Quoin

Workflow state for stateless coding agents.

A workflow memory toolkit for Claude Code that structures your coding agent workflow using lightweight persistence files — so every session starts informed and every decision is tracked.

What is Quoin?

Quoin gives Claude Code a disciplined engineering partner: planning, architecture review, quality gates, cost discipline, and institutional memory — across sessions, tasks, and team members.

Without Quoin, each Claude session starts cold. With Quoin, sessions share accumulated knowledge through structured artifacts: plans, critic reviews, session state, lessons learned, and a cost ledger.

Why Quoin?

• Cost discipline — §0 model dispatch preamble routes each skill to the right tier (Haiku/Sonnet/Opus). /cost_snapshot shows live spend. ccusage fallback tracks every session.
• Planning rigor — /thorough_plan runs a plan→critic→revise convergence loop before a single line of code is written.
• Audit trail — every phase produces a structured artifact: architecture.md, current-plan.md, critic-response-N.md, review-N.md, cost-ledger.md.
• Six stages of foundation — stages 1–6 of the Quoin foundation hardened the system: §0 dispatch preamble (1), ccusage fallback (2), stage-subfolder path resolution (3), architect Phase 4 critic loop (4), native Haiku summarizer (5), and this rebrand + QUICKSTART relocation (6).

Install

git clone https://github.com/FourthWiz/quoin
cd quoin
bash quoin/install.sh

Note: GitHub auto-redirects from the old FourthWiz/claude_dev_workflow URL — existing clones continue to work.

30-Second Start

/init_workflow   ← one-time project bootstrap
/architect       ← design the solution
/thorough_plan   ← converge on a plan with critic review
/implement       ← write the code (explicit — you decide when)
/review          ← verify implementation against the plan
/end_of_task     ← push the branch (explicit — you decide when to ship)

Skills

Planning & Architecture

Command	Model	What it does
/architect	Opus	Deep architectural analysis; internal Phase 4 critic loop
/plan	Opus	Detailed implementation plan (single-pass)
/thorough_plan	Opus	Triages task size; runs plan→critic→revise convergence loop
/critic	Opus	Reviews a plan for gaps, risks, integration issues
/revise	Opus	Revises plan from critic feedback (strict/Large mode)
/revise-fast	Sonnet	Revises plan from critic feedback (Medium mode, cost-efficient)

Implementation & Review

Command	Model	What it does
/implement	Sonnet	Writes code from the plan (explicit command only)
/review	Opus	Verifies implementation against the plan; production-ready check
/gate	Sonnet	Automated quality checkpoint between phases; requires your approval
/rollback	Sonnet	Safely undoes an implementation phase or specific tasks

Session Lifecycle

Command	Model	What it does
/init_workflow	Opus	One-time project bootstrap — creates .workflow_artifacts/, runs /discover
/discover	Opus	Scans all repos; maps architecture, dependencies, git log
/start_of_day	Haiku	Morning briefing — restores context from daily cache
/end_of_day	Haiku	Saves session state; promotes insights to daily cache
/end_of_task	Sonnet	Pushes branch, captures lessons, marks task complete (explicit only)

Utilities

Command	Model	What it does
/run	Opus	End-to-end pipeline orchestrator; pauses at each gate
/cost_snapshot	Haiku	Live cost: today, lifetime, per-task breakdown
/triage	Haiku	Routes your request to the right skill
/weekly_review	Haiku	Aggregates the week's progress into a structured review
/capture_insight	Haiku	Logs a pattern or gotcha to the daily scratchpad
/expand <path>	Sonnet	Re-renders a terse workflow artifact in readable English

Architecture

Each Claude session is stateless by nature. Quoin bridges sessions using structured file artifacts under .workflow_artifacts/. Skills read these files at startup (session bootstrap) and write them on completion — so the next session, whether in 5 minutes or 5 days, picks up exactly where you left off.

For full rules, model assignments, artifact formats, and cost-tracking conventions, see quoin/CLAUDE.md.

Cost & Model Discipline

Every Quoin skill declares a model tier (haiku, sonnet, opus). When invoked from a session running on a more expensive model, the skill self-dispatches via the Agent tool to its declared tier (§0 model dispatch preamble).

• Fails gracefully: if dispatch is unavailable, the skill proceeds at the current tier with a one-line warning — never aborts your invocation.
• Tracked: every session appends a row to cost-ledger.md. /cost_snapshot reads it live. ccusage is the fallback if the ledger is unavailable.
• Three artifact tiers: Tier 1 (always English — CLAUDE.md, reviews, architectures), Tier 2 (English + side-file), Tier 3 (terse; use /expand to read).

Typical Flows

Large feature (full pipeline):

/discover → /architect → /thorough_plan large: ... → /implement → /review → /end_of_task

Medium task (default):

/thorough_plan add caching layer → /implement → /review → /end_of_task

Small bug fix:

/thorough_plan small: fix null pointer in auth handler → /implement → /review → /end_of_task

Start / end of day:

/start_of_day          ← morning: restores context, checks git state
/end_of_day            ← evening: saves state, consolidates insights

Documentation

• quoin/QUICKSTART.md — command reference table (21 skills)
• <your-quoin-clone>/Workflow-User-Guide.html — interactive walkthrough with scenarios (open in browser)
• quoin/CLAUDE.md — full workflow rules, model assignments, artifact formats

Contributing

Bug reports and PRs welcome. The workflow is its own guinea pig — all Quoin development uses Quoin.

License

MIT