agent-skills · Addy Osmani · as of September 30, 2026
With agent-skills, you run the pipeline. That’s the point.
Many AI methods aim to take you out of the loop. agent-skills writes the opposite into its rules: you move the work from spec to launch one phase at a time, and at every phase the agent works the way a senior engineer would, right up to a rollback plan before anything ships.
This guide is written by EnsoDynamics, which makes the Enso Method. We compare our method with others as fairly as we can, and we recommend another method when it fits you better.
Two philosophies at a glance
agent-skills
Run the lifecycle yourself, phase by phase, while the agent works like a senior engineer at every step, through to a rollback plan before launch.
The Enso Method
Define the whole body of work first, keep it as the source of truth, and keep building on recorded assumptions while the people who decide confirm them.
Where they differ
- Stops and asks Decides, records it, keeps going agent-skills sits toward “Stops and asks”; the Enso Method sits at the “Decides, records it, keeps going” end.
- Specs serve one change Specs live as long as the system agent-skills sits in the middle; the Enso Method sits at the “Specs live as long as the system” end.
- Tests optional, or after the code Test-first, enforced agent-skills sits at the “Test-first, enforced” end; the Enso Method sits toward “Tests optional, or after the code”.
- You steer each step It runs on its own agent-skills sits toward “You steer each step”; the Enso Method sits at the “It runs on its own” end.
What they share
- Starts from a prompt Intent written down first agent-skills sits at the “Intent written down first” end; the Enso Method sits at the “Intent written down first” end.
- Light, few gates Heavy: roles and approvals agent-skills sits in the middle; the Enso Method sits in the middle.
Neither end of a line is the right one. Every comparison uses the same lines, so you can compare across methods.
Your judgment belongs between the steps
“The user (or a slash command) is the orchestrator.” That rule sits in agent-skills’ guide to orchestration, next to the thing it declines to build: an agent that runs every phase for you. Such an agent, the guide says, would skip the checkpoints that catch wrong-direction work early, and remove “user agency at exactly the points where judgment matters most.”
The reasoning starts from a line in the README: “AI coding agents default to the shortest path,” which “often means skipping specs, tests, security reviews.” So you hold the checkpoints, and the agent gets the discipline. Its skills are senior engineers’ workflows that end in evidence, and most carry a table of the excuses an agent uses to skip a step, each answered. Many draw on Google’s published engineering practices: Hyrum’s Law for APIs, the Beyonce Rule for tests, Chesterton’s Fence before deleting code.
From /spec to /ship
The lifecycle is a set of commands you run:
/specasks what you’re building and for whom, then writesSPEC.md: objective, commands, structure, code style, testing, and boundaries (always do, ask first, never do). You approve it./planwrites no code. It cuts the spec into vertical slices, each with acceptance criteria and a check, with a review point every two or three tasks./buildtakes the next task test-first, commits it, and stops./build autoruns the whole approved plan, pausing on failures, ambiguity, or anythinggit revertcan’t undo./reviewchecks correctness, readability, architecture, security and performance./shipruns code, security and test reviewers in parallel for a GO or NO-GO: “The rollback plan is mandatory before any GO decision.”
Along the way the agent shows its working: “ASSUMPTIONS I’M MAKING: … → Correct me now or I’ll proceed with these.”
Built for engineers who want to sign off every phase
agent-skills fits an engineer, or a team, who wants to approve each phase and wants security, performance and rollout handled as part of the work. Its own comparison guide says it is built for teams: named phases, review personas and shared checklists give several engineers and their agents one standard. Its examples lean towards web applications.
It asks for your attention at every gate, and its parallel reviews cost more tokens than a single pass; its advice is to load skills by phase, not all 25 at once. It suits you less if you’d rather hand off a large piece of work and walk away. Its own guide points that reader elsewhere.
Where the Enso Method made a different call
Both write intent down before code (“Code without a spec is guessing”), cut work into vertical slices, and read the code rather than ask about it. They part ways on three things.
When the spec runs out
agent-skills
“STOP. Do not proceed with a guess.” It names the gap, offers options and waits: requirements are “the human’s job.”
The Enso Method
It researches, picks the best-supported answer, records business questions as assumptions, and keeps building while their owners confirm or correct them.
Choose by situation: if you decide, agent-skills keeps every call yours. If the people who decide aren’t in the chat, the Enso Method keeps the build moving.
What the spec is for
agent-skills
A living SPEC.md for the work in flight. Keeping it after merge is optional.
The Enso Method
The requirements and design stay the system’s source of truth, amended before every change.
Choose by situation: agent-skills keeps documents light. The Enso Method costs upkeep, and pays off when someone asks, years later, what the system is supposed to do.
What proves it works
agent-skills
A failing test before any code, mostly fast unit tests, with outside services mocked.
The Enso Method
Tests come with the code. A later pass maps each acceptance criterion to evidence and swaps boundary mocks for real systems.
Choose by situation: agent-skills proves each test can fail. The Enso Method proves the integrations actually work.
Where agent-skills shines
- It doesn’t stop at the merge. Parallel reviewers, a GO/NO-GO, a mandatory rollback plan and numeric thresholds for a staged rollout. The Enso Method has no release or rollback step.
- It guards the bar against the agent.
CONSTRAINTS.mdputs your quality bar in numbers, and a diff check flags skipped tests, silenced checkers and lowered thresholds. The Enso Method forbids weakening a test but doesn’t check for it mechanically. - Its skills are tested. CI checks that each skill triggers when it should, and graded runs check that agents follow it. The Enso Method’s skills aren’t tested.
- Production breadth. Security, performance, observability and deprecation each get a full workflow, and accessibility a checklist.
How the Enso Method differs, and why
When the requirements run out, agent-skills stops and asks, so every judgment call stays with the engineer at the keyboard, phase by phase. The Enso Method answers open questions from research instead, records the business ones as assumptions for the people who own them to confirm, and keeps building, with the requirements kept as the system’s source of truth.
Get started
agent-skills is free and open source (MIT). It installs with one command (npx skills add addyosmani/agent-skills) or as a native plugin for Claude Code, Codex, Gemini CLI and other agents; the instructions are in the README at github.com/addyosmani/agent-skills.