prompts

📋 EU Parliament Monitor — Agentic Workflow Prompt Library

Classification: Public · Owner: Hack23 AB · Version: 2.0 (2026-04-21)

Purpose

Ten focused prompt files, one bounded context each, composed by every news-*.md gh-aw workflow. This replaces the 1,789-line SHARED_PROMPT_PATTERNS.md monolith.

File Index

#	File	Purpose
00	00-scope-and-ground-rules.md	Workspace scope, forbidden/allowed edits, neutrality, one-PR rule
01	01-data-collection.md	EP MCP feeds + direct fallbacks, deep-fetch, IMF primary economic + WB non-economic
01a	01a-data-fanout.md	Stage-A fan-out helpers — forward-statements registry seed, multi-day foreseen activities, Monday urgency sweep, quarter+ horizon fan-out (imported only by forward-looking / long-horizon workflows)
02	02-analysis-protocol.md	Pipeline, methodologies/templates, mandatory 2-pass
02a	02a-rerun-merge.md	Re-run improve/extend rule (same-day folder resume)
03	03-analysis-completeness-gate.md	Agent-side completeness gate over artifacts, IMF evidence, and metadata readiness
04	04-article-generation.md	Deterministic aggregator render, article.md, SEO title/description/keywords contract
05	05-analysis-to-article-contract.md	AI-first artifact contract and read-before-render duties
06	06-pr-and-safe-outputs.md	Single-PR rule, analysis-only fallback, noop diagnostics
07	07-mcp-reference.md	EP / IMF (primary economic) / WB (non-economic) tool tables, parameter corrections, reliability matrix
08	08-infrastructure.md	Required frontmatter, mcp-setup.sh, client env vars
09	09-troubleshooting.md	AWF firewall diagnostic, error→root-cause table

Import Order for Article Workflows (Stage A → D)

Every news-*.md except news-translate.md reads prompts in this order:

1. 00-scope — before anything else: what you may and may not touch.
2. 08-infrastructure — frontmatter you inherited from the workflow.
3. Stage A — Data Collection → 01-data-collection.md + 07-mcp-reference.md
4. Stage B — Analysis (2 passes) → 02-analysis-protocol.md
5. Stage C — Completeness Gate → 03-analysis-completeness-gate.md
6. Stage D — Deterministic article render → PR → 04-article-generation.md + 05-analysis-to-article-contract.md + ../../Article-Generation.md → 06-pr-and-safe-outputs.md
7. On error → 09-troubleshooting.md

Analysis Artifact Integration

Every article-generating workflow produces deep political analysis before rendering article output. The chain is:

Stage A · Data Collection (01-data-collection.md + 07-mcp-reference.md)
      → Stage B · Analysis Artifacts (02-analysis-protocol.md)
        authoritative protocol: analysis/methodologies/ai-driven-analysis-guide.md (10 steps)
        master artifact map:    analysis/methodologies/artifact-catalog.md
        per-artifact rules:     analysis/methodologies/per-artifact-methodologies.md
        39 templates:           analysis/templates/  (14 agentic-workflow [incl. 6 reusable framework] + 25 per-artifact)
    → Stage C · Completeness Gate (03-analysis-completeness-gate.md)
          per-artifact floors:  analysis/methodologies/reference-quality-thresholds.json
       → Stage D · Deterministic Article Render (04-article-generation.md + 05-analysis-to-article-contract.md + Article-Generation.md)
             Read-Before-Render: agent MUST read every artifact in analysis/daily/<run>/ before invoking the aggregator
             artifact → section map: 04-article-generation.md § 7.1 and Article-Generation.md § Templates and Artifact-to-Article Mapping
        → Stage E · Single PR (06-pr-and-safe-outputs.md) — one safeoutputs___create_pull_request call

An article that does not cite a specific analysis/daily/<run>/… artifact for each analytical section fails Stage C and is blocked from PR creation. methodology-reflection.md is the final artifact of every run (after workflow-audit.md).

Upstream gh-aw documentation (authoritative)

Every change to .github/workflows/*.md, .github/agents/*.md, or .github/prompts/*.md should be anchored against the upstream gh-aw doc sets:

• Abridged: https://github.github.com/gh-aw/llms-small.txt
• Full: https://github.github.com/gh-aw/llms-full.txt
• Blog series: https://github.github.com/gh-aw/_llms-txt/agentic-workflows.txt
• Repository: https://github.com/github/gh-aw

The pinned gh-aw version for this repo lives in .github/workflows/compile-agentic-workflows.yml.

The Single-PR Rule (summary)

Every article-generating workflow calls safeoutputs___create_pull_request exactly once, at the very end of the run, after all files are written. The only exception is news-translate.md, which uses multi-call flush with max-patch-size + re-calls. Checkpoint PRs, keep-alive heartbeats, and progressive safe outputs are banned and CI-lint-enforced by scripts/lint-prompts.js.

Full rationale: 06-pr-and-safe-outputs.md.

Drift-guard Lint (npm run lint:prompts)

scripts/lint-prompts.js enforces four rules across every .github/workflows/news-*.md file (news-translate.md is the single, fully-exempt workflow):

1. safeoutputs___create_pull_request appears at most once per workflow.
2. None of the banned phrases appear: checkpoint pr, checkpoint-pr, keep-alive, keepalive, keep alive, heartbeat, progressive safe output (case-insensitive).
3. No safeoutputs___push_repo_memory references.
4. Analysis-awareness — every news-*.md must either directly reference both analysis/methodologies/ai-driven-analysis-guide.md and 03-analysis-completeness-gate.md, or import .github/agents/news-generation.agent.md (which provides both anchors).

Execution surface:

• Local: npm run lint:prompts
• CI: job compile-agentic-workflows.yml runs node scripts/lint-prompts.js before compiling .md → .lock.yml.
• Unit test: test/unit/lint-prompts.test.js locks the rule set against regressions.

SHARED_PROMPT_PATTERNS.md — redirect stub

The former 1,789-line monolith has been decomposed into this 10-file library. SHARED_PROMPT_PATTERNS.md now exists only as a redirect stub mapping old sections → new files. New contributions must land in the appropriately-numbered file above; do not re-expand the monolith.

Token Discipline

• Each prompt file ≤ 300 lines, one bounded context.
• No file repeats content from another file; cross-links only.
• Bash diagnostics live in scripts/, not in prompts.
• Reliability tables live in 07-mcp-reference.md only.