La feuille de route des écarts de compétences : transformer un poste cible en plan d’apprentissage sur 90 jours
Why this prompt matters
Self-directed career pivots fail most often not from lack of effort but from lack of sequencing — people spend months on the wrong material, build a portfolio project that doesn't map to what interviewers actually screen for, or discover three weeks before an interview that they never covered a prerequisite skill. A structured, prioritized roadmap with checkpoints turns an open-ended 'learn ML' goal into a plan that can be executed and audited week by week, and it tells you honestly, up front, if your timeline needs to change instead of leaving you overconfident walking into a role you're not ready for.
What we use it for
A backend engineer with four years of Python and PostgreSQL experience wants to move into a Machine Learning Engineer role within four months, has 8 hours a week to study, and has no idea whether to start with math fundamentals, a specific framework, or a portfolio project first.
Prompt
Act as a career learning strategist and technical mentor who has helped hundreds of professionals successfully transition into new roles and skillsets. Context: I currently work as [YOUR CURRENT ROLE AND KEY SKILLS], with [X YEARS] of experience. My goal is to become qualified for [TARGET ROLE OR SKILLSET] within [YOUR TIMELINE]. I can dedicate [HOURS PER WEEK] to learning, and I learn best through [YOUR LEARNING STYLE: hands-on projects / structured courses / reading and practice / pairing with a mentor]. Task: 1. Identify the specific skill gaps between my current profile and the target role. Be precise, not generic — 'learn Python' is not acceptable; name the actual sub-skills, tools, libraries, and depth of knowledge required. 2. Rank the gaps by priority: which gaps block progress on others (prerequisites), and which are 'must-have' versus 'nice-to-have' for the target role. 3. Build a week-by-week learning roadmap covering my full timeline, broken into clear phases (for example: foundations, applied practice, portfolio or proof-of-work, interview or evaluation readiness). 4. For each phase, recommend 2-3 specific high-quality resources (courses, books, official documentation, practice platforms, or project ideas) — prefer resources known for depth and accuracy over general popularity. 5. Define one concrete deliverable or checkpoint per phase that proves the skill was actually acquired — a shipped project, a working demo, a completed technical write-up, or a benchmark passed. Not 'finished watching the course.' Constraints: - Do not recommend more hours per week than I specified. - Do not give generic advice like 'practice more' or 'read the docs' without naming the specific practice method or documentation. - If my timeline is unrealistic for the skill gap you've identified, say so explicitly and propose a realistic alternative timeline instead of pretending the original one works. - Flag any skill gap that typically takes longer than my specified timeline to close credibly, so I don't walk into interviews or a new role overconfident. Output Format: ## Skill Gap Analysis [Table: Skill | Current Level | Target Level | Priority (Must-Have / Nice-to-Have) | Blocks Which Other Skills] ## Roadmap ### Phase 1: [Name] (Weeks X-Y) - Focus: - Resources: - Checkpoint/Deliverable: ### Phase 2: [Name] (Weeks X-Y) [repeat structure for each phase through the full timeline] ## Reality Check [Honest assessment of whether the stated timeline is achievable given the identified gaps, and what to deprioritize or extend if not]
Result
Skill Gap Analysis
| Skill | Current Level | Target Level | Priority | Blocks |
|---|---|---|---|---|
| Linear algebra & probability for ML | Basic (college-level, unused in 4 years) | Working fluency for model debugging | Must-Have | Everything downstream |
| PyTorch | None | Can build, train, debug a custom model | Must-Have | Portfolio project |
| Classical ML (scikit-learn, feature engineering) | None | Can justify model choice vs. deep learning | Must-Have | Interview readiness |
| ML system design (serving, monitoring, drift) | Adjacent (you already know backend infra) | Can design a production ML pipeline | Nice-to-Have (leans on existing strength) | — |
| MLOps tools (MLflow, Kubeflow) | None | Basic familiarity | Nice-to-Have | — |
Roadmap
Phase 1: Math and Python Foundations (Weeks 1-3)
Focus: Rebuild linear algebra and probability intuition specifically for ML, not abstract math. Get comfortable with NumPy and pandas at a level beyond basic scripting.
Resources: 3Blue1Brown's "Essence of Linear Algebra" series, StatQuest's probability playlist, and the NumPy/pandas sections of the official documentation worked through with real datasets, not just read.
Checkpoint: Reimplement linear regression and logistic regression from scratch using only NumPy, no libraries. If you can't derive the gradient update by hand, this phase isn't done.
Phase 2: Applied PyTorch (Weeks 4-8)
Focus: Move from math to a working deep learning toolkit. Build, train, and debug increasingly complex models.
