Pro Workflow
Persistent correction memory, research wikis, and quality gates for coding assistants that need to improve across sessions.
Per-dimension scores and reasoning
Evidence shows permission management (permission-tuner, safe-mode, allow/deny rules), but no default least-privilege configuration or specific permission boundaries. User confirmation exists (approval-gated agents, user-approved rules), but not all external actions require confirmation. Data flow transparency: some description (SQLite storage, hooks), but no detailed data flow or external API calls. Sensitive data handling: secret-scan and secret detection, but no key storage/transmission security. Dependency security: only one runtime dependency better-sqlite3, but no vulnerability scanning or lock file. External effects: commit-validate, pre-push-check, but not comprehensive. Rollback: compact-guard and state preservation, but no data rollback mechanism. Source attribution: author and repo info, but publisher unverified.
Self-consistency: README and package.json version match (v3.4.0), skill/command counts consistent (41 skills, 23 commands), but some numbers in comparison table inconsistent (e.g., skills 34 vs 41). Dependency availability: better-sqlite3 available on npm, but version compatibility not verified. Failure messages: /doctor health check, but no detailed error handling docs.
Audience and scenarios: clearly for Claude Code users, multiple scenarios (self-correction, knowledge base, research). Capability boundaries: lists skills and commands, but no explicit limits per skill. Trigger precision: skill descriptions as triggers, but no trigger examples. Environment fit: supports multiple agents and OS, but no detailed configuration guide.
Information architecture: clear directory structure (skills, commands, hooks). Install notes: multiple install methods (marketplace, skills add, manual). Naming stability: consistent command/skill names, but version updates may change. Examples and FAQ: 60-second tour and community tips, but no FAQ. Known limitations: not explicitly listed. License: MIT, but no LICENSE file content. Versioning/changelog: version number, but no CHANGELOG. Maintenance responsibility: author info, but no maintenance commitment.
Output usability: outputs markdown, HTML, directly usable. Marginal value: self-correcting memory and knowledge base, unique value vs others. Cost-benefit: cost tracking and budget control, but no performance benchmarks.
Claim traceability: README claims lack specific evidence (e.g., test results). Cross-source corroboration: no third-party verification. Fact-inference separation: some claims (e.g., '50 sessions') not distinguished as fact or inference.
- Publisher unverified; review code before installation.
- Dependency better-sqlite3 requires build; some marketplaces may skip build step.
- Auto-research loop may make external network requests; watch budget and privacy.
- Permission rule generation may reduce security; use with caution.
What does this agent do, and when should you use it?
Pro Workflow is a Claude Code workflow plugin and cross-agent skill bundle built around a SQLite store for learning rules, sessions, and wiki data. It provides 41 skills, eight specialized agents, 23 slash commands, and 37 hook scripts spanning 24 event types. Its knowledge plane creates persistent FTS5-indexed wikis with BM25 search, optional vector retrieval, a budget-capped research loop, and multi-provider deliberation. Commands, agents, and skills are layered into workflows for development, commits, session wrap-up, context management, and parallel worktrees. Claude Code has a native plugin path, while Codex, Cursor, and other listed tools use skills add to install the skill and command bundle.
/learn-rule stores a user correction as a searchable rule, while session-start.js loads learnings and registered wikis at SessionStart. /wiki init creates a wiki, /wiki page adds a page, and /wiki ask performs BM25 retrieval; prompt-submit.js injects top wiki hits during UserPromptSubmit when a prompt matches the index. Optionally, /wiki embed builds embeddings with OPENAI_API_KEY or VOYAGE_API_KEY, and /wiki hybrid combines BM25, vectors, and RRF; /wiki research runs a bounded BFS over web, arXiv, GitHub, or custom fetchers. /wiki council performs a three-phase deliberation using Anthropic, OpenAI, OpenRouter, Fireworks, or a custom endpoint and saves the transcript as a wiki page; a single provider failure does not abort the run. /develop runs Research → Plan → Implement → Review with validation gates, while smart commits, secret scanning, destructive Git guards, and PreCompact/PostCompact hooks cover commit and compaction stages.
