From vibe-taker
Shared behavior, persona, voice, and hygiene rules for the vibe-taker plugin's command skills. Internal-only — referenced by capture/plant/list skills before doing the command's work; not user-invocable.
How this skill is triggered — by the user, by Claude, or both
Slash command
/vibe-taker:guideThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
This file is loaded **first** by every command skill (`capture`, `plant`, `list`) before any work. It defines voice, persona handling, the hygiene rules every command honors, and the cross-references to the four operational reference files.
This file is loaded first by every command skill (capture, plant, list) before any work. It defines voice, persona handling, the hygiene rules every command honors, and the cross-references to the four operational reference files.
Surface area, in one sentence: vibe-taker reads files in the user's repo, writes files to a cross-repo shelf at
~/.vibe-taker/library/, and writes files into the user's repo only after the user explicitly confirms a diff.
vibe-taker speaks builder-to-builder, second person, sentence case — same baseline as the rest of the marketplace storefront copy. The agent's personality is the user's shared.preferences.persona from ~/.claude/profiles/builder.json. Defer to that field; fall back to baseline if absent.
~/.vibe-taker/library/bg-remover/v1/. 6 artifacts. Run :list to see the shelf." beats a paragraph of preamble.Every command skill applies these. They're load-bearing for v1 quality.
For deliverables longer than ~300 words (rare in vibe-taker — most output is short stdout), write directly to a file first, then reply with the path + 2-sentence summary + next action. Never both.
In practice for vibe-taker: bundle artifacts (README.md, architecture.md, notes.md) are always file writes; the stdout summary after capture is always short (≤10 lines).
Always verify pwd (or the equivalent) before running git, gh, or any cross-repo command after any cd. When working across multiple repos in one session, prefer absolute paths over relying on cwd.
In practice for vibe-taker: :capture reads from cwd; :plant writes to cwd. If the agent has done any cross-directory navigation in the session, it must verify cwd before resolving the target path.
When a sub-step's reading contradicts a prior reading or earlier-in-session conclusion, re-verify before incorporating. Don't paper over the gap; name the contradiction and resolve it with evidence.
Don't speculate about external system behavior (vendor API tiers, third-party rate limits, network state) without evidence. Say "I don't know" and ask the user.
Match the scope of the user's ask. If they want a quick capture of one file, don't pivot to architecture review. If they want full plant ceremony, don't degrade to a partial diff.
The four reference files in ./references/ carry the load-bearing operational data. Skills consult these instead of inlining the lists.
| Reference | What it carries | Consumed by |
|---|---|---|
references/bundle-schema.md | Bundle directory layout, contract.json fields, index.json shape, versioning rules. | capture, plant, list |
references/secret-patterns.md | Glob patterns matched at capture-time and skipped from reference/. Load-bearing detection rules. | capture |
references/stack-match.md | Framework-family taxonomy + the high/low/hard match table. Decision tree for code-lift vs spec-driven vs decline. | plant |
references/interview-gate.md | When the interview fires, what it asks, the substantive-item heuristic. | capture |
references/error-contract.md | Three exit classes (0/1/2) with required recovery-message discipline. | capture, plant, list |
Schemas live in ./schemas/; artifact templates live in ./templates/.
Two skill folders ship as documented placeholders in v1: skills/session-logger/ and skills/friction-logger/. Each documents its contract and reserved data path. Command skills do not invoke them in v1. Reserved for v2 lighting-up; reserved on disk in v1 so v2 doesn't need a directory restructure.
Pattern #13 (ecosystem-aware composition) is implemented in v1 — the plant skill checks for mcp__626Labs__manage_decisions at the end of a successful plant, calls it if present, succeeds silently if absent.
Every command skill opens with the equivalent of:
1. Read this file (`skills/guide/SKILL.md`) — voice, hygiene, references.
2. Read the appropriate `references/*.md` files for the command's job.
3. Do the command's work.
4. Print the outcome per the [error contract](./references/error-contract.md).
The reference files are read on demand — capture doesn't load stack-match.md, plant doesn't load secret-patterns.md. Keep the working set small.
Provides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub estevanhernandez-stack-ed/vibe-taker --plugin vibe-taker