Atelier Litteraire : Redaction litteraire, avec idéation et processus itératif. plusieurs agents : idéation, rédaction, jury/correcteurs.
PassAudited by ClawScan on Apr 30, 2026.
Overview
This is a local, instruction-only literary writing workflow with no code or install scripts, but it does create local project files and uses persistent project memory and multi-agent handoffs.
This skill appears appropriate for local literary ideation and drafting. Before installing, be aware that it may create and retain local project files, read a MEMORY.md file, pass real-world source dossiers between Hermine and Herbert, and use Anthropic models for processing. Use a dedicated workspace, review retained memory, and avoid providing unnecessary sensitive personal details.
Findings (6)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Once started, the workflow may run multiple model-intensive steps before asking for a later creative choice.
The skill can automatically continue to an Opus-based reasoning step after H3 without waiting for a separate approval. This is disclosed and aligned with the ideation workflow, but it can consume model resources autonomously after invocation.
Déclenchement : Systématique, immédiatement après H3. Toujours produit, sans attendre la validation de l'auteur sur les axes 1-2-3. Modèle : Opus
Confirm model budget and stopping points before starting a large project, especially if Opus usage has cost or quota implications.
Users have less external provenance information to verify the package origin and exact published version.
The registry provenance is limited, and the packaged openclaw-skill.json declares version 1.0.6. There is no executable install path, so the practical risk is limited, but the source/version mismatch is worth noticing.
Source: unknown; Homepage: none; Version: 1.0.5
Install only if you trust the registry publisher, and compare the registry version with the packaged openclaw-skill.json before relying on it in important workflows.
Old project notes or poisoned memory content could influence future drafts, and sensitive writing context may persist locally longer than expected.
Herbert is instructed to read and use persistent long-term memory and current project context. This is coherent for iterative literary work, but persistent memory can carry sensitive drafts or stale instructions into later sessions.
Lire `MEMORY.md` — les projets en cours et les leçons apprises ... Mémoire long terme dans `MEMORY.md`
Use a dedicated project workspace, review MEMORY.md periodically, and remove sensitive or outdated memory entries between unrelated projects.
Private or real-person source material may be shared between the skill's agents as part of the writing pipeline.
The skill intentionally passes a dossier of real-world facts from Hermine to Herbert. The handoff is purpose-aligned and the file says the dossier is confidential, but it is still a sensitive inter-agent context transfer.
Hermine livre à l'auteur et à Herbert : ... Le dossier du réel — transmis à Herbert pour nourrir ses bibles
Avoid providing unnecessary personal details, and review the dossier before passing it from ideation into drafting.
If the initial structure is wrong, many generated sections may need revision before the issue is caught.
The workflow is long and iterative, so an early bad assumption in the pitch, bible, outline, or source dossier can propagate through many units. The pipeline includes jury scoring and coherence checks, which are useful containment measures.
BOUCLE × 40-50 seeds ... Iter 1 → Iter 2 → Iter 3 ... CONTRÔLE DE COHÉRENCE — tous les 5 seeds
Review and approve the pacte narratif, bibles, outline, and first few seeds carefully before letting the workflow continue through all 40–50 units.
Drafts, pitches, or real-world source material may be processed by the configured model provider even though final files are saved locally.
The local-output wording is paired with a clear cloud model requirement. This is not deceptive in context, but users should not mistake local DOCX/PDF generation for fully offline processing.
This skill operates locally by default ... Requirements: Anthropic Claude API (Sonnet for standard tasks, Opus for H4 counter-axes)
Read 'locally' as referring to saved outputs, and check your model provider's data handling terms before entering highly sensitive material.