Slide Analyzer — observabilité sk-agent (Marp vs PPTX + prompts)
Origine : fichier
.claude/agent-memory/slide-analyzer/MEMORY.md(relocalisé en c.1301+210 dans le cadre de l’EPIC #9535, item 7 — agent-memory relocalisation en doc pérenne). Le contenu est conservé tel quel (anti-régression : cf triage jsboigeEpita, durable vs scratch) ; seul un sommaire FR est ajouté.Statut : durable, avec un outil périmé. sk-agent n’expose plus d’outil
analyze_image(mesuré 2026-09-29 : le serveur déclare 9 outils, dontcall_agent(prompt, attachment, model_override, ...), et aucunanalyze_image) — l’appel est à adapter encall_agent(prompt=..., attachment=<chemin de l'image>). Le contenu documente (a) le comportement observé desk-agent analyze_image(routageglm-4.6v, indépendamment du paramètremodel) ; (b) deux prompts canoniques (primaire FR + retry FR court) qui contournent les hallucinations sur slides logo-heavy — toujours valables, l’outil de vision ayant simplement changé de nom et de signature ; (c) une grille de comparaison Marp vs PPTX (deck01-introduction) qui sert de baseline pour les audits futurs de slides. Ce sont des invariants d’environnement, pas du scratch daté.
Sk-agent analyze_image : comportement observé
Model Used
- sk-agent routes requests to
glm-4.6vregardless of themodelparameter value. - Specifying
qwen3-vl-8b-thinkingdoes NOT change the underlying model.
Hallucination Risk on slide.017
- On the “Qui fait de l’IA” logo-grid slide, the standard English prompt triggered a hallucination: the agent returned a fake web URL (
searxng-web_url_read) instead of analyzing the image. - Root cause likely: logo-heavy slides with many brand images confuse the model’s tool-calling heuristics.
- Fix: use a shorter French prompt — this reliably forces direct image description.
Prompt Effectiveness (tested 2026-02-21)
| Prompt style | Result | Notes |
|---|---|---|
| Long English (5 questions + rating 1-5) | Works for clean slides, fails on logo-heavy | slide.003 and slide.033 OK, slide.017 hallucinated |
| Short French “Decris les visuels… note /10” | Reliable fallback | Used for retry, succeeded on slide.017 |
Recommended Primary Prompt (French, concise)
TEXTE EXTRAIT:
{texte}
---
Analyse UNIQUEMENT la mise en forme et les visuels.
1. VISUELS: Diagrammes, images, icones - lesquels, qualite ?
2. MISE EN FORME: disposition, equilibre texte/visuel
3. LISIBILITE: note /10 pour projection amphitheatre
4. 2 SUGGESTIONS concretesRecommended Retry Prompt (on empty/hallucinated response)
Decris les visuels de cette slide et note la lisibilite /10.PPTX vs Marp Rendering Quality (deck 01-introduction)
General observations
- Marp strips French diacritics inconsistently: “réflexe” -> “reflexe”, “académique” -> “academique”
- Marp renders logos as image blocks at bottom (not inline with text as in PPTX)
- Marp loses the PPTX two-column layout: text and diagrams stack vertically instead of side-by-side
- Marp footer “I - Introduction” is visible but truncated on slide.003
Slide-specific issues
- slide.003 (Sommaire): Marp version cuts the bullet list (content truncated after “Apprentissage”), book image is large and fills the right half but text list is incomplete.
- slide.017 (Qui fait de l’IA): Marp version loses inline logos; logos move to a separate bottom row, Fujitsu logo partially cut off at bottom edge.
- slide.033 (Agent reflexe): Marp version is the best conversion - two-column layout preserved, diagram and pseudocode box visible; book image small but present. Minor: accents stripped.
File Locations
- PPTX renders:
slides/01-introduction/extracted/renders/slide_NN.png - Marp renders:
slides/01-introduction/output/marp_renders/slide.NNN.png - Extracted text:
slides/01-introduction/extracted/content.md - Analysis output:
slides/01-introduction/analysis/visual_review.md