CustomerNewsletter: Instructions file for Codex

AGENTS.md

Primary AI instruction surface for briancl2-customer-newsletter (platform-agnostic). Monthly GitHub customer newsletter pipeline: six core phases plus optional polish, video-enrichment, and editorial stages, driven by LLM skills, scoring, and intelligence mining.

Operating Protocol

Every change follows this numbered workflow. No exceptions.

1. Hypothesize — State a testable prediction with explicit PASS criteria (HIGR)
2. Score — Define a scoring rubric with acceptance threshold
3. Plan — Design the approach before building
4. Build — Implement the change
5. Test — Validate against PASS criteria using layered scoring (cheapest first)
6. Fix — Iterate until all criteria pass
7. Review — Run make review on ALL changes before committing. --no-verify is NEVER permitted.
8. Validate — Run quality checks (scoring battery)
9. Document — Update HYPOTHESES.md, LEARNINGS.md, and HANDOFF.md. For published changes, also record scope and validation in the PR description.
10. Clean up — Delete old implementations (Deletion Discipline), archive stale artifacts
11. Report + Handoff — Write HANDOFF.md for session continuity

Core Principles

• Deletion Discipline: Replace directly. No _v2 suffixes, no parallel code paths.
• LLM-First: Agent-orchestrated workflows over deterministic scripts.
• Skills-First: Check .github/skills/ before implementing ad-hoc procedures.
• HIGR: Every change is a testable hypothesis with PASS criteria.
• Trust Disk, Not Self-Reports: Verify by reading files, not trusting agent claims.
• Layered Scoring (Cheapest First): Structural → Heuristic → Selection → Editorial.
• Feed-Forward Learnings: Every finding becomes an L-number in LEARNINGS.md.
• Benchmark-Grounded: Score against benchmark data, not intuition.

Agents (4)