Trust Checkpoint for AI Code
AI made code cheap. It didn't make trust cheap. The fix isn't better reviewers - it's moving the gate from PR diff to approved intent.
AI Writes Code. Where Is the Proof?
Proofpack chains contract, implementation, and audit into a single verifiable record. Why proof artifacts are the missing primitive of AI code generation.
Herald v2: Local-First News Intelligence for AI Agents
How I built a 4-stage news pipeline that clusters articles into stories using title similarity, all in stdlib Python with SQLite.
The Contract Is the Context: How Signum Makes AI Code Verification Principled
Why running AI-generated code through more AI reviewers doesn't solve the reliability problem — and what a contract-first pipeline changes about it.
11 Plugins, One Marketplace: Building an AI Agent Toolkit from Scratch
How I built a plugin ecosystem for Claude Code — from scattered scripts to a full lifecycle with scaffolding, quality gates, multi-AI review, and one-command install.
First Agent Skills Benchmark: What Works, What Doesn't, and Why Context Matters
Analyzing SkillsBench — the first systematic benchmark for Agent Skills. 7,308 trajectories, critical review, and why skills are context engineering for agents.
Git is not for agents
Why git breaks AI agents and how jj solves every single one of these problems
Agent-friendly web: context engineering at internet scale
How content format determines whether an AI agent can see your site. Research data, real standards, and what to do right now.
Which model for which agent: metrics over intuition
Research on Claude model selection for multi-agent teams. Why Opus can be cheaper than Sonnet, and Haiku is dangerous for agentic tasks.
Gas Town: Multi-Agent Orchestrator Cheatsheet
Reference guide for Gas Town — a system for parallel management of 20-30 Claude Code agents. Commands, concepts, workflows.
Evoidea: A Memetic Algorithm for Ideas
Applying evolutionary algorithms to startup idea generation with AI agents
Loadout: Dependency Management for AI Skills
How to solve skill drift in AI agents. Manifest + lock + symlinks — a pattern from package managers applied to context management.
Claude Code Architecture: Why a Simple Loop Beat Complex Graphs
Breaking down Claude Code architecture based on PromptLayer founder's talk. Why while-loop, Bash, and context management matter more than complex workflows.
Semantic Catalog: How Enterprise Teams Engineer Context
Why Text-to-SQL and direct REST API mapping fail, and how a semantic graph of business entities solves the context delivery problem in enterprise.
Context Engineering: First Steps
Introduction to context engineering — an engineering approach to working with LLMs. Why prompts stop working and what to do about it.