Mathias STRASSER
aedde6707e
Les enseignants ont besoin de moyennes à jour immédiatement après la publication ou modification des notes, sans attendre un batch nocturne. Le système recalcule via Domain Events synchrones : statistiques d'évaluation (min/max/moyenne/médiane), moyennes matières pondérées (normalisation /20), et moyenne générale par élève. Les résultats sont stockés dans des tables dénormalisées avec cache Redis (TTL 5 min). Trois endpoints API exposent les données avec contrôle d'accès par rôle. Une commande console permet le backfill des données historiques au déploiement.
3.0 KiB
Quality Dimensions — Quick Reference
Seven dimensions to keep in mind when building skills. The quality scanners check these automatically during quality analysis — this is a mental checklist for the build phase.
1. Outcome-Driven Design
Describe what to achieve, not how to get there step by step. Only add procedural detail when the LLM would genuinely fail without it.
- The test: Would removing this instruction cause the LLM to produce a worse outcome? If the LLM would do it anyway, the instruction is noise.
- Pruning: If a block teaches the LLM something it already knows — scoring algorithms for subjective judgment, calibration tables for reading the room, weighted formulas for picking relevant participants — cut it. These are things LLMs do naturally.
- When procedure IS value: Exact script invocations, specific file paths, API calls with precise parameters, security-critical operations. These need low freedom because there's one right way.
2. Informed Autonomy
The executing agent needs enough context to make judgment calls when situations don't match the script. The Overview establishes this: domain framing, theory of mind, design rationale.
- Simple utilities need minimal context — input/output is self-explanatory
- Interactive/complex workflows need domain understanding, user perspective, and rationale for non-obvious choices
- When in doubt, explain why — an agent that understands the mission improvises better than one following blind steps
3. Intelligence Placement
Scripts handle plumbing (fetch, transform, validate). Prompts handle judgment (interpret, classify, decide).
Test: If a script contains an if that decides what content means, intelligence has leaked.
Reverse test: If a prompt validates structure, counts items, parses known formats, compares against schemas, or checks file existence — determinism has leaked into the LLM. That work belongs in a script.
4. Progressive Disclosure
SKILL.md stays focused. Detail goes where it belongs.
- Stage instructions →
./references/ - Reference data, schemas, large tables →
./references/ - Templates, config files →
./assets/ - Multi-branch SKILL.md under ~250 lines: fine as-is
- Single-purpose up to ~500 lines (~5000 tokens): acceptable if focused
5. Description Format
Two parts: [5-8 word summary]. [Use when user says 'X' or 'Y'.]
Default to conservative triggering. See ./references/standard-fields.md for full format.
6. Path Construction
Only use {project-root} for _bmad paths. Config variables used directly — they already contain {project-root}.
See ./references/standard-fields.md for correct/incorrect patterns.
7. Token Efficiency
Remove genuine waste (repetition, defensive padding, meta-explanation). Preserve context that enables judgment (domain framing, theory of mind, design rationale). These are different things — never trade effectiveness for efficiency. A skill that works correctly but uses extra tokens is always better than one that's lean but fails edge cases.