BotShelf Vampire — AI Team Registry

Public source for the BotShelf Vampire Build Library and its AI Team implementation/evidence layer: prompt packs, workflow recipes, code sketches, and MCP-oriented tool guidance for local and self-hosted AI workflows. Each evidence record is designed to make the job, runtime, exact implementation, verification state, human approval boundary, failures, and known limits inspectable.

Browse the canonical Build Library · Canonical team pages · Machine-readable registry · AI-readable project map · CodeMeta 4.0

Choose a job

| Job | Source directory | Typical input | | --- | --- | --- | | Research a question from sources | Deep Research | Question and labeled sources | | Review a code change | Coding Review | Diff and expected behavior | | Review documentation | Doc Review | Draft and prerequisites | | Analyze pasted data | Data Analysis | Table or CSV excerpt | | Draft from notes | Writing | Notes and intended audience | | Prepare a search-focused brief | SEO Brief | Keyword and page context | | Triage a support request | Support Triage | Message and escalation rules | | Review a status update | Monitoring | Status or log excerpt | | Build a lead brief from public facts | Lead Research | Public facts you provide | | Prioritize a day | Personal Assistant | Tasks and constraints |

Choose an implementation

Open the team's directory, then choose the runtime you already use:

| Runtime | Materials | What to expect | | --- | --- | --- | | Ollama | Modelfile and prompt | Local model instructions | | LM Studio | System prompt and local-server example | Local prompt workflow | | Open WebUI | System prompt and knowledge notes | Instructions to adapt in your UI | | n8n | Workflow JSON and setup notes | Recipe requiring configuration | | CrewAI | Python sketch | Multi-agent starting point to adapt | | LangGraph | Python graph sketch | Stateful workflow starting point | | MCP | Prompt/tool guidance and server hints | Integration guidance; inspect the files for what is implemented |

Runtime directories are variants of a job, not additional distinct teams. A prompt pack is not automatically a multi-agent system, and an MCP-oriented pack is not necessarily a deployable MCP server. Current inventory and status belong in the canonical registry, rather than a fixed marketing count here.

First run

1. Pick one job and inspect its team.yaml, README, and runtime files.
2. Check the prerequisites and replace example settings for your environment.
3. Start with a small, non-sensitive input whose expected result you can inspect.
4. Compare the output with that input. Record the model, runtime version, source revision, output, and any failure.
5. Keep actions that affect other people, accounts, money, or production behind the job's human approval boundary.

For a concrete local starting point, open Deep Research for Ollama. Its Modelfile contains the build and run commands. The system prompt asks for findings, open questions, and next checks without inventing citations.

Ollama walkthrough · Run-evidence template

For workflows split across writer and reviewer agents, use the exact-artifact review gate: copyable prompts, a dependency-free Python checker, and failure-case tests. It catches a draft changed after review; it does not authenticate reviewers or authorize publication.

Verification status

Treat an implementation as Untested unless evidence identifies the exact runtime, model, version, input, and result. Source files and website status can differ between revisions. A website's verification badge does not automatically verify this GitHub snapshot.

A JSON parse, Python syntax check, or file-presence check is structural validation. It is not proof that a model or workflow executed correctly. Prompt instructions also do not replace permissions enforced by the runtime.

What the evidence layer records

Canonical BotShelf pages are designed to make implementation claims inspectable before adoption. The machine-readable evidence-record schema defines the fields needed to pin a claim to a job, implementation, source revision, runtime, model, checks, approval boundary, failures, limits, and supporting artifacts. Schema validity makes a record easier to inspect; it does not by itself prove accuracy or safety.

The artifact-gate evidence record is a conservative machine-readable example: it is marked Partial, records the checks that ran, keeps unrun model and publisher checks visible, and does not promote a deterministic unit test to a verified AI workflow.

The Research Desk failure record preserves the September 2026 result as Blocked: 0 of 9 answers were accepted, the observed attribution and numerical failures remain visible, and unexecuted external actions are not presented as successful.

For an evidence-first overview designed for independent editorial review, see the public media brief.

For consistent directory submissions, use the public listing kit, its machine-readable directory manifest, the consolidated directory route status, and its machine-readable status data. The status index separates verified free paths, paid-only routes, blockers, proposed product surface, and current submission state; public CI checks the JSON schema and lifecycle counts.

For the current free Launching Next route, use the verified submission package. It remains Prepared / Not Submitted until a final duplicate, contact, fee, and approval check is completed.

For Product Hunt, use the free-launch route audit and owner-review package. The Marketplace is the proposed product, the Build Library remains separate supporting evidence, advertising is excluded, and the status is Audited / Prepared / Not Submitted / Not Scheduled.

For Hacker News, use the Show HN route audit. The Build Library is the proposed technical artifact and the Marketplace remains separate. The package is Audited / Prepared / Not Submitted; posting requires an owner-authorized maker to rewrite the comment in their own words, participate without generated text, and never solicit votes or comments.

