klaude
Silverfell/klaude
Klaude is a lightweight discipline layer for Claude Code sessions that keeps project context current, enforces explicit decision logs, and reduces agent drift.
Instalação
npx convez-ai Silverfell/klaude1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Klawde
Klawde is a lightweight harness for Claude Code that makes sessions more reliable.
It does four things:
- forces an explicit session start
- keeps a living project brief up to date
- leaves behind an append-only decision log for the next session
- compresses older change history when the journal gets too large
The goal is simple: less agent drift, less forgotten context, less vague "the model probably remembers" nonsense.
What problem this solves
Coding agents are useful, but they are sloppy by default.
They lose context across sessions, invent assumptions, skip verification, and leave behind work with no durable record of what changed or why. Klawde adds a minimal operating discipline around that workflow.
It is not a framework. It is not an IDE plugin. It is not a memory layer.
It is a small harness that gives Claude Code a stricter contract for how a session begins, how decisions are recorded, how a session closes, and how older history stays usable instead of bloating into noise.
Who this is for
Klawde is for people who already use Claude Code and want:
- a consistent session protocol
- a living project brief
- an append-only project journal
- stricter rules around honesty, verification, and scope control
- a cheap way to reduce context drift between sessions
It is especially useful on projects where architecture, scope, and decision history matter more than just the final diff.
What's included
| File | Purpose |
|---|---|
template/CLAUDE.md | Project rules, coding discipline, verification rules, and output contract |
template/klawde.md | /klawde command: reads or creates session documents at session start |
template/close.md | /close command: updates session records before ending work |
template/compresschanges.md | /compresschanges command: compacts older CHANGES.md history into summaries |
setup.sh | First-time setup: copies the harness files into your project |
upgrade.sh | Upgrade: overwrites harness files with the latest defaults, converts legacy CHANGES.md entries, and retires the legacy /init command |
How it works
/klawde
Run this at the start of every Claude Code session. (The command is named /klawde rather than /init to avoid colliding with Claude Code's built-in /init.)
It makes Claude Code:
- check whether
BRIEFING.mdandCHANGES.mdexist, and create them if missing - read
BRIEFING.mdcompletely - ask for the project's purpose and current scope if the brief is empty
- read at least the last 30 lines of
CHANGES.md - check that recent changes do not contradict the brief
- confirm readiness, then stop
This forces the session to begin from actual project state instead of hand-waving.
/close
Run this at the end of a session.
It makes Claude Code:
- review the current
CHANGES.mdandBRIEFING.md - review what changed in the session
- append typed journal entries for any unrecorded decisions, plans, docs, scope shifts, notes, or code that needs project-level explanation
- update
BRIEFING.mdif purpose, scope, key decisions, non-goals, or breaking-change context changed, or if its fields are still empty - confirm what was recorded, then stop
This gives the next session a real handoff.
/compresschanges
Run this when CHANGES.md has grown too large.
It makes Claude Code:
- read
CHANGES.mdcompletely - identify entries older than 30 days
- group those older entries by month
- replace older low-level history with summary lines by major theme or area of work
- preserve decision entries, scope changes, and breaking-change context verbatim
- leave the last 30 days untouched
- keep the file header and format hint intact
- report how many lines were compressed and preserved, then stop
This keeps the journal useful instead of letting it become an unreadable dump.
The two core documents
BRIEFING.md
A concise living brief for the project.
Use it for:
- purpose
- current scope
- key decisions
- non-goals
- breaking-change context
A new session should be able to read this and understand what the project is trying to do.
CHANGES.md
An append-only journal for project-level context.
This is not a git log. It is where you record things git commits usually fail to preserve cleanly:
- architectural decisions
- scope changes
- revised plans
- documentation shifts
- handoff notes
- code changes that need project-level explanation
Format:
YYYY-MM-DD [type] description
For example: 2026-05-12 [decision] Switched queue from Redis to Postgres SKIP LOCKED; one less service to operate
Types:
decisionplandocscopecodenote
Installation
Clone this repo somewhere local:
git clone https://github.com/Silverfell/klaude.git
Move into the project where you want to install the harness:
cd /path/to/your/project
Run the setup script from the Klawde repo:
/path/to/klaude/setup.sh
That will copy:
CLAUDE.md.claude/commands/klawde.md.claude/commands/close.md.claude/commands/compresschanges.md
Then start Claude Code and run:
/klawde
Upgrading
To overwrite an existing installation with the latest defaults:
/path/to/klaude/upgrade.sh
The upgrade script also:
- overwrites the harness files with the latest versions
- removes the legacy
.claude/commands/init.md(the entry protocol is now/klawde) - converts legacy
CHANGES.mdlines fromYYYY-MM-DD: descriptiontoYYYY-MM-DD [note] description - adds the format hint to
CHANGES.mdif it is missing - backs up every file it overwrites or removes first
Example workflow
Start session:
/klawde
Claude reads the briefing and recent changes, then confirms the session is ready.
Do the work.
End session:
/close
Claude records decision and scope changes, updates the briefing if needed, and leaves the project in a state that the next session can pick up cleanly.
If the journal gets too large:
/compresschanges
Claude compacts older history into monthly summaries while preserving high-value records verbatim.
Design principles
Klawde is opinionated.
It assumes that better agent output comes from stricter process, not from pretending the model is magically dependable.
Core principles:
- read real files before making claims
- say "I don't know" when you do not know
- verify before reporting completion
- keep diffs minimal
- avoid speculative improvements unless asked
- preserve durable project context outside the model
When not to use this
Do not use Klawde if:
- you want a fully autonomous agent stack
- you do not care about session continuity
- your projects are too trivial to justify written context
- you want hidden magic instead of explicit process