About ameliore-systeme
ameliore-systeme is a Claude Code skill that reviews local session history and turns recurring errors, repeated instructions, workflows, and tasks into specific improvement proposals. It is intended for users who want to improve Claude Code rules, create skills, or automate recurring work based on evidence from their conversations.
Key features
- 01Digest session history for a chosen period
- 02Identify recurring mistakes and instructions
- 03Propose CLAUDE.md rules and skills
- 04Suggest automation for repeated tasks
- 05Keep history local and read-only
Use cases
- 01Find Claude Code mistakes repeated across sessions
- 02Turn recurring user instructions into project rules
- 03Design a skill from repeated multi-step workflows
- 04Identify repeated tasks suitable for automation
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ melvynx/ shared-skills --skill ameliore-systemeAfter installing, type/ameliore-systemein Claude Code to confirm it loaded.
Bundled files
References and scripts installed along with SKILL.md.
README.mdagents/openai.yamlassets/codex-icon.svgscripts/digest.mjsDetails
From the project READMEameliore-systeme
Every Claude Code conversation is saved on the user's computer, one .jsonl file per session in
~/.claude/projects/<folder>/. That history shows what Claude got wrong, what the user had to say
twice and the work they redo every day. This skill reads it and turns it into fixes.
What the user typed after /ameliore-systeme: $ARGUMENTS (a number = how many days to read, default 7).
Answer in the language the user writes in (French by default).
1. Read the history
Run the digest script of this skill folder. It only reads, and writes one Markdown file:
node <this skill folder>/scripts/digest.mjs --days <N> --out /tmp/ameliore-digest.md
Then read /tmp/ameliore-digest.md in full. Each session has its folder, its date, its title, the
slash commands used, the tool errors and every message the user typed. Never open the raw .jsonl
files one by one: they hold the whole conversation and fill the context. Open one only to check a
precise point, with grep.
If the digest is cut (tronqué), run it again with fewer days.
2. The report
Write the report in chat, with these 4 sections. Every point quotes the history: a date, a folder and the user's own words. Never a generic tip that the history does not show.
- Ce que Claude fait mal: the mistakes that come back (the user says « non », « pas ça »,
« je t'ai déjà dit », redoes the same request, stops Claude). For each one: how many times, 2 quotes,
and the fix: one rule to add to
CLAUDE.md(written out, ready to paste). - Ce que tu répètes: instructions the user gives in many sessions (a style, a tool, a check).
Each one becomes a
CLAUDE.mdrule or a skill. - Les skills à créer: workflows of several steps done again and again. For each one: its name, what it would do in 3 lines, the sessions it comes from.
- Les tâches répétées à automatiser: the tasks that come back every day or every week (same folder, same kind of request). For each one: the dates and folders where it appears, the time it takes today (number of messages and tools in those sessions) and the automation that would do it: a skill, a script, a scheduled task or a hook.
End with the top 3 changes, ranked by time saved, and ask which one to build.
3. Build what the user picks
- A
CLAUDE.mdrule: show the exact lines, then add them to~/.claude/CLAUDE.md(every project) or the project'sCLAUDE.md(one project) on the user's yes. Never remove a line already there. - A skill: write
~/.claude/skills/<name>/SKILL.md(frontmatternameanddescriptionwith the phrases that should trigger it, then the steps taken from the sessions). Show it before saving. Tell the user to type/<name>to use it. - An automation: write the script or the scheduled task, run it once, show the result.
Never delete or edit a session file. Never send the history anywhere: everything stays on the computer.
Notes
- Run it once or twice a week: each report only reads the days asked.
- The digest keeps only what the user typed, not Claude's answers, so it stays small. Tool errors are counted per session, which shows where Claude struggled.