For Indie Hackers, use the Products Database route audit. The Marketplace is the proposed product and the Build Library remains separate supporting evidence. The signed-out Add Your Product route requires login and does not explicitly confirm a $0 submission, so the status is Audited / Prepared / Blocked / Not Submitted.

For Microlaunch, use the route audit. Paid options are verified, but the current public pages do not explicitly confirm that the Regular route is free, so the status remains Audited / Not Submitted.

For AlternativeTo, use the free-queue submission package. The Marketplace is the listing target, the Build Library remains separate evidence, and the status is Prepared / Not Submitted.

For Dev Hunt, use the free-slot route audit. Weeks below the platform's capacity threshold are free, full weeks are $49, and the status is Audited / Not Submitted.

For SaaSHub, use the route audit. The normal submission tool is officially described as free, the featured placement is $99/month, and the status is Audited / Not Submitted.

For StartupBase, use the route audit. The standard no-badge community queue is free with an estimated 10+ week wait; badge verification and paid tiers remain excluded, and the status is Audited / Not Submitted.

For Tiny Startups, use the route audit. The official free review queue is separate from paid upgrades; the Marketplace is the proposed listing target and the Build Library remains its separate evidence layer. Status: Audited / Not Submitted.

For SideProjectors, use the free-route audit. Basic project submissions are officially free, the Build Library is prepared as a Showcase while the Marketplace remains separate, and authenticated duplicate checking is still required. Status: Audited / Prepared / Not Submitted.

For Open Launch, use the route audit. Free launches are currently fully booked into 2027, so the available $12 Premium and $59 SEO paths are excluded. Status: Audited / Paid-Only Currently / Not Submitted.

For Micro SaaS Examples, use the route audit. The current official form shows only $3, $29, and $147 paid tiers, so no submission or backlink was made. Status: Audited / Paid-Only / Not Submitted.

For Dev Resources, use the route audit. Its current terms describe a free curated directory and its official GitHub repository accepts rule-compliant contributions; paid advertising remains excluded. The Build Library is the proposed resource and the Marketplace stays separate. Status: Audited / Prepared / Not Submitted.

For Startup Buffer, use the route audit. Ordinary review is currently available without payment; paid priority and visibility packages remain excluded. The Marketplace is the proposed listing and the Build Library remains its separate implementation and evidence layer. Status: Audited / Prepared / Not Submitted.

For Toolfolio, use the route audit. The current public Submit tool route exposes only paid listing and promotion tiers; free claims apply only to existing listings, and no public BSV listing was found. The Build Library is the proposed future subject while the Marketplace remains separate. Status: Audited / Paid-Only for a New Listing / Not Submitted.

For Ctrl Alt CC, use the route audit. The public signed-out pages do not expose an official current fee schedule, and the site's terms prohibit automated or non-human access. No account or submission was created. The Build Library is the proposed subject while the Marketplace remains separate. Status: Audited / Route Not Officially Verifiable Without an Account / Not Submitted.

• Job: the outcome the AI Team is meant to produce.
• Runtime and implementation: the actual model/runtime path, required setup, and whether the material is a prompt, workflow recipe, code sketch, or tool integration.
• Status: Verified only for the exact run evidenced; otherwise Untested, Partial, Blocked, or Unverified as applicable.
• Human approval boundary: actions such as publishing, sending, spending, pushing, merging, or deploying remain behind an explicit human gate.
• Failure evidence: rejected outputs and unavailable seats remain visible instead of being converted into success claims.

Two public examples show the difference between a claim and evidence:

• Software Team evidence records a partial local run, unavailable seats, and a WAITING_ON_HUMAN stop before push, merge, or deploy.
• Research Desk evaluation preserves the tested configuration's failed evaluation rather than generalizing it to every model or runtime.

How this fits BotShelf Vampire

| Surface | Role | | --- | --- | | This GitHub repository | Copyable source and distribution for Build Library materials | | Build Library | Canonical job and implementation pages | | Ready-to-use Marketplace | Separate free and paid AI team product listings | | Marketplace prompt source | Free packs for Grok Bot, Claude Code, and ChatGPT |

Use a Library workflow, adapt it to a specific job, and document a real run before considering a Marketplace submission. Copying a runtime variant does not automatically create or publish a product.

Layout

teams/<team_id>/
  team.yaml
  README.md
  ollama/
  lm-studio/
  open-webui/
  n8n/
  crewai/
  langgraph/
  mcp/

License and links

The registry uses free-use-at-own-risk; read the LICENSE. It is not labeled MIT, Apache-2.0, or OSI-approved.

BotShelf Vampire · Build Library · X: @botshelfvampire

BotShelf Vampire is independent of the AI runtime vendors named here.