This folder contains the complete automation kit for producing deep research reports on any topic using Hermes Agent. It was derived from the actual workflow used to produce "The State of Artificial Intelligence in Europe: A 2026 Report".

deep-research-pipeline/
├── README.md                          ← You are here
├── docs/                              ← Analysis & blueprints
│   ├── workflow_analysis.md           ← Pipeline reconstruction
│   └── automation_blueprint.md        ← Templates & orchestrator
├── templates/                         ← Reusable prompt templates (topic-agnostic)
│   ├── reusable_pass_prompt_template.md
│   ├── prompt_to_produce_pass_prompt_from_template.txt
│   └── prompt_files/                  ← Parameterized stage prompts
│       ├── initial.md                 ← Stage 1: Initial research
│       ├── assessment.md              ← Stage 2: Quality assessment
│       ├── cleanup.md                 ← Stage 7: Cleanup
│       └── meta_prompt.txt            ← Meta-prompt generator
├── scripts/                           ← Python scripts
│   ├── orchestrator.py                ← Full pipeline orchestrator
│   ├── generate_charts.py             ← Chart generation (matplotlib)
│   └── integrate_charts.py            ← Chart insertion into document
├── config/                            ← Pipeline configuration
│   └── cron_job_config.yaml           ← Cron job definitions for chaining
├── references/                      ← Reference materials (skill copies, source docs)
│   └── deep-research-framework/     ← Copy of the deep-research-framework skill
│       └── SKILL.md                 ← The methodology that drives this pipeline
└── case-studies/                    ← Example case studies
    └── european-ai-2026/              ← The Europe report (example)
        ├── artifacts/                 ← Report versions, charts, exports
        ├── analysis/                  ← Quality assessment, improvement plan
        ├── prompts/                   ← Original prompts used
        └── pass-prompts/              ← Generated pass prompts

The entire pipeline is built on the deep-research-framework skill (a copy of which is available in references/deep-research-framework/SKILL.md). This skill defines the 10-component quality methodology that drives every stage of the pipeline — from the initial research persona and PICO framing, through the source quality hierarchy, to the quality assessment criteria used in Stage 2.

The skill is kept as a reference copy in this folder so the pipeline is fully self-contained: you can reproduce the workflow on any machine without depending on the installed skill set.

# Run the full pipeline for a new topic
python3 scripts/orchestrator.py \
  --topic "Your Research Topic" \
  --output-dir /path/to/output \
  --passes 4

# Dry run to see what would happen
python3 scripts/orchestrator.py \
  --topic "Your Research Topic" \
  --output-dir /path/to/output \
  --dry-run

The pipeline consists of 9 sessions across 7 logical stages:

Session 1:  Initial Research → report.md
Session 2:  Quality Assessment → assessment.md + improvement_plan.md
Session 3A: Pass 1 Meta-Prompt → pass_1_prompt.md
Session 3B: Pass 1 Execute → pass1 report
Session 4A: Pass 2 Meta-Prompt → pass_2_prompt.md
Session 4B: Pass 2 Execute → pass2 report
Session 5A: Pass 3 Meta-Prompt → pass_3_prompt.md
Session 5B: Pass 3 Execute → pass3 report + charts
Session 6A: Pass 4 Meta-Prompt → pass_4_prompt.md
Session 6B: Pass 4 Execute → final report
Session 7:  Cleanup → clean report + exports

For each improvement pass, there are two separate Hermes sessions:

1. Meta-Prompt Generation — Reads the reusable template + improvement plan, extracts tasks for the specified pass number, and generates a complete pass prompt. This is the prompt_to_produce_pass_prompt_from_template.txt command.

2. Pass Execution — Reads the generated prompt + current document, executes the tasks, and produces the revised document.

