improve-system
/improve-system
Use when the user says "improve system", "/improve-system", "what do you keep getting wrong", "review my week", "what should be a skill", or a twice-weekly scheduled run fires. Reads the local Claude Code transcripts across every chat for a window of days and reports repeat mistakes, repeat tasks worth a skill, and rules that would have prevented them.
- Author
- kennykennethkkang ↗
- Repository
- switchboard ↗
- Licence
- Apache-2.0
- Last repository change
- 10.11.2026
Assessment and security
- Files in this package
- 3
- Scripts
- 2
- HTML comments
- 1
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 10.11.2026, without references to retired models or APIs.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
- 2 scripts in this package. Review them before installing, as Claude may execute them.
- SKILL.md zawiera komentarze HTML (1). Nie widać ich w podglądzie Markdown, a model je czyta.
Scripts: scripts/scan.mjs, scripts/scan.test.mjs
What it does and when to use it
Use when the user says "improve system", "/improve-system", "what do you keep getting wrong", "review my week", "what should be a skill", or a twice-weekly scheduled run fires. Reads the local Claude Code transcripts across every chat for a window of days and reports repeat mistakes, repeat tasks worth a skill, and rules that would have prevented them.
Triggers from the description
Original author description
Use when the user says "improve system", "/improve-system", "what do you keep getting wrong", "review my week", "what should be a skill", or a twice-weekly scheduled run fires. Reads the local Claude Code transcripts across every chat for a window of days and reports repeat mistakes, repeat tasks worth a skill, and rules that would have prevented them.
How it works
- Arguments
- Steps
- What counts
- Report
- Drafting a skill
- Running it twice a week
SKILL.md preview
--- name: improve-system description: Use when the user says "improve system", "/improve-system", "what do you keep getting wrong", "review my week", "what should be a skill", or a twice-weekly scheduled run fires. Reads the local Claude Code transcripts across every chat for a window of days and reports repeat mistakes, repeat tasks worth a skill, and rules that would have prevented them. --- # Improve system Looks across many chats for patterns: what Claude got wrong more than once, what the user keeps asking for by hand, and what context would have stopped it. `aar` reviews one chat; this one reviews the week. It only reads and proposes. It never edits CLAUDE.md, rules, skills or settings. Fully local: transcripts stay on this machine, no network calls, nothing sent anywhere. Under `~/.claude` it reads only `projects/*/*.jsonl`, memory files and `CLAUDE.md`; never `settings*.json`, `.credentials*` or `.claude.json`. ## Arguments - Days: default 7. "last 3 days" means `--days 3`. - Project: optional, a part of the project folder name, `--project switchboard`. ## Steps 1. Scan. Never read whole transcripts into context; the script does the reading. ``` node <this skill>/scripts/scan.mjs --days 7 [--project name] ``` It prints one JSON object: `projects` (sessions, typed turns, tool errors, interrupts), `retries` (a failed tool call run again unchanged), `corrections` (typed turns that read like a correction or frustration), `errors` (tool errors grouped, with counts and projects), `repeated` (4 to 8 word phrases typed in two or more sessions) and `openers` (the first ask of each session). Secrets are redacted before printing, best effort: private keys, URL passwords, `sk-`/`rk-`/`pk-` and Stripe `sk_live_`/`rk_test_` keys, AWS `AKIA`, GitHub `gh*_`, Slack `xox*-`, Google `AIza`, JWTs, `Bearer` values, `key`/`token`/`secret`/`password` assignments (plain or JSON-quoted), hex runs of 32+ and mixed runs of 40+ outside file paths. Anything else can slip through; treat the output as private. 2. Check what is already known, so the report does not repeat it: - `hub_recall` for the main topics found, if that tool is available. - If a vault exists at `~/Documents/vault`: its `wiki/aar-*.md` and `wiki/improve-system.md` (earlier runs). - The project memory folders' `MEMORY.md` and `~/.claude/CLAUDE.md`. - The open Switchboard to-do, if Switchboard is installed, so the same fix is not filed twice. 3. When a hit needs context, grep the one transcript for it (`grep -n "phrase" ~/.claude/projects/<dir>/<id>.jsonl | head`) and read a few lines around it. Do not open whole files. 4. Write the report (below). 