SkillWeaveCraft intelligence from small, focused pieces.

A runtime — and emerging open standard — for composing LLM tasks from small, focused, testable micro-skills. The reference pipeline follows Ask → Validate → Judge → Learn.

Quick start →

Architecture

GitHub

🧩

Micro-skills, not mega-prompts

LLMs produce better, more reliable output given a narrow, well-defined job. Each skill declares does / does_not, an input/output contract, and its own assertions.

Skill contract →

⚖️

A boundary judge at every probabilistic step

The orchestrator auto-inserts a groundedness judge after every probabilistic skill — no manual wiring. Deterministic skills carry zero overhead.

Reliability layer →

🔁

Retry with negative context

A failing skill is re-invoked with its prior output and the failure reason (budget 2), then halts with full diagnostics if it can't recover.

How retry works →

🔌

Provider-neutral with automatic fallback

A provider-adapter layer (Claude · Gemini · OpenAI · local Ollama) with capability profiles, tier-based model selection, and primary→fallback — plus a deterministic offline heuristic.

Provider layer →

📊

SigMap adapters + health grading

Wraps SigMap's CONTEXT / COST / OBSERVE primitives over local artifacts (no shell-out). A composite 0–100 health score grades every run on SigMap's A–D scale.

SigMap adapters →

🌱

Built on a proven pattern

SkillWeave generalises SigMap's production ask → validate → judge → learn workflow from code to any domain.

Primitives →

The chain

The reference pipeline maps SigMap's proven ask → validate → judge → learn pattern onto documents:

parse-input  →  validate-coverage  →  extract-highlights  →  memory-update
(sigmap ask)    (sigmap validate)     (probabilistic)         (sigmap learn)
                                      ↑ auto-judged · confidence-routed · retried

60-second start

bash

npm install
npm start                            # run the built-in chain (offline heuristic judge)
npm start -- --inject hallucination  # ungrounded → judge RETRIES → recovers
npm test                             # node:test suite (137 tests)

Or drive it from the skillweave CLI:

bash

npm run cli -- doctor                # readiness report — start here if you're new
npm run cli -- run pipelines/document-grounding.pipeline.yaml
npm run cli -- health                # composite 0–100 health score + grade
npm run cli -- providers             # provider/model capability table
npm run cli -- visualise pipelines/document-grounding.pipeline.yaml  # ASCII/Mermaid diagram
npm run cli -- validate pipelines/document-grounding.pipeline.yaml
npm run cli -- list

To run the boundary judge on a real model, set one of ANTHROPIC_API_KEY, GEMINI_API_KEY, or OPENAI_API_KEY — see the Multi-LLM judge guide.

Where to go next

New here: Quick start · Architecture
The model: Primitives · Skill contract
The v0.2.0 headline: Reliability layer
Running it: Multi-LLM judge · CLI
What's next: Roadmap