powerloop
Zbrooklyn/powerloop
A Claude Code skill that interviews you into a complete, self-stopping /loop before it runs.
Instalação
npx convez-ai Zbrooklyn/powerloop1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
powerloop
A Claude Code skill that interviews you into a complete, self-stopping loop before it runs — the way plan mode interviews you into a plan.
The problem
Claude Code's /loop will run a prompt on repeat. In practice most loops end up looking
like:
/loop 1m Continue Fixing
/loop 3m make them all 10 out of 10
/loop 1m Continue Designing
These have no done-condition, no verification contract, no scope, and no stop. The
loop runs until you notice and kill it — you're the off-switch. A fixed 1m/3m
interval is also usually the wrong tool: it's a clock, meant for polling external
state, but the actual work (design, browser, build) takes longer than the interval, so
the clock paces nothing.
An audit of 89 real
/loopinvocations over 11 days found ~95% were/loop 1m|3m <2-word vague continue>with no stop condition. That's what this fixes.
What /powerloop does
Type /powerloop and it:
- Reads the live context — the conversation, your
cwd,git status, the active project — and drafts a full loop spec, pre-filling everything it can infer. - Computes a risk tier —
max(write-impact, attendance); anything irreversible forces the strictest tier. - Interviews only the gaps — for each required slot it can't infer, it asks one plan-mode multiple-choice question with its best guess pre-loaded. An attended polish loop is usually 1–2 questions; everything else is inferred and shown for override.
- Confirms the assembled spec, then launches
/loopwith the done-condition embedded — so the loop stops itself.
The one inviolable rule: no loop launches without a done-condition.
The loop spec — 10 slots
Core (always required)
- Objective — the real outcome, never "Continue Fixing"
- Done-condition — the test that ends the loop
- Verification — proof produced each tick (screenshot / test / inspection)
Shape
- Per-iteration unit — what one tick does ("one fix", "one phase", "one item")
- Scope / boundary — what's in, what's off-limits
- Mode + cadence — dynamic self-pace (default) vs fixed-interval clock-poll
- Backlog source — inline prompt vs a project
loop.mdfile
Guardrails
- Commit / push policy
- Escalation trigger — what makes it stop and ask you
- Budget cap — max iterations / token ceiling / wall-clock
Adaptive tiers
| Tier | When | Required slots |
|---|---|---|
| Lean | attended + read-only / local | Core 3 |
| Standard | edits + commits, or stepping away | Core + unit, scope, mode |
| Full | push / deploy / prod, unattended, or irreversible | all 10 |
Install
powerloop is a single skill file. Install it as a user-global skill so it's available in every project:
macOS / Linux
git clone https://github.com/Zbrooklyn/powerloop.git
mkdir -p ~/.claude/skills/powerloop
cp powerloop/skills/powerloop/SKILL.md ~/.claude/skills/powerloop/SKILL.md
Windows (PowerShell)
git clone https://github.com/Zbrooklyn/powerloop.git
New-Item -ItemType Directory -Force "$HOME\.claude\skills\powerloop" | Out-Null
Copy-Item powerloop\skills\powerloop\SKILL.md "$HOME\.claude\skills\powerloop\SKILL.md"
Or drop it in a project: copy SKILL.md to .claude/skills/powerloop/SKILL.md in
the repo.
The skill is picked up on the next Claude Code session (often the current one).
Usage
/powerloop
…then answer the few questions it asks. That's it. It assembles the spec and launches the loop for you.
Requirements
- Claude Code with the built-in
/loopskill. - The
AskUserQuestiontool (used for the plan-mode-style interview).
How it works
The full skill logic lives in
skills/powerloop/SKILL.md. Design notes are in
docs/design.md.
License
MIT — see LICENSE.