# Controlled Parallel Work

> Complete software work with the explorer and forge custom agents under tight scope. Use only when user explicitly requests subagents, parallel work, exploration, or delegated implementation. Avoid small or sequential tasks.

- Skill: `davidbasilefilho/controlled-parallel-work` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add davidbasilefilho/controlled-parallel-work`
- Raw SKILL.md: https://api.skillmd.com/api/skills/davidbasilefilho/controlled-parallel-work/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: davidbasilefilho (https://skillmd.com/u/davidbasilefilho)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/davidbasilefilho/controlled-parallel-work

---


# Controlled Parallel Work

Use only configured `explorer` and `forge` agents.

## Rules

- Use at most one `explorer` and one `forge`; one or zero may suffice.
- Forbid nested delegation.
- Give narrow, non-overlapping scopes. Never duplicate work.
- `explorer`: read-only evidence, paths, symbols, constraints, risks.
- `forge`: implement one approved plan with smallest complete patch. Never use without approved plan.

## Workflow

1. Read plans in `.agents/plans/`. If none exist, stop. If one exists, use it. If multiple exist, ask user which plan to use and stop.
2. Read selected approved plan, applicable `AGENTS.md`, and repo instructions.
3. Spawn `explorer` only when focused investigation helps. Verify returned evidence.
4. Give `forge` selected plan, exact scope, constraints, and checks.
5. Wait for each required result. `forge` owns implementation; parent owns synthesis and final verification.
6. Inspect patch. Run narrow checks. Fix only caused failures.
7. Never commit, push, open PRs, or deploy unless requested.

Stop after implementation and targeted checks. Report agent scopes, changes, validation, unresolved risks. Skip optional work.