Expertise you can
snap on.
Carapaces, skills, and tools — three layers of composable knowledge that give any bot instant domain expertise without redeployment.
Three layers of composable knowledge.
A code review bot uses the code-review carapace. A research assistant uses the researcher carapace. A DevOps monitor uses a custom carapace with monitoring skills. You can create your own for any domain.
Expertise bundles
A carapace packages skills, tools, and a system prompt fragment into a reusable unit.
Assign carapaces: [qa, code-review] in bot YAML and the bot gains
those capabilities instantly.
Knowledge documents
Markdown files that give bots domain knowledge. Three injection modes:
pinned (always loaded), rag (semantic match),
or on_demand (fetched via get_skill()).
Agent capabilities
Python functions, MCP servers, and client-side actions. Tool RAG selects only the most relevant tools per query via cosine similarity.
Fragment-as-Index architecture.
A carapace's system prompt fragment doesn't dump all knowledge into context. It acts
as a routing table — telling the bot when to fetch each skill via
get_skill(). Bots can have thousands of pages of expertise without
context window bloat.
# system_prompt_fragment acts as routing table:
When the user asks about test coverage
or mentions regression testing:
→ get_skill("qa-test-planning")
When reviewing a pull request
or asked for code feedback:
→ get_skill("code-review-checklist")
When debugging a failure
or investigating a stack trace:
→ get_skill("debugging-methodology")Composable expertise bundles.
12 carapaces ship with Spindrel — 12 built-in and 0 from integrations. Assign them to any bot or use them as templates.
Core
Full orchestration capability — delegation, channel management, system administration, and cross-bot coordination. Apply to any bot to make it an orchestrator via channel carapaces_extra.
Development
Systematic debugging and bug resolution — reproduce, isolate, assess blast radius, test, fix, verify, and document.
Structured code review focusing on correctness, security, and maintainability. Provides actionable feedback with severity ratings.
Full quality assurance workflow — test planning, execution, coverage analysis, and bug reporting. Includes code review via composition.
Research
Multi-angle web research with cross-referencing and citations. Structured briefings with source reliability assessment.
Fast file scanning and pattern extraction. Cheap bulk reading: find usages, list endpoints, extract config values.
Compress large inputs into concise digests. Conversation history, docs, status updates, verbose tool output.
Creative
Image generation, editing, and visual analysis. Handles iterative creation workflows: generate from prompts, edit uploaded images, chain edits across turns, compare versions, and deliver polished results. Apply to any vision-capable bot.
Marp slide deck creation. Takes a brief and produces polished presentations (HTML/PDF/PPTX) with visual focus.
Long-form content creation: docs, reports, proposals, READMEs. Audience-aware, polished output with concrete examples.
Domain
Sourdough, pastry, and fermentation assistant — real-time bake guidance, starter tracking, photo-based dough assessment, recipe scaling, and proactive scheduling. Sees photos natively for instant visual diagnosis.
Seasonal gardening assistant — plant tracking, care scheduling, pest diagnosis, project planning, and proactive reminders. Equipped with research, scheduling, weather, and a persistent garden journal.
Knowledge documents that ship with Spindrel.
9 core skills available in skills/. Integrations provide additional skills
auto-synced from integrations/*/skills/.
Core
context_mastery Context window management, reference file authoring, research delegation patterns, and cold start sequences.
delegation Using delegate_to_agent effectively with model tiers — when to delegate vs self-handle, tier selection, and common patterns.
Workspaces
channel_workspaces Operational guide for managing persistent state files in channel workspaces — active vs archived files, schemas, and cross-channel search.
workspace_files File tool guide for reading, writing, and editing workspace files — when to use file vs exec_command, memory patterns, and edit safety.
workspace_member Guide for bots operating in shared workspaces — executing commands, calling server API, managing files, and coordinating with other bots.
Files
attachments File attachment handling and image processing — generating, editing, sending files, referencing uploads, and vision analysis.
Harnesses
claude_code_harness Delegate work to Claude Code via the harness system — execution modes, prompt construction, response parsing, and error handling.
History
search_history Searching historical channel messages across all sessions — tool signatures, result shapes, filtering, and proactive usage.
Security
prompt_injection_and_security Prompt injection detection and defense for untrusted input — trust hierarchy, pattern recognition, sanitization, and response patterns.
Installable capability packages.
packages/slides/
Marp slide deck generation — create polished HTML, PDF, and PPTX presentations.
Provides the create_slides tool and slide design skills used by the Presenter carapace.
Packages live in packages/ and can be installed independently.
More packages coming soon.
Create your own carapace.
Define a YAML file with skills, tools, and a system prompt fragment.
Drop it in carapaces/ or integrations/*/carapaces/
and it's available on the next restart.
# carapaces/my-expert/carapace.yaml
id: my-expert
name: My Expert
description: Domain expertise for ...
includes:
- researcher
skills:
- id: domain-guide
mode: on_demand
tools:
- domain_search
- domain_action
system_prompt_fragment: |
When the user asks about ...
→ get_skill("domain-guide")