Give Claude Code a Memory That Compounds With pro-workflow

You correct Claude the same way a dozen times, re-explain your conventions every session, and re-research the same topic because nowhere durable exists for the answer to land. pro-workflow puts a single SQLite store underneath every session: corrections become searchable rules, research becomes persistent wikis, and 33 hook scripts across 22 events keep quality gates on. The README's inventory: 41 skills, 8 agents, and 23 slash commands.

Why This Skill Matters

Three layers share that store. Self-correction memory turns every correction you approve into a rule that loads on session start. The knowledge plane adds FTS5-indexed research wikis that can grow themselves through a budget-capped auto-research loop. Quality gates — LLM-powered hooks, deterministic git and secret guards, cost tracking — run around the edges. UserPromptSubmit auto-injects the top-3 wiki hits whenever your prompt mentions an indexed topic, so the wiki you built last week answers this week's question without being asked.

Installation

On Claude Code, use the native plugin marketplace:

/plugin marketplace add rohitg00/pro-workflow
/plugin install pro-workflow@pro-workflow

For Cursor, Codex, Gemini CLI, OpenCode, and other agents — 32+ in total via skills add — install the same bundle:

npx skills add rohitg00/pro-workflow

The README notes two things: use the owner/repo form (skills add resolves providers from it, not from marketplace slugs), and run skills sync after installing to register the skills with the target agent's config.

Real Workflow: Turn a Correction Into a Rule, Then Grow a Wiki

1. Smoke-test the install:

/doctor

If /doctor reports KB: missing, the README documents the fix: cd ~/.claude/plugins/*/pro-workflow && npm install && npm run build — some marketplaces skip the SQLite build step. 2. Capture the correction you just made:

/learn-rule

It becomes a rule, FTS5-searchable, auto-loaded on every future SessionStart. 3. Give research a durable home. The wiki command supports 9 flavors — research, paper, domain, product, person, organization, project, codebase, and incident:

/wiki init agent-memory --title "Agent Memory" --flavor research
/wiki seed agent-memory "memory consolidation in agents"
/wiki research agent-memory --max-pages 5 --budget-usd 0.50

The research loop is budget-capped and opt-in — it runs only when you seed it. 4. End sessions with intention:

/wrap-up

It audits changes, persists learnings, and writes a handoff document for the next session. 5. Any auto loop obeys the kill switch: touch ~/.pro-workflow/STOP.

Tips

• Everything lives under ~/.pro-workflow/: data.db for learnings and wiki pages, wikis/<slug>/ for wiki content, plus council transcripts and fetcher scripts. Project-scope wikis live in the repo at .claude/wikis/<slug>/ and are committable.
• Before a pull request, /smart-commit runs quality gates, staged review, and a conventional commit message.
• Stuck on a hard bug? /develop walks Research → Plan → Implement with validation gates between phases.
• Hybrid retrieval (BM25 + vector) and the multi-LLM council are optional and need API keys.
• The repo is MIT-licensed, and the manual clone-and-copy path in the README works on any OS.

When Not to Use This

This is a harness, not a toggle: 33 hook scripts across 22 events reshape how sessions start, stop, and submit prompts. If your workflow is a handful of conventions, a plain CLAUDE.md is lighter and has nothing to maintain. The payoff the README describes is compounding across dozens of sessions — if you code in one-off environments that reset weekly, the store never gets the repetitions it learns from.

See the leaderboard for more skills.