A small development process with roles, for AI coding agents. You brainstorm the idea, write a plan, review it with fresh agents, implement it, and review the code the same way. A "step back" agent asks if the work is worth it. You see only the judgment calls.
flowchart LR
brainstorm["💬 brainstorm"] --> plan["📝 plan"] --> r1("🔍 review") --> implement["🔨 implement"] --> r2("🔍 review") --> ship(["🚀 ship"])
classDef think fill:#fef3c7,stroke:#d97706,color:#1f2937
classDef build fill:#d1fae5,stroke:#059669,color:#1f2937
classDef review fill:#e0e7ff,stroke:#4f46e5,color:#1f2937
class brainstorm,plan think
class implement build
class r1,r2 review
Run /devs:start with an issue, a file, or a sentence, or with nothing and
it asks what to work on. Each stage then loads the next one by itself. It
stops only for your answers, a CHANGE COURSE or STOP verdict, or a failed
build. It never merges or pushes.
For a small change (a list of nits, a bug with a known cause), start
offers the small path: plan, implement, and one neutral reviewer on the
code, with no plan review and no step back. It takes that path only when
you say yes.
You can also start any stage alone with its command (/devs:brainstorm,
/devs:plan, /devs:review, /devs:implement) or in plain words, for
example "review this with security and minimalist". Say "only the review" to
stop after that stage.
| Skill | What it does |
|---|---|
start |
Takes the input (issue, file, text) or asks for it, picks the first stage, and runs the flow to the end. Continues an existing .work/<feature>/ from its next stage. |
brainstorm |
A two-way talk about the idea: the model brings options, opinions, and prior art, and grills you one question at a time until every decision is settled. Uses Matt Pocock's grilling method (skills/brainstorm/grilling.md). |
plan |
Takes any input, grills only the open decisions, writes plan.md. |
review |
Fresh reviewers → sort findings → apply clear fixes → step back verdict → your decisions. Works on the plan and on the code. |
implement |
Builds from plan.md, runs the project checks until they pass, commits. |
flowchart LR
work["📄 your plan<br/>or code"] --> neutral["🧐 neutral"] & minimalist["✂️ minimalist"] & skeptic["👎 skeptic:<br/>the cons"] & lenses["🛡️ security · ⚡ perf · 🎨 ux"]
neutral & minimalist & skeptic & lenses --> stepback["🔭 step back:<br/>is it worth it?"] --> you(["🙋 you decide"])
classDef review fill:#e0e7ff,stroke:#4f46e5,color:#1f2937
classDef think fill:#fef3c7,stroke:#d97706,color:#1f2937
class neutral,minimalist,skeptic,lenses review
class stepback think
| Persona | What it does |
|---|---|
🧐 neutral |
A careful senior developer who sees the work for the first time. Checks that it is clear, complete, and does what it says. |
✂️ minimalist |
The laziest senior developer in the room. "The best code is the code never written." Looks for what should not exist and what can be smaller. |
👎 skeptic |
Reads only the plan and gives the cons. It has no checklist. |
🛡️ security |
Reports only problems with a real path: who the attacker is, what they control, and what they get. |
⚡ performance |
Finds the real load of the project, today and after months of use. Judges only against that. |
🎨 ux |
Speaks for the person who uses the result. Taste is not a finding. |
🔭 step back |
Was not in the room. Asks if this is the right thing, if it is worth it, and if a cheaper way exists. |
All except the skeptic can report that they found nothing. They never invent a finding to have something to report.
- Reviewers run as fresh subagents with no chat history, on the same
model as your session unless you name another.
neutralandminimalistalways run.security,performance, anduxrun when the change touches that area or when you ask. - Skeptic. One more fresh agent gets a single job: read the plan and give the cons, the reasons why this is bad. It has no checklist and no protected decisions. It runs on every plan, and on code when you ask. The main session does not filter its list: you see every con with a one-line answer, and you pick what changes.
- Nobody is protected.
plan.mdtags each decision(user)or(agent). Before the review,planlists the(agent)decisions for you in plain words. Reviewers question both kinds. A finding against one of your own decisions always comes back to you as a question. - Persona files are in
skills/review/personas/. Each has an "On a plan" and an "On code" section, plus "Tuning notes" for your own rules. The lists in a file are examples. A reviewer reports anything that fits its purpose. - Sorting. The main session opens the cited code and labels each finding:
- FIX: right and clear, so it is applied. A FIX only corrects or removes. A fix that adds a mechanism, an option, or a dependency comes to you as a question.
- CONTEST: wrong, with evidence you can check.
- ESCALATE: needs your judgment, so you get a question.
- DECLINE: right but too small to matter.
- Step back. One more fresh agent reads the plan and the findings and says GO, CHANGE COURSE, or STOP. It can add remarks: parts it would drop, change, or check first, also on a GO. You see each remark as written and you answer it.
- You see the verdict, one count line, the skeptic's cons and the step
back remarks, and the escalated questions one at a time. The details are
in
.work/<feature>/plan-review.mdorcode-review.md. - Status. A review round ends with a
Status: doneline after your last answer. A round without it is not finished, andstartreturns to it.
skills/principles.md holds your coding principles: simplicity,
maintainability, understandability, elegance, and comments that explain why.
plan and implement carry a copy inline, and review pastes it into the
neutral and minimalist reviewers' briefs. Security,
performance, ux, skeptic, and step back do not, so they stay in their own lanes. A
project's own AGENTS.md or CLAUDE.md wins where it differs.
.work/<feature>/
plan.md # the plan, the source of truth
plan-review.md # plan review rounds
base-commit # commit before implement; the code review diffs from here
code-review.md # code review rounds
Commit .work/ or add it to .gitignore. Each project decides.
/plugin marketplace add urza/devs
/plugin install devs@urza
Other harnesses: the skills are plain SKILL.md files with no Claude Code tool
names. Copy skills/ into the harness's skills folder. The review skill
needs a harness that can start fresh subagents.
urza/devs is the stable version that you install. Work happens in the fork
d3mnjx/devs and reaches urza/devs through pull requests.
skills/brainstorm/grilling.mdis copied unchanged from mattpocock/skills at4ae2317, MIT license (seeskills/brainstorm/grilling-LICENSE).skills/review/personas/minimalist.mdadapts the review rules of Ponytail by Dietrich Gebert, MIT license (seeskills/review/personas/minimalist-LICENSE).- The fresh reviewers and the finding labels come from Google's Elephant-Goldfish Model.