Contenu du pack

Un fichier SKILL.md (le guide pratique) + un modèle de départ. Aperçu :

Terminal AI Assistant Skills

When to use

• You live in the terminal and want AI help without leaving it — reading files, editing code, running commands, and explaining output in place
• Your AI sessions drift: the assistant forgets the goal, rewrites files you never asked about, or runs commands you would not have approved
• You want repeatable prompt patterns for everyday terminal work instead of improvising a new prompt every time
• You need a session you can hand to someone else — or to your future self — and have them pick up exactly where you left off
• You work across several projects and want the same session discipline in each one, not a different habit per repo

Workflow

1. Set up the session (input: project directory + task → output: session brief)
   Open the terminal in the project directory and state the task in one sentence, the files in scope, and the files that are off-limits. Write that down as the session brief before the first prompt — it is the contract the rest of the session answers to. A session without a brief becomes a ramble; a session with one has a finish line.
2. Teach it the repo (input: session brief → output: context map)
   Have the assistant build a short context map: what the project is, how it is laid out, where the entry points are, and which files matter for this task. Ask it to state what it does *not* know instead of guessing. If the map is wrong, fix the map before any work starts — every later step stands on it.
3. Shape prompts as commands (input: vague request → output: command-shaped prompt)
   Rewrite each request the way you would write a command: verb first, target named, constraints listed, expected output stated. "Fix the login bug" becomes "Diagnose the login failure in auth.py, list the two most likely causes with evidence, change nothing yet." Command-shaped prompts get command-shaped answers: scoped, checkable, and safe to act on.
4. Run file workflows in small loops (input: command-shaped prompt → output: verified change)
   Work in the loop: read → change one thing → run the check → show the result. One file per loop when you can manage it, one behaviour per change, and a runnable check after each change — a test, a build, a lint pass, or simply running the program. If a loop produces no checkable result, the loop is not done; do not stack a second change on top of an unverified first one.
5. Review diffs before they land (input: proposed change → output: accepted diff)
   Before anything is treated as finished, look at the actual diff: every line added, changed, or deleted, and the reason for each. Reject changes outside the session brief, even good-looking ones — scope creep in a diff is still scope creep. You approve the diff; the assistant proposes it. That separation is the whole safety model.
6. Close the session with a log (input: session history → output: session log)
   End by writing a short log: what was attempted, what changed, what was verified, what is still open, and the exact command to resume. The log is what turns a one-off session into a repeatable skill — next time, the session starts from the log, not from zero.

Aperçu du modèle

Terminal AI Assistant — Session Template

One session = one brief, one log. Fill this in before the first prompt.

Session brief

Task (one sentence): {what you want done}

Project directory: {path}

In scope: {files / directories the assistant may read and change}

Off-limits: {files / directories it must not touch}

Definition of done: {what "finished" looks like, in checkable terms}

Context map

What this project is: {one paragraph}

Layout: {key directories and what lives in each}

Entry points: {where the program / site / pipeline starts}

Files relevant to this task: {list}

Le pack de compétence lui-même (SKILL.md + modèle) est rédigé en anglais.

Acheter le pack complet — $4.99 Pro — les 511 compétences