Customer Newsletter Generator

A fully automated, skills-based system for generating monthly GitHub customer newsletters. Built with hypothesis-driven development, layered scoring, and iterative editorial intelligence mining.

Quick Start

# Validate repository health
make validate-structure
make validate-all-skills

# Canonical production command
copilot --model gpt-5.5 --allow-all --deny-tool agent --no-ask-user --stream off \
  -p "$(bash tools/render_product_run_prompt.sh 2026-02-14 2026-04-16 production)"

Convenience wrapper over the same canonical command:

make newsletter-gen START=2026-02-14 END=2026-04-16 MODE=production

Retained proof-run wrapper over the same oracle:

make newsletter-proof-run START=2026-02-14 END=2026-04-16 MODE=production

The proof-run wrapper snapshots canonical workspace/ and output/ artifacts into retained private run evidence without replacing the raw prompt-rendered production authority.

Live operator paths:

• make newsletter-gen START=2026-02-14 END=2026-04-16 MODE=production
• make newsletter-proof-run START=2026-02-14 END=2026-04-16 MODE=production

Experiment-only optimization helpers:

• make newsletter-cost-profiler
• make newsletter-hotspot-auditor
• make newsletter-phase-experimenter
• make newsletter-orchestrated-proof

newsletter-orchestrated stays diagnostic-only for phase-local debugging and regression diagnosis. newsletter-orchestrated-proof packages that diagnostic path as retained evidence with phase-session telemetry for cost experiments.

This repository keeps the runnable pipeline, selected outputs, and release bundle material. Raw run logs and session evidence are not tracked here.

VS Code flow: open the repo, select the customer_newsletter agent, then run:

please generate the newsletter from scratch for the dates  to 

Benchmark command on the same prompt-rendered surface:

bash tools/prepare_newsletter_cycle.sh 2025-12-05 2026-02-13 --no-reuse
copilot --model gpt-5.5 --allow-all --deny-tool agent --no-ask-user --stream off \
  -p "$(bash tools/render_product_run_prompt.sh 2025-12-05 2026-02-13 benchmark)"

Primary Regression Benchmark

Public benchmark-mode contract: config/benchmark_modes/feb2026_consistency.json

Use this before relying on the diagnostic-only phase harness for regression claims:

bash tools/prepare_newsletter_cycle.sh 2025-12-05 2026-02-13 --no-reuse
copilot --model gpt-5.5 --allow-all --deny-tool agent --no-ask-user --stream off \
  -p "$(bash tools/render_product_run_prompt.sh 2025-12-05 2026-02-13 benchmark)"
bash tools/validate_pipeline_strict.sh 2025-12-05 2026-02-13 --require-fresh --benchmark-mode feb2026_consistency

Rule:

• treat the single-shot prompt-rendered copilot -p run as the production-like oracle
• use tools/run_newsletter_orchestrated.sh only after the single-shot benchmark fails and you need phase-local diagnosis