Notre avis

Spindle effectue une classification et un tri rapides de fragments provenant d'une étape de dérive antérieure, en les catégorisant par pertinence et intensité de signal pour filtrer le bruit et extraire des thèmes en vue d'un traitement ultérieur.

Points forts

Classification rapide avec un système de catégories clair
Sélectivité personnalisable par cycle
Traitement spécial des lacunes et frictions
Extraction de thèmes à partir de fragments connexes

Limites

Impossible de générer de nouveaux fragments
Classification basée sur un contexte limité
Les seuils de sélectivité peuvent écarter des fragments utiles s'ils sont trop stricts

Quand l'utiliser

À utiliser lorsque vous disposez d'un grand ensemble de retours utilisateurs ou de fragments mélangés qui doivent être triés par pertinence pour une compétence ou un domaine spécifique.

Quand l'éviter

À éviter lorsque vous devez générer de nouvelles idées ou lorsque les fragments sont déjà bien organisés et ne nécessitent aucun filtrage.

Exemples

Classify and filter user feedback fragments

I have a bag of fragments from the drift stage containing user feedback about my AI agent. Run spindle with selectivity=0.6 to classify each fragment as HIT, EDGE, GAP, FRIC, ADJ, or NOISE, assign signal strength, and produce a sorted manifest with themes.

Extract themes from support tickets

Take these 20 support ticket summaries and apply spindle classification. Group fragments that touch the same area into themes with priority levels. Flag any contradictions where the skill documentation says one thing but users expect another.

name: spindle description: "Stage N2. Rapid classification bursts — sorting, tagging, filtering fragments. The workhorse stage at ~45% of cycle time. Most fragments die here." context: fork user-invocable: false allowed-tools:

Read

Spindle — Stage N2

The sorting room. Takes the noisy fragment bag from drift, classifies each fragment, kills irrelevant noise, extracts themes.

Inputs

From orchestrator: fragment_bag (from N1), target_skill (path), selectivity (0-1, default 0.6), cycle_number.

Procedure

1. Load Target Skill

Read SKILL.md. Build mental model of: what it does, what it handles, what it assumes, what it declines.

2. Classify Each Fragment

One category + one signal strength (0-1) per fragment:

| Category | Code | Description | |----------|------|-------------| | Direct hit | HIT | Skill explicitly covers this | | Edge case | EDGE | Within domain but near boundaries | | Gap | GAP | Within domain but NOT covered | | Friction | FRIC | User difficulty with documented workflow | | Adjacent | ADJ | Related domain, different skill | | Noise | NOISE | Irrelevant or unrealistically synthetic |

3. Filter

Survival threshold = selectivity × 0.8. Exceptions:

FRIC and GAP survive at signal > 0.3 (always interesting)
Synthetic fragments get +0.1 bonus (give creativity a chance)

| Cycle | Selectivity | Expected Kill Rate | |-------|------------|-------------------| | 1 | 0.7 | 60-70% | | 2 | 0.5 | 40-50% | | 3 | 0.3 | 20-30% | | 4+ | 0.2 | 10-20% |

4. Extract Themes

If 2+ fragments touch the same area, elevate to a theme with priority (high/medium/low). Themes are primary input for N3.

Flag contradictions (skill declines something users expect) and escalations (fundamental structural issues).

5. Output Sorted Manifest

sorted_manifest:
  target_skill: /path/to/SKILL.md
  input_fragments: 14
  surviving_fragments: 6
  kill_rate: 0.57
  themes: [{name, fragment_ids, priority, reason}]
  survivors: [{id, category, signal, content, tags, theme, notes}]
  discarded: {count, categories: {NOISE: 4, ADJ: 2, HIT: 2}}
  escalations: []

Behavioral Rules

Be decisive. If you can't classify quickly, it's NOISE.
Favor false positives. Better to pass mediocre fragments to N3 than kill one that would have produced insight in REM.
Don't generate. Spindle classifies, it doesn't create new fragments.

Spindle - Classification Rapide

Recommandé pour