This means no pre-written prompt files are needed. The improvement plan is the single source of truth. If you update the plan, all pass prompts regenerate automatically.

python3 scripts/orchestrator.py \
  --topic "Your Research Topic" \
  --output-dir /path/to/output \
  --template-path templates/reusable_pass_prompt_template.md \
  --meta-prompt-path templates/prompt_to_produce_pass_prompt_from_template.txt \
  --passes 4

1. Import the cron job definitions from config/cron_job_config.yaml
2. Replace {TOPIC}, {OUTPUT_DIR}, {TEMPLATE_PATH}, and {PLAN_PATH} with actual values
3. Run the jobs in sequence using context_from to chain them

1. Run Session 1 with templates/prompt_files/initial.md
2. Run Session 2 with templates/prompt_files/assessment.md
3. For each pass N (1-4): a. Run the meta-prompt with templates/prompt_to_produce_pass_prompt_from_template.txt b. Run the pass with the generated pass_{n}_prompt.md
4. Run Session 7 with templates/prompt_files/cleanup.md

To customize for different topics:

1. Replace the topic — Change {TOPIC} in the prompt files
2. Adjust research parameters — Modify depth, source mix, time period, etc.
3. Customize chart generation — Edit scripts/generate_charts.py for topic-specific charts
4. Adjust cleanup rules — Modify templates/prompt_files/cleanup.md for topic-specific cleanup
5. Add/remove passes — Change --passes in the orchestrator or adjust the cron config

1. Fresh sessions per pass — Each session is completely isolated. No cross-session memory dependency. This respects the context limit (up to 180K).

2. Two-level automation — Pass prompts are generated at runtime from the template and improvement plan. No pre-written prompts needed.

3. Sequential execution — Passes run in order because each pass builds on the previous one.

4. Improvement plan is the single source of truth — If you update the plan, all pass prompts regenerate automatically.

1. The two-level automation is the key insight — separating prompt generation from prompt execution means no pre-written prompts are needed.

2. The meta-prompt is the most reusable artifact — prompt_to_produce_pass_prompt_from_template.txt is a generic, parameterized command that works for any pass number.

3. The 4-pass strategy works — Separating concerns (methodology → prose → visualization → QA) produced a significantly better result than trying to fix everything at once.

4. Self-contained prompts are essential — Every pass execution session successfully read all three input files and produced the correct output.

5. Quality assessment is the bottleneck — Stage 2 produced the improvement plan that guided all subsequent work. A poor assessment would have led to poor improvements.

6. Chart integration is mechanical — The integrate_charts.py script proved that chart insertion is a deterministic operation that can be automated.

7. Cleanup is topic-specific — The diagnosis of what to remove from the final report depends on the internal artifacts generated, which vary by topic and research depth.

topic + parameters
        │
        ▼
┌─────────────────────────────────────────────────────────────┐
│  STAGE 1: INITIAL RESEARCH                                  │
│  hermes-cli run --prompt-file templates/prompt_files/initial.md       │
│  Output: report_<topic>.md                                  │
└─────────────────────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────────────────────┐
│  STAGE 2: QUALITY ASSESSMENT                                │
│  hermes-cli run --prompt-file templates/prompt_files/assessment.md    │
│  Output: assessment.md + improvement_plan.md                │
└─────────────────────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────────────────────┐
│  STAGE 3-6: IMPROVEMENT PASSES (x4)                         │
│  For each pass N:                                           │
│    hermes-cli run --prompt-file templates/prompt_to_produce_pass...   │
│    hermes-cli run --prompt-file pass_{n}_prompt.md          │
│  Output: report_<topic>_pass{1-4}.md                        │
│         charts/ (if applicable)                             │
└─────────────────────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────────────────────┐
│  STAGE 7: CLEANUP                                           │
│  hermes-cli run --prompt-file templates/prompt_files/cleanup.md       │
│  Output: report_<topic>_clean.md                            │
│         report_<topic>.docx                                 │
│         report_<topic>.pdf                                  │
└─────────────────────────────────────────────────────────────┘

This automation kit is derived from the actual workflow used to produce the "State of AI in Europe" report. It is provided as-is for reuse and customization.