Skip to content

Latest commit

 

History

History
59 lines (47 loc) · 2.81 KB

File metadata and controls

59 lines (47 loc) · 2.81 KB

AGENTS.md

Compound Beads v3.0 | Project: Alien Eyes | Initialized: 2026-03-04

Methodology: Compound Beads

Core loop: START ROUND > WORK > COMPOUND (Arc + learnings) > CLOSE (push + update)

Files: CLAUDE.md (handoff) | .compound-beads/QUICKSTART.md (continuity) | context.md (state) | rounds.jsonl (history) | learnings.md (insights)

Round types: feature | bug_fix | triage | polish | infrastructure Sizing: 30min-4hr. Break larger work into multiple rounds.

Session Protocol

Start: Read QUICKSTART.md > context.md > scan recent learnings End: git commit > session intelligence capture > update tracking files > regenerate QUICKSTART.md > git push Rule: Work is not done until pushed AND tracking files updated.

Auto-Triggers

Condition Action
Session start + .compound-beads/ exists Load QUICKSTART.md and context.md
Context window > 80% full Run handoff protocol
Round has >5 file modifications Update context.md
Round marked complete Capture Arc + extract learnings
Session ending detected Run session close protocol
Significant work completed Update CLAUDE.md
Pattern discovered Add to learnings.md
Bead open > 7 days Prompt: close, defer, or update?
Completion signals ("that worked") Capture learnings
3+ decisions made Capture rationale
Error or dead end Record for future avoidance

Narrative Capture

Every round: We started believing > We ended believing > The transformation

Project Rules

  • Separation of concerns: The tool that BUILDS must never have access to the test scenarios. This is cardinal.
  • Probabilistic default: Satisfaction scores (0-1), not boolean pass/fail. Deterministic only when legally required.
  • Agent-native first: Every feature accessible to both humans and agents. Parity is non-negotiable.
  • Dual scoring: Every audit produces both a Human-Native Score and an Agent-Nativeness Score.
  • Privacy by default: Users can opt out of storing results. Rhumb-bound data is anonymized.
  • Reuse GMPF patterns: Envelope, browser pool, model router, phased orchestrator — don't reinvent.

Skills

(pending discovery — run /compound:discover)

Tools & MCPs

playwright | Browser automation | screenshots, form testing, network interception, multi-device firecrawl | Web scraping | site crawling, content extraction exa | Search API | semantic search, competitor discovery, code context brave-search | Web search | research, current events replicate | ML models | image generation, analysis glif | Creative AI | image generation workflows context7 | Documentation | library docs, code examples