Resources: The official PyTorch "Deep Learning with PyTorch: A 60 Minute Blitz" followed immediately by fast.ai's Practical Deep Learning course, plus Andrej Karpathy's "Neural Networks: Zero to Hero" for the parts fast.ai glosses over.
Checkpoint: Train an image classifier on a non-trivial dataset (not MNIST) to a documented accuracy, including a written note on what you tried that didn't work and why.
Phase 3: Portfolio Project (Weeks 9-13)
Focus: One end-to-end project that demonstrates the full pipeline — data collection, training, evaluation, and a simple deployment — using your existing backend skills as the deployment layer.
Resources: Your own project idea, scoped to 4 weeks; Full Stack Deep Learning's free course material for the deployment section specifically.
Checkpoint: A working, documented repository with a README explaining your modeling decisions — this becomes your interview talking piece.
Phase 4: Interview Readiness (Weeks 14-17)
Focus: ML system design questions and classical ML fundamentals, since your deep learning portfolio project won't cover model selection reasoning on its own.
Resources: "Designing Machine Learning Systems" by Chip Huyen, and mock interviews focused specifically on 'why this model over that one' questions.
Checkpoint: Can answer, unscripted, why you'd choose gradient boosting over a neural net for a given tabular dataset.
Reality Check
Four months at 8 hours a week is roughly 140 hours total. That is enough to reach a credible junior-to-mid ML engineer portfolio and pass first-round technical screens at companies that value applied skills over research depth. It is not enough to be competitive for roles requiring published research or deep theoretical specialization — if the target role specifically asks for that, extend Phase 1 by 4 weeks and push the timeline to 5 months, or accept you'll be targeting applied ML engineer roles rather than research scientist roles at this pace.
La plupart des reconversions autodirigées échouent pour une raison prévisible : le plan est une liste de lectures, pas une feuille de route. « Apprendre Python », « améliorer son ML », « étudier la conception système » – ces objectifs semblent réalisables mais ne fournissent ni séquencement, ni moyen de vérifier si une compétence est réellement acquise, ni un contrôle honnête sur le réalisme de l’échéance.
Ce prompt est construit autour de cinq décisions de conception qui font la différence entre un vague plan d’étude et un plan que vous pouvez réellement exécuter.
Pourquoi il force la spécificité sur les écarts de compétences
La première instruction interdit explicitement les noms de compétences génériques. « Apprendre Python » n’est pas un écart de compétence – cela masque si vous avez besoin de maîtrise des pandas, de programmation asynchrone ou de connaissances en packaging. En exigeant du modèle qu’il nomme les sous-compétences, les outils et la profondeur de connaissance, le résultat devient quelque chose que vous pouvez réellement cocher, et non quelque chose dont vous pouvez vous convaincre d’avoir terminé sans rien changer.
Pourquoi le classement par priorité compte plus qu’une liste plate
Les écarts de compétences ont des dépendances. Vous ne pouvez pas déboguer la boucle d’entraînement d’un modèle PyTorch si vous ne comprenez pas les gradients, et aucun tutoriel sur les frameworks ne corrige cela. Le prompt force le modèle à identifier quels écarts sont des prérequis pour d’autres – de sorte que la feuille de route qui en résulte soit ordonnée par ce qui débloque réellement la progression, et non par ce qui est le plus excitant à apprendre en premier.
Pourquoi les jalons valent mieux que le suivi d’achèvement
« Avoir terminé le cours » n’est pas une preuve de compétence. Le prompt exige un livrable concret par phase – un projet fonctionnel, une implémentation from scratch, un benchmark documenté – car ce sont les mêmes types de preuves qu’un recruteur ou un responsable du recrutement demandera réellement. Si vous ne pouvez pas produire le jalon, la phase n’est pas terminée, quelles que soient les heures passées.
Pourquoi la vérification de réalité est non négociable
La plupart des plans d’apprentissage supposent silencieusement que le calendrier annoncé tient la route. Ce prompt instruit le modèle de le dire explicitement si ce n’est pas le cas – et de proposer une alternative réaliste plutôt qu’une alternative rassurante mais fausse. Cette seule instruction empêche quelqu’un de se présenter à un entretien dans quatre mois avec encore des compétences manquantes dont personne ne l’avait averti.
Remplissez votre rôle actuel, votre rôle cible, votre budget de temps hebdomadaire et votre style d’apprentissage, et le résultat est une feuille de route par phases que vous pouvez commencer à exécuter le jour même – avec un verdict honnête sur la faisabilité de votre échéance.