- A Claude Code user repeatedly correcting the same testing or code-convention mistake can capture it with
/learn-ruleand have it loaded in later sessions. - An engineer maintaining a long-running research topic can use
/wiki init,/wiki seed, and/wiki research --max-pages 5 --budget-usd 0.50to build a searchable local knowledge base. - A team implementing a complex feature can use
/developto sequence research, planning, implementation, and review behind validation gates. - A Git user with several independent changes can use
/parallel,parallel-worktrees, andbatch-orchestrationto organize concurrent worktree tasks. - A developer who needs staged review, secret checks, and a conventional commit before pushing can use
/commitor/smart-commit.
What are this agent's strengths and limitations?
- It puts learning rules, session data, and wiki indexes in one SQLite store with FTS5 retrieval across sessions.
- The knowledge plane covers manual wiki pages and queries as well as budget-bounded BFS research, optional embedding retrieval, and a shareable single-file HTML wiki viewer.
/developlinks research, planning, implementation, and review with validation gates, while commit, secret-scan, and compaction hooks address concrete development checkpoints.- It has a native Claude Code path and a documented
skills addroute for Codex, Cursor, and other listed agents.
- The native plugin and many event hooks are designed around Claude Code; other agents depend on skills add writing into their native skill directories, so feature parity should be validated locally.
- The SQLite components require a build step; the documentation notes that some marketplaces skip it, requiring an npm build when
/doctorreportsKB: missing. - Auto-research, embeddings, and council deliberation add network access, API credentials, and possible model costs; the research loop can cap a dollar budget, but provider pricing still determines the cost.
- The README uses differing counts in places, including 41 skills and 34 skills in its comparison table, so the installed inventory should be checked before standardizing on it.
How do you install or deploy this agent?
For native Claude Code installation:
/plugin marketplace add rohitg00/pro-workflow
/plugin install pro-workflow@pro-workflowFor listed tools including Cursor, Codex, Copilot CLI, and Gemini CLI:
npx skills add rohitg00/pro-workflowThen run skills sync to register the skills with the target agent configuration. For a first check, run /doctor, then /wrap-up. Manual installation requires cloning the repository, running npm install && npm run build, and copying the skills, commands, templates, and hook configuration; Node.js and npm are therefore required.
How do you use this agent?
Start with /doctor. Use /learn-rule to retain a correction and /wrap-up at the end of a session. For a research topic, run /wiki init <slug>, /wiki page <slug> <path> --type concept, and /wiki ask "<question>" --wiki <slug>. Auto-research is opt-in: run /wiki seed <slug> "<topic>", then /wiki research <slug> --max-pages 5 --budget-usd 0.50; creating ~/.pro-workflow/STOP stops every research loop. Hybrid retrieval needs OPENAI_API_KEY or VOYAGE_API_KEY, and multi-model council runs need a supported provider credential or custom endpoint credential.
How does this agent compare with similar options?
The repository compares Pro Workflow with Superpowers, everything-claude-code (ECC), gstack, and Get Shit Done (GSD). Its table positions the project around SQLite+FTS5 self-correction memory, persistent research wikis, budget-capped research, hybrid retrieval, multi-provider deliberation, and LLM-powered hooks; it also lists different skill, agent, and command counts across the projects.
FAQ
Can auto-research run indefinitely or exceed a budget?
/wiki research supports --budget-usd plus page and depth limits, and the documented loop includes convergence detection. Creating ~/.pro-workflow/STOP halts every research loop.Which API credentials are needed?
OPENAI_API_KEY or VOYAGE_API_KEY; council runs can use ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY, FIREWORKS_API_KEY, or custom endpoint credentials.What happens if a council provider fails?
llm-council uses Promise.allSettled, so one provider failure does not abort the whole deliberation.What local data does it create?
~/.pro-workflow/, including data.db, wiki directories, council transcripts, fetchers, and the stop file. Project-scoped wikis can live in <project>/.claude/wikis/<slug>/ and be committed.Can I use it directly with Codex?
npx skills add rohitg00/pro-workflow, followed by skills sync. Claude Code is the only platform with an explicitly documented native plugin installation path.