loop

Estrelas0

Forks0

Atualizado15 de julho de 2026 às 07:34

Run the Vibin v2 build loop — pull a work item into the sprint, then retrieve → do → verify → capture → review → close, mostly inline. Use to build the backlog or resume work.

Instalação

Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.

Fonte

dxlbnl

dxlbnl/list

Abrir repositório GitHub Ver repositórios do creator

Download

Ocupações relacionadasSOC

Baseado na classificação ocupacional SOC

Especialistas em gestão de projetosNegócios e Operações Financeiras·SOC 13-1082

SKILL.md

readonly

name	loop
description	Run the Vibin v2 build loop — pull a work item into the sprint, then retrieve → do → verify → capture → review → close, mostly inline. Use to build the backlog or resume work.

The loop — Vibin v2 build cycle

You (the working session) run a work item end to end, mostly inline. There is no spec→test→implement relay: the card is the spec, tests are the spec, and you do the work yourself. Only review and research are separate (cold) agents.

Working state — the sprint

wiki/sprint.md is the one working-state file: the sprint's task list + a disposable run log.

Resuming: read wiki/sprint.md first — it says what's in flight. Tasks remain → continue them. No sprint file or all tasks done → run Sprint start (below) before touching any card.
Log as you go: the run log is append-only — add one-liners at meaningful steps; never overwrite or trim entries mid-sprint (it's the narrative the archive preserves). Put a short Next pointer just below the Tasks for resumability — not in the log. Knowledge does NOT live here.

Sprint start — plan, then get the go-ahead

The sprint boundary is the checkpoint; the sprint interior is autonomous. Don't plow from "what's this sprint about?" straight into execution — compose the plan and hard-pause for a nod first:

Goal — settle the one-line sprint goal with the user (a short conversation is fine).
Compose — pick the candidate cards from wiki/backlog/ and write the plan into wiki/sprint.md: the goal + the task list (slugs) + rough order. sprint.md is the plan.
Show + wait — present the composed sprint (goal, the cards = the workload, order) and wait for an explicit go-ahead before running any card. The user may swap / drop / reorder. This is the one blocking checkpoint per sprint — cheap to eyeball, and a mis-composed sprint is the costliest miss.
On go-ahead → run the items (Per item, below), interior autonomous until blocked.

Execute

Now follow the run skill (.claude/skills/run/SKILL.md) for every open task. It covers per-card execution (retrieve → do → verify → capture → commit → close) and sprint close (retro → archive → commit). The run skill is the canonical execution spec — don't re-derive it here.

One addition for human-driven loops: if a card has flags:[review] or a real fork appears mid-card, pause and discuss with the user before continuing. The run skill is otherwise autonomous.

Rules

Knowledge lives in wiki/knowledge/ atoms — never native memory, never the sprint log.
Links flow work → knowledge, never knowledge → work. One canonical home; replace (delete + repoint) when wrong.
New work discovered mid-task → /intake a card (slug-named); don't inline-patch.
Research items → spawn the researcher agent (writes atoms, no code).
Run until blocked within an agreed sprint: once the user has OK'd the plan (Sprint start), keep pulling items until the sprint is done, a review card, or a real fork. Composing a new sprint is itself a boundary — stop and get the go-ahead, don't auto-start the next one.

Mais deste repositório

mesmo repositório

bootstrap

dxlbnl/list

Primary entry point for a freshly cloned Vibin seed. Interviews the user, seeds the knowledge graph, scaffolds the stack, writes the permission profile, strips seed-meta, and hands off to the loop skill. Use when starting a new project, or when the user says "bootstrap" or "let's begin".

2026-07-150

intake

dxlbnl/list

File a new work item as a card in wiki/backlog/ (v2 — card-is-spec, slug filename, proportional detail). Use whenever the user reports a bug, asks for a feature, or proposes new work mid-task — instead of inline-patching.

2026-07-150

migrate-vibin

dxlbnl/list

Bring an existing Vibin project up to date with the latest seed by diffing the project's recorded seed commit against the latest Vibin on GitHub and applying the changes. Use when the user says "migrate", "migrate-vibin", "upgrade Vibin", "apply the latest seed changes", or after a new seed release. Reads the seed commit hash from .vibin-version, audits the project's actual state, applies the migrations added since, and reconciles seed-owned files while preserving local customizations.

2026-07-150

orchestrate

dxlbnl/list

Run multiple sprints in parallel — partition backlog cards, create git worktrees per sprint, spawn background subagents, merge when done. Use when you want to parallelise independent concerns across concurrent sprints.

2026-07-150

run

dxlbnl/list

Execute a pre-composed sprint — retrieve, do, verify, capture, commit, close. No planning, no user checkpoints. Used by the loop skill after planning, and by the sprint-runner agent for parallel sprints.

2026-07-150

status

dxlbnl/list

One-screen v2 status — current sprint tasks + run log tail, backlog cards, knowledge-graph size. Read-only.

2026-07-150

The loop — Vibin v2 build cycle

Working state — the sprint

wiki/sprint.md is the one working-state file: the sprint's task list + a disposable run log.

Resuming: read wiki/sprint.md first — it says what's in flight. Tasks remain → continue them. No sprint file or all tasks done → run Sprint start (below) before touching any card.

Log as you go: the run log is append-only — add one-liners at meaningful steps; never overwrite or trim entries mid-sprint (it's the narrative the archive preserves). Put a short Next pointer just below the Tasks for resumability — not in the log. Knowledge does NOT live here.

Sprint start — plan, then get the go-ahead

Goal — settle the one-line sprint goal with the user (a short conversation is fine).

Compose — pick the candidate cards from wiki/backlog/ and write the plan into wiki/sprint.md: the goal + the task list (slugs) + rough order. sprint.md is the plan.

Show + wait — present the composed sprint (goal, the cards = the workload, order) and wait for an explicit go-ahead before running any card. The user may swap / drop / reorder. This is the one blocking checkpoint per sprint — cheap to eyeball, and a mis-composed sprint is the costliest miss.

On go-ahead → run the items (Per item, below), interior autonomous until blocked.

Execute

One addition for human-driven loops: if a card has flags:[review] or a real fork appears mid-card, pause and discuss with the user before continuing. The run skill is otherwise autonomous.

Rules

Knowledge lives in wiki/knowledge/ atoms — never native memory, never the sprint log.

Links flow work → knowledge, never knowledge → work. One canonical home; replace (delete + repoint) when wrong.

New work discovered mid-task → /intake a card (slug-named); don't inline-patch.

Research items → spawn the researcher agent (writes atoms, no code).

Run until blocked within an agreed sprint: once the user has OK'd the plan (Sprint start), keep pulling items until the sprint is done, a review card, or a real fork. Composing a new sprint is itself a boundary — stop and get the go-ahead, don't auto-start the next one.