# Compound Development Loop

> Agentic development operating loop for non-engineer-led builds: expand the idea, research surrounding advances, produce artifacts, plan, build, review, and compound learnings back into the 2nd brain and CRM.

- Skill: `cleanexpo/compound-development-loop` (Agent Skill)
- Install (CLI): `npx skillmds@latest add cleanexpo/compound-development-loop`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cleanexpo/compound-development-loop/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- Author: cleanexpo (https://skillmd.com/u/cleanexpo)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/cleanexpo/compound-development-loop

---


# Compound Development Loop

Use this skill whenever the operator presents an idea, asks what to build next,
requests a feature, or asks the system to improve a product. The operator is
not expected to provide an engineer-complete brief. The system must expand the
idea, research the surrounding opportunity, and produce the missing structure
before implementation.

## Core Rule

Do not treat the operator's words as the full specification.

Treat the words as a signal. Expand the signal into a better operating packet:
adjacent opportunities, research lanes, risks, implementation candidates,
approval tasks, and durable 2nd-brain context.

## Loop

1. **Strategy Anchor**
   - Read the relevant project charter, strategy doc, 2nd-brain notes, CRM
     context, and current repo state when available.
   - If no anchor exists, state the missing anchor as a gap instead of
     inventing one silently.

2. **Idea Expansion**
   - Run or mirror `scripts/idea_expansion.py` for raw operator ideas.
   - Produce a direct ask, strategic read, adjacent opportunities, second-order
     effects, research lanes, implementation candidates, risks, and CRM-ready
     approval tasks.
   - Write outputs to the active 2nd brain when the workflow has file access.

3. **Research Grounding**
   - Search the local repo and 2nd brain first.
   - For current external context, research GitHub repos, Hugging Face models or
     Spaces, papers, product examples, docs, and security/licensing concerns.
   - Do not install, vendor-onboard, or production-wire external tools from
     research alone.

4. **Requirements / Decision Brief**
   - Define WHAT should exist before deciding HOW to build.
   - Name the user/problem, scope boundaries, success criteria, non-goals,
     approval decisions, and open questions.
   - Use `forward-planner` for large, ambiguous, or system-level work.

5. **Implementation Plan**
   - Use `technical-plan` for software execution.
   - The plan must include affected files, sequencing, risks, test scenarios,
     rollback/operability notes, and approval gates.
   - If the plan touches restricted files or external integrations, stop and
     request human approval instead of auto-shipping.

6. **Build**
   - Implement in small verifiable units.
   - Preserve existing project patterns and AGENTS boundaries.
   - Do not overbuild hidden adjacent ideas; adjacent ideas become research or
     approval tasks unless explicitly approved for the current build.

7. **Multi-Agent Review**
   - Run specialist review before merge for meaningful changes.
   - At minimum cover correctness, tests, scope, security/data boundaries,
     integration behavior, and operational readiness.
   - Synthesize through a Senior PM ledger: task, implemented status, evidence,
     concrete gap, owner, and next action.

8. **Compound Learning**
   - After completion, capture what was learned while the context is fresh.
   - Write a 2nd-brain note or solution record for repeated patterns, failures,
     or decisions that future agents should reuse.
   - Promote to a SKILL.md or AGENTS guidance only when the pattern is repeated,
     bounded, testable, and safe.

9. **Product Pulse**
   - Feed CRM approvals, user outcomes, errors, deployment signals, and operator
     decisions back into the next idea expansion and planning pass.

## Operating Boundaries

- The loop increases ceremony intentionally for ambiguous and strategic work.
  For trivial fixes, use a compact version: scope, implement, test, review,
  record any learning.
- Expansion is not permission to build every adjacent idea.
- CRM tasks generated by the loop are blocked/approval-required by default.
- External capabilities stay sandbox-only until license, security,
  maintenance, and product fit are reviewed.
- The 2nd brain is the knowledge source and evidence trail; CRM is the
  operational approval surface.

## Required Output Shape

For any non-trivial invocation, produce or reference:

- `strategic_read`
- `research_lanes`
- `decision_or_requirements_brief`
- `implementation_plan`
- `verification_plan`
- `review_ledger`
- `compound_learning_target`

If any item is missing, name it as a gap. Do not claim the loop is complete
until those artifacts exist or are explicitly waived by the operator.

## Native Pi-CEO Integration Points

- Idea expansion: `scripts/idea_expansion.py`
- Phone/Telegram idea intake: `scripts/process_ideas_inbox.py`
- Capability research loop: `scripts/run_capability_loop.py`
- Foresight and dependency planning: `skills/forward-planner/SKILL.md`
- Technical implementation planning: `skills/technical-plan/SKILL.md`
- Agentic iteration: `skills/agentic-loop/SKILL.md`
- Review: `skills/agentic-review/SKILL.md`, `skills/launch-review/SKILL.md`
- Outcome learning: `skills/analyzing-customer-patterns/SKILL.md`

## Human-Friendly Framing

The operator supplies direction, not an engineering-ready spec. The system's job
is to turn direction into a better artifact, verify it against reality, and
compound the result so the next build starts from a higher level.