aice_skill

This skill evaluates user and agent confidence across TECH OPS JUDGMENT COMMS ORCH domains, delivering dual scoring to improve collaboration.

Python

2.6k

GitHub Stars

5

Bundled Files

2 months ago

Catalog Refreshed

3 months ago

First Indexed

Readme & install

Copy the install command, review bundled files from the catalogue, and read any extended description pulled from the listing source.

Installation

Preview and clipboard use veilstrat where the catalogue uses aiagentskills.

npx veilstrat add skill openclaw/skills --skill aice

_meta.json287 B
CHANGELOG.md2.1 KB
confidence.template.json4.4 KB
README.md748 B
SKILL.md13.9 KB

Overview

This skill implements the AI Confidence Engine (AICE), a bidirectional scoring system across five domains: TECH, OPS, JUDGMENT, COMMS, and ORCH. It produces agent and user scores, pool-level aggregation, and runtime-aware team metrics to surface trust, maturity, and actionable recovery plans. Designed for live scoring, hub sync, and automated post-task evaluations.

How this skill works

AICE tracks events and triggers (puntúa, auto-score, task-complete, idea-validate, criteria-evolution) to compute per-domain deltas, streaks, clusters, and caps. Scores range −100% to +100% and combine domain weights into a global score; pools aggregate by runtime. The engine logs anti-/pro-patterns, enforces ops rules (noise reduction, escalation), and optionally syncs minimal event metadata with a Hub while preserving privacy.

When to use it

After completing a task to auto-evaluate performance (task-complete)
When a correction or validation occurs to auto-score impact (auto-score)
To validate user ideas or evolving criteria (idea-validate, criteria-evolution)
To report and diagnose cross-agent delegation or pool issues
Before day close to seal scores and run maturity checks

Best practices

Use short, specific triggers (e.g., “puntúa”, “score”) to avoid duplicate scoring
Prefer ADR-like user inputs (what, why, scope) to minimize vagueness penalties
Apply pro-patterns (ANTICIPATE, CLEAN_FIX, SMART_SILENCE) to gain consistent +3 deltas
Limit retries and push resolved events only; follow anti-noise rule (2 silent retries then alternative)
When hub-sync active prefer hub cachedState for authoritative displays and keep local events anonymized

Example use cases

An agent auto-evaluates after finishing a code task and posts an event to update pool maturity
A user says “Eso estuvo bien” to apply a quick correct rating via natural language
Team lead queries /aice pool to inspect runtime-level scores and identify delegation failures
Agent detects SELECTIVE anti-pattern and triggers a trust-recovery plan with corrective steps
Registering a runtime with the Hub, then sending minimal per-event metadata for cross-platform leaderboard tracking

FAQ

Only minimal event metadata: domain, eventType, severity, patternCode, quadrant, trigger, sessionId, timestamp. No conversation content, prompts, or system instructions.

How does AICE handle repeated errors?

Reincidence in the same session escalates severity (up to −10). Escalation rules: correct → warning → enforcement → redesign, and a trust-recovery plan is required below 20% in a domain.