using-vibekit

name	using-vibekit
description	Use when starting any conversation — establishes vibekit's auto-trigger discipline so pipeline skills (brainstorm, brief, plan, exec, verify, review, finish, memory) actually fire at their trigger points instead of being silently skipped.

If you were dispatched as a subagent to execute a specific task, skip this skill. Subagents follow their RTCO brief, not the orchestration discipline. If you think there is even a 1% chance a vibekit skill applies to what you are about to do, you MUST invoke it via the Skill tool.

This is not negotiable. You cannot rationalize "the task is too small," "I already know the answer," or "it would be faster to just do it." Skipping a guardrail skill is the failure mode this plugin exists to prevent.

Instruction priority

1. User's explicit instructions (CLAUDE.md, AGENTS.md, GEMINI.md, direct messages) — highest.
2. Vibekit skills — override default behavior where they conflict.
3. Default system prompt — lowest.

If the user says "skip the brainstorm" or "just write the code," follow the user. Otherwise, follow the trigger map below.

Auto-trigger map

Trigger condition	Skill that MUST fire
User says "vibe" or gives a short intent ("add X", "build Y", "fix Z")	vibe
About to start any creative or implementation work, before code is written	brainstorm-lean
Spec is approved, implementation has not started	plan-write
Plan is approved, before any implementation dispatch	isolate
About to dispatch any subagent / Task / Agent call	brief-compiler (compile RTCO brief first)
Executing an approved plan task-by-task	exec-dispatch
Just received output from a subagent, tool chain, or dispatched task	report-filter
About to claim work is done, fixed, complete, or passing	verify-gate
verify-gate returned ready, before any outward-facing action	review-pack
A vibe run is complete and the user wants a security pass on the diff (optional, before finish-branch)	security-review
review-pack returned yes and user signed off on the diff	finish-branch
Driving a vibe run autonomously across iterations	ralph-loop
Storing, recalling, or auditing durable project knowledge	memory-dual
verify-gate returns not satisfied/partial, or an exec-dispatch task fails its test/build	debug-recovery
User reports a bug, failing test, broken build, or asks "why is this failing"	debug-recovery
User reports vibekit feels broken, skills not firing, or cross-runtime drift	vibekit-doctor

Hard gates (never bypass)

• No code before brainstorm-approved design. brainstorm-lean has a HARD-GATE; respect it on every project regardless of perceived simplicity.
• No "done" claim before verify-gate. Evidence before assertions, always.
• No merge / push / PR before review-pack sign-off. finish-branch is the only outward-facing endpoint.
• security-review is an optional peer gate. When the user runs it and it returns blocked (CRITICAL/HIGH finding), finish-branch is unavailable until the user fixes or waives — the same block-halts-handoff rule as review-pack.
• No subagent dispatch without a compiled brief. brief-compiler runs first; report-filter runs on return.

How to access skills

• Claude Code / Copilot CLI: Skill tool. The skill content loads — follow it directly. Never Read a skill file as a substitute for invoking it.
• Codex: skills are referenced from AGENTS.md; invoke by following the named workflow.
• Gemini CLI: activate_skill tool.
• opencode: see the per-runtime adapter shim.
• Pi: prompts/vibe.md is registered via the pi key in package.json; skills auto-discovered from skills/ and primed by the .pi-plugin extension. Invoke via /skill:<name> or by following the auto-trigger map.

Compression policy reminder

Compress agent chatter, briefs, capped reports, progress logs. Never compress questions to the user, user answers, plan files, TDD markers, verification evidence, error quotes, code blocks, destructive-operation warnings, or final commit/PR text. If a compression would reduce guardrail clarity by any amount, do not compress.

Anti-patterns

• "This is too simple to need brainstorm." → Wrong. Run brainstorm-lean. The design can be three sentences, but it must exist and be approved.
• "I'll skip verify-gate, the diff is obviously right." → Wrong. The diff being obvious is not evidence. Run the verification.
• "I'll dispatch this subagent with a quick inline prompt." → Wrong. Run brief-compiler to produce an RTCO brief, every time.
• "The subagent's report looks fine, I'll just use it." → Wrong. Run report-filter to enforce schema and strip restatements.