5. File each proposed rule or skill change with `mcp__switchboard__todo_add` (without Switchboard, list them at the end of the report): the exact target file and the exact text to add or replace. Point at the source a rule or skill is built from, not an installed copy: if a `claude-mods` clone exists, that is `kit/rules.md`, `kit/skills/<name>/SKILL.md` or `kit/agents/<role>.md` there, not the copies in `~/.claude` or `switchboard/extension/kit`. Never apply them. 6. If a vault exists at `~/Documents/vault`, append the dated report to its `wiki/improve-system.md` following its `README.md` (add it to `wiki/index.md` if the page is new, a dated line in `wiki/log.md` either way). Otherwise the report in the reply is the record. One-line facts that are already settled go to memory the way `aar` does it. ## What counts - A mistake is a pattern when it shows up in two or more sessions, or three times in one. One-offs go nowhere. - A correction keyword alone is not a mistake. "again" can mean "run it again". Read the text before counting it. - A repeat task is the same ask in two or more sessions that takes the same steps each time: "ask all questions first", "add the commit message", "watch this reel with gemini". Phrases that are just how the user talks are not tasks. - A rule proposal must name the file, the text and the mistake it would have stopped. Same bar as `aar`: no "be more careful". ## Report ``` Improve system: <since> to <today>, <n> sessions, <projects> Repeat mistakes - <what went wrong>: <n> times in <projects>. Why: <mechanism>. Fix: <to-do id or "known, see <memory/page>"> Make it a skill or automation - <task>: asked in <n> sessions. Draft: <skill name, or hook/schedule/alias> Rules that would have helped - <file>: "<text>" stops <mistake>. Filed: <to-do id> Nothing new: <areas checked that came up clean> ``` Keep it under a screen. Zero findings in a section is a fine answer. ## Drafting a skill Only when the user asks for it, or picks one from the report: draft a `SKILL.md` for the top repeat task in the reply (or under the scratchpad), with `name`, a `description` that starts "Use when" and names the user's own trigger phrases from `repeated`, and the steps as they were done in the sessions. It is saved (to `claude-mods/kit/skills/<name>/` if that clone exists, else `~/.claude/skills/<name>/`) only after the user approves it. Follow `superpowers:writing-skills` if installed. ## Running it twice a week Switchboard does it: with Settings, Learn automatically on (the default), one window runs `claude -p` naming this skill in the switchboard folder twice a week, when usage is under the pause threshold. That run passes each proposal to `self-improve`; Switchboard applies clear wins and files the rest. Without Switchboard, a cron or launchd job running `claude -p "/improve-system"` works; a cloud `/schedule` routine cannot read the local transcripts. <!-- switchboard generic copy -->
Files in this package (3)
- SKILL.md 6 KB
- scripts/scan.mjs 9 KB
- scripts/scan.test.mjs 7 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · kennykennethkkang/switchboard | 100 | Apache-2.0 |
| lucas-saldanha-werneck/self-improve | 100 | MIT |
| HarryBurd33/codex-improve-system | 100 | MIT |
Database snapshot: 13.11.2026 · package version 5ee13a9f443d (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100chat-review
Chat review ("Bigg Boss") - audit past Claude Code sessions on a topic for mistakes the user corrected, prove which repeat across sessions, propose fixes to skills, rules, tools or agents, and apply only the approved ones. Use when the user says "Bigg Boss", asks to review or… - 87skill-harvest
Load when the user wants to find, create, improve or publish agent skills from the work already done across ALL their Claude Code sessions, or wants a chat that "watches every chat" for skills. It reads every past session on the machine in minutes, finds the methods that repeat… - 87skill-harvest
Load when the user wants to find, create, improve or publish agent skills from the work already done across ALL their Claude Code sessions, or wants a chat that "watches every chat" for skills. It reads every past session on the machine in minutes, finds the methods that repeat… - 100repeat-tax
Use when the user wants to lower Claude Code or Codex costs in a repo, asks what agents keep getting wrong here, or wants facts for CLAUDE.md or AGENTS.md learned from past sessions. - 100tune
Harness retro — turn recurring agent mistakes into one-line fixes. Mines recent session transcripts (plus the bugfix log and past runs) for repeated corrections, permission friction, and rules that get violated anyway, then proposes ranked changes one at a time: a CLAUDE.md… - 100alice-learn-from-outcomes
Alice's weekly dream — read what became of every PR she opened (merged ones for the steering and pushback in their reviews, closed and reverted ones for her mistakes), her raw session transcripts for where she was blocked, stalled or repeated an error, and her memory, then land…