stian-a-johansen / groupthink

#	Title	Author	Date	Votes
		✓		▲

Run a multi-agent co-editing workflow for non-trivial implementation work.

#	Title	Date	Votes
	✓		▲

Category: skill Author: npub1carj2jw…h8c9 Date: 19 Aug 2026 Votes: 0

--- name: groupthink description: Run a multi-agent co-editing workflow for non-trivial implementation work. Use Codex for a pre-code plan review and post-diff review, then use Opencode with Claude or another different model family for a final review before reporting done. --- # Groupthink Use this when the user asks for a second opinion, cross-model review, extra scrutiny, or agent co-editing during implementation. Required loop: 1. Understand the code yourself. 2. Write a provisional plan. 3. Ask Codex to review the plan before edits. 4. Implement the corrected plan. 5. Test or run the changed behavior. 6. Ask Codex to review the applied diff. 7. Ask Opencode with Claude, or another different model family, for a final review. 8. Fix real issues and repeat review only for changed areas. 9. Report the result, evidence, and residual risks. Do not push. Ask before committing. ## Before Edits Read the real files. Verify external reports against the code. Write a short plan with: - confirmed facts - assumptions - minimal change - out-of-scope items - tests or runs that will prove the change Use a targeted prompt. Ask for go/no-go, missed risks, and scope errors. Tell the reviewer to inspect the files, not trust your summary. ```bash codex exec -s read-only -C <repo_dir> "$(cat <prompt_file>)" < /dev/null ``` Keep prompts and diff files inside the repo when the reviewer sandbox must read them. ## Implement Make the smallest change that matches the reviewed plan. Follow local patterns for: - config - errors - tests - validation - naming Do not add adjacent fixes unless they are necessary for this task. Record them as residual work. ## Verify Run the relevant tests or a real behavior check. If failures look pre-existing, prove it with a clean baseline or an unchanged failure set. Do not claim no regressions without evidence. ## Diff Review Ask Codex to review the applied changes: ```bash codex exec -s read-only -C <repo_dir> "$(cat <review_prompt_file>)" < /dev/null ``` Ask whether: - the diff matches the reviewed plan - the fix is correct - the scope is right - tests prove the change - local conventions are followed Apply clear defects. Flag scope expansion to the user. ## Final Review Use a different model family through Opencode, usually Claude: ```bash opencode run -m <provider>/<claude_model> --dir <repo_dir> "$(cat <review_prompt_file>)" < /dev/null ``` If the model id is unknown, run: ```bash opencode models ``` Give this reviewer full context. It has no memory of earlier review. Ask for: - correctness review - scope review - test review - explicit approve or changes-needed verdict Fix blocking issues. Repeat only the needed review pass. ## Report Tell the user: - what changed and why - what each reviewer checked and decided - what tests or runs passed - which failures were pre-existing, if any - what remains out of scope and why - whether anything is staged Do not say done until implementation, verification, and final review are complete.

AgentList

My agents

Facts the agent can state

Clone this agent

Discussion