ohugonnot
UserBattle-tested Claude Code skills by Odilon Hugonnot: quality-gated feature delivery, senior review, issue/MR workflow, branch wrap-up, book distillation, context saving, human-sounding tech writing, and a learning mentor.
Categories
Indexed Skills (10)
book-distill
Reads a book (PDF or EPUB) and produces a faithful, pedagogical markdown reading note: maps the book, reads with intent, verifies every citation word-for-word against the text, distills by themes into linked ideas, with a reasoned rating and counterpoints. Use whenever the user provides a book file or asks to summarize or distill a book — "book summary", "distill this book", "reading note", « résume ce livre », « fiche de lecture », « extrais la substance » — even if they never say the word "note".
branch-wrap-up
Use when the user says the work is done and wants to close out a branch — orchestrates the wrap-up of already-written work: gathers uncommitted/branch changes, delegates the code review to senior-review (standard tier by default), proposes a conventional commit message in the project's own style, suggests push + MR/PR on GitLab (glab) or GitHub (gh) auto-detected from the remote, and runs a CLAUDE.md/memory knowledge-capture check. Propose-only on git: never runs add/commit/push itself, the user validates each action; never mentions AI. NOT for the deep review itself (use senior-review), nor for building a feature (use feature-loop), nor for creating the issue/branch/MR up front (use issue-mr). Flags: --no-review (review already done), --quick/--deep (tier passed through to senior-review).
feature-loop
Use when the user wants a non-trivial feature implemented with autonomous, iterative, quality-gated delivery — best when quality matters more than raw speed. A mother agent sizes the difficulty tier and delegates to SEPARATE subagents (code writer ≠ test-writer ≠ blind reviewer; tests written from the spec). Objective gate (build/lint/typecheck/tests + test-must-go-red mutation check) BEFORE any LLM review; runnable features get a mandatory LIVE smoke test before any "tested" SUCCESS. Loops until all quality axes ≥ 8/10 with zero criticals (max 3 iters), keeps the BEST version, emits a markdown report. Subcommands: status (runs dashboard), learn (propose-only self-improvement). Flags: --fast, --paranoid, --worktree, --max-iter=N, --threshold=N, --judge, --issue=N (GitLab/GitHub issue as spec). NOT for quick edits (do them directly) nor review-only (use senior-review).
issue-mr
Use when the user wants to turn a task description into a well-formed issue with its branch and MR/PR — on GitLab (glab) or GitHub (gh), auto-detected from the remote. Three behaviors: SCAFFOLD (clean tree → issue + branch + MR/PR shell), ISOLER (uncommitted work gets isolated into its own issue/branch/MR — file list and commit message validated first), ANALYSE (--analyse or vague/non-trivial task: explore the code, settle design with the user, produce a structured spec issue — constat/pourquoi/périmètre/plan — ready to feed feature-loop --issue=N). Discovers project conventions live (labels, default branch, commit scopes) instead of hardcoding; validates everything with the user BEFORE creating anything; never mentions AI. Flags: --analyse, --issue-only. NOT for implementing the task (use feature-loop) nor for wrapping up an existing branch (use branch-wrap-up).
rediger-cir
Rédige le dossier justificatif technique du Crédit d'Impôt Recherche (CIR) d'une entreprise pour une opération de R&D donnée, autour du verrou scientifique ou technique qui doit tenir face à un contrôle fiscal. À utiliser quand l'utilisateur veut produire, compléter ou noter un CIR (ex. « on fait le CIR 2025 », « rédige la fiche-opération », « note ce dossier »). Connaît la trame officielle 2025 (CIROCO, 4 sections, Frascati), le style technico-administratif et le fil rouge pluriannuel, génère les figures (SVG) et auto-note le dossier sur un radar 8 axes.
senior-review
Use when the user wants a thorough senior-level fresh-eyes review of changes (working tree, staged, branch, or PR) before merge: bug hunting, security threat-model, design, tests, spec-alignment vs the linked ticket. Use it whenever the user says things like "review my changes", "relis mon code", "revue de code", "can I merge this", "is this ready to ship", "check this MR/PR", even if they never say "senior-review". Research-grounded pipeline: context-first assembly (cross-file call sites, conventions, git history, ticket), objective tool gate, parallel blind reviewers (reviewer is never the author), verification gate (every critical finding needs a receipt: grep, execution, or a red test), adversarial refute-panel on critical/uncertain findings only. Optimizes signal/noise over recall: caps nits, explicit what-NOT-to-flag, silence when nothing material. Tiers: --quick / standard / --deep. Flags: --pr N, --base ref, --staged, --ticket N, --security, --fix, --comment, --no-tools. NOT for building features (use
code-mentor
Use when the user wants to learn while coding, not just get code shipped — slows down at each important decision, explains it from zero (the why, not just the what), leaves the thinking to the user, and saves a course-like markdown lesson for later review. Trigger on "teach me while we build this", "explain your choices as you go", "I want to understand, not just copy", "level up", "mentor mode", even when the word "mentor" never appears. NOT for when the user just wants the code fast (use a normal session).
skill-builder
Use when the user is designing, reviewing or debugging a Claude Code skill and needs the judgment behind a good SKILL.md — should this even be a skill (vs CLAUDE.md / subagent / hook / MCP), why a description does or doesn't trigger, how to keep the body skimmable, how to package it. Carries the opinionated principles plus a review checklist. Trigger on "review my SKILL.md", "why doesn't my skill trigger", "principles of a good skill", "should this be a skill or a hook", "is my skill description good", even when the word "skill" is only implied. NOT for interactive scaffolding from scratch (use Anthropic's Skill Creator), nor for general coding.
tech-article
Use when the user is writing or publishing a technical blog article that should read like a human engineer wrote it, not an AI. A story-first method — open on a real problem, keep opinions sharp, cut AI-slop phrasing, control sentence rhythm, and add a diagram only when it earns its place. Trigger on "write a blog article", "draft a dev post", "turn this into an article", "make this not sound like AI", even when the word "article" never appears. NOT for documentation, READMEs, or reference pages.
vide-contexte
Persists a conversation's non-obvious insights into memory files before its context is cleared or compacted — nothing is cleared by this skill, only saved. Use when the user says "vide le contexte", "clear context", "vide le context", "sauvegarde avant /clear", "save context before clearing" or similar, even when they only announce the /clear without asking to save. NOT for closing out a branch (use branch-wrap-up, which runs its own knowledge-capture pass).
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.