How I use AI

My Claude Code workflow.

The skills, agents and instructions I use as a professional Software Engineer. It's called Craft, and it's a plugin you can install.

View the Craft source on GitHub
Craft v4.10.0 6 Nov 2026 MIT
$claude plugin install craft --marketplace juanwmedia/craft
To keep it up to date automatically, turn on autoUpdate. Here's how
Three principles

What never changes.

  1. 01

    You decide what the agent cannot infer: the domain, the architecture, the structure of what gets built.

  2. 02

    Code is cheap once the uncertainty is gone. The decisions are written down and the agent builds on its own.

  3. 03

    Every step is reviewed in a fresh context: the documents against the code before you see them, the code against the documents once it is built.

The loop

Five steps. Two are human.

You decide what the agent cannot infer. After that, the loop runs on its own and every step is reviewed in a fresh context. Click the right or left side of the stage, or use the arrow keys.

An idea. You and Claude Code, nothing decided yet.

The loop starts with what the agent cannot infer: the domain, the architecture, the structure.

Shape: what we build and why.

It interviews you until nothing blocking is left open. Anything the code can answer, it looks up in the code.

A fresh agent audits the draft.

Evaluate, in a fresh context, checks every claim against the code. Evaluator-optimizer: fix it, audit it again.

Plan: slices that never share a file.

Each slice with its own check. This is where your time inside the loop ends.

Implement: one agent per slice, in parallel.

If a check fails, another fresh agent tries. Nobody asks you anything.

Review: another fresh context, against the documents.

Whatever is not confirmed gets fixed and reviewed again.

Try: live, with the evidence recorded.

What fails goes back to plan.

Close: the documents tell the truth again.

It reconciles what was built with the original proposal. One yes: commit, push and pull request.

You
shape
plan
implement
try
close
Human phases
← → or click
The skills

Nine skills (five are the main loop) and four agents.

The loop 5

/craft: shape

You: What we build and why. Agent: It interviews you until nothing blocking is left open; anything the code can answer, it looks up in the code.

/craft: plan

You: That the slices are the right ones. Agent: It splits the work into slices that never touch the same file, each with its own check.

/craft: implement

You: Nothing: your time in the loop ends here. Agent: It builds the slices in parallel, one agent per slice, passes every check and reviews the whole change.

/craft: try

You: Which proof counts. Agent: It tests the build live, end to end and in a browser when it applies, and records the evidence.

/craft: close

You: One yes. Agent: It reconciles the documentation with what was built, and then commits, pushes and opens a pull request.

When you are lost 1

/craft: wtf

You: Saying you're lost. Agent: It repeats the last answer until it lands, without moving the work forward.

Above the loop 3

/craft: detached

You: Whether it works without you. Agent: Every question a phase would ask you goes to act-as-me, and detached.md keeps them for the morning.

/craft: status

You: Which piece of work to pick up. Agent: It says where each open piece of work stands, where it came from and what the next step is.

/craft: bar

You: Showing it or hiding it. Agent: It shows the line above the prompt with the work, the phase, the slices and the agent running now.

The agents 4

build

You: Nothing. Implement launches it, one per slice. Agent: It builds its slice in a fresh context, changes only the files the slice lists, passes its check and reports.

evaluate

You: Nothing. Shape, plan and close launch it. Agent: It audits a document in a fresh context and returns every claim the code does not back, with file and line. The evaluator-optimizer pattern: the draft is fixed and audited again.

review

You: Nothing. Implement launches it. Agent: It tries to refute the change against the documents and the repo's conventions, and returns confirmed, refuted or uncertain, with evidence.

act-as-me

You: Whether you let it answer for you. Agent: In detached mode it answers what a phase would ask you, as a pragmatic senior would: it checks the code first, answers with the why, or says it cannot decide.

Changelog

The latest releases.

v4.10.0 6 Nov 2026

The craft bar leaves this plugin and becomes craft-bar, installed from the same marketplace with /plugin install craft-bar@craft. The Claude plugin directory accepts mods only from Verified organizations, so craft ships no mod and the directory can list it. Updating craft alone removes the bar until craft-bar is installed; /craft:bar is now /craft-bar:bar.

v4.9.1 23 Oct 2026
  • The craft bar shows the work, the phase and detached for a phase the model starts with the Skill tool, the same as for one you type. Before, a phase started that way (a routine asking in words, a session moving on to the next phase) showed only craft.
  • /craft:detached answers one line, whether the mode is on or off in this repository, without the line about routines.
  • The bar draws above the prompt in the terminal and in the desktop app. The web (claude.ai/code) does not draw it.
v4.9.0 23 Oct 2026

Detached mode: craft carries a piece of work with nobody at the keyboard.

  • /craft:detached on turns it on and /craft:detached off ends it. A routine starts with /craft:detached on as its first message, or asks for the mode in words, since both /craft:detached and /craft:bar can now be called by the model.
  • While it is on, every question a phase would put to you (shape, plan, implement, try and close, approvals and confirmations included) goes to the new agent craft:act-as-me, which runs on Fable 5.1 in a fresh context and answers as an extremely pragmatic senior software engineer who takes nothing for granted. It checks the code first, answers with one option, given or its own, and why, and says Cannot decide with what is missing instead of inventing. When it cannot decide, or does not start, the phase stops and names the question.
  • The mode is the flag craft.detached in the repository's local git config: per repository, outside the tree, never committed.
  • Each answer is appended to detached.md in the work's folder with its phase, the question, the answer and why, in order. It is the file to read in the morning, next to the pull request.
  • Close commits, pushes and opens the pull request when craft:act-as-me says yes. The yes is logged in detached.md before the commit, so the commit carries it, and close says who gave it. You review the pull request.
  • The craft bar shows detached after the phase while the mode is on, and ends with an x that hides it, the same as /craft:bar off.
  • Four evals grade the detached path of shape, plan, try and close on the trace and the files: no AskUserQuestion call, craft:act-as-me started for the questions, the entry in detached.md.
  • Without the flag the human answers every question, as before. Every rule now asks as references/ask.md says, instead of naming the human or AskUserQuestion itself.
Full history on GitHub ↗
Free guide

The 51 essentials, as a guide.

One page per tip. Five chapters. What I actually use daily in production. No theory, no fluff.

  • I. Getting started 10 tips
  • II. Awareness 3 tips
  • III. Mastery 22 tips
  • IV. Autonomy 10 tips
  • V. Comparison 6 tips
Are you a professional Web developer?

You'll receive the guide by email · You join the Gravitas newsletter · Unsubscribe anytime

of 51
#

Wmedia · 51 Tips
Free guide · 51 tips · 5 chapters

The 51 essentials, as a guide.

Are you a professional Web developer? · Unsubscribe anytime