daily-intel-brief

name: daily-intel-brief description: Produce Trevor's Daily Intelligence Brief — a dated, structured intelligence product with ten regional sections (Europe, North America, Central America & Caribbean, South America, Africa, Middle East, Central Asia, South East Asia, Oceania, Prediction Markets), each carrying calibrated key judgments and prediction bands. Trigger this skill on "daily brief", "morning brief", "today's brief", "INTSUM", "daily intel", "Trevor brief", "run the brief", "kick off today's product", "where are we today", or any standing-product request that implies the daily cadence. Composes existing analyst skills (sat-toolkit, source-evaluation, indicators-and-warnings, bluf-report, geospatial-osint, chartgen, mermaid, pdf-report) and the analyst/ scaffold; routes analysis to deepseek-v4-pro per ORCHESTRATION.md escalation criteria. Do not use this skill for ad-hoc single-region deep dives, single-incident analysis, or non-security topics — those should run through the standard analytic-workflow playbook. metadata: trevor: constraints: no_visual_assets: true # principal directive 2026-05-22 — text-only no_social_pipeline: true # principal directive 2026-05-22 — paused delivery: agentmail-email # principal directive 2026-05-22 — no PDF requires: bins: [python3] env: [DEEPSEEK_API_KEY] env_optional: [] composes: skills: - source-evaluation - indicators-and-warnings - bluf-report analyst_assets: - analyst/playbooks/analytic-workflow.md - analyst/playbooks/scenario-triage.md - analyst/playbooks/quality-gates.md - analyst/templates/bluf-report.md - analyst/templates/indicators-and-warnings.md - analyst/templates/ach-matrix.md - analyst/meta/sources.json

Daily Intelligence Brief

A standing product, not a one-off. Run once per day at the user's chosen hour (default: 06:30 local). Every run produces the same shape so that the principal recognises it on sight — same regions, same order, same field structure, same length budget. Variation is in the content, not the form.

Why this is a discrete skill

analyst/playbooks/analytic-workflow.md covers single-question assessments well. The daily brief is a different shape: multiple parallel mini-assessments (one per region), strict length budget, ruthless cadence. Trying to bend the analytic-workflow playbook into this shape every morning loses time and drift-prone consistency.

This skill orchestrates three subagents — collector, analyst, visuals — and assembles their outputs into the daily product. Each subagent has its own prompt under agents/ so the work parallelizes.

Product specification

A finished daily brief is a PDF (with the source .docx retained for revision) containing, in this exact order:

Cover page — date, classification banner ("UNCLASSIFIED // FOR OFFICIAL USE — TREVOR DAILY"), product name, BLUF.
Executive summary — five key judgments. Each has a Sherman Kent verbal anchor and a 7-day prediction percentage.
Six regional sections, in fixed order:
1. Europe
2. Asia
3. Middle East
4. North America
5. South & Central America (incl. Caribbean)
6. Global Finance
Annex A — sources & methodology (one line per source feed).
Annex B — I&W status board (running indicators that have fired or moved since yesterday).

Each regional section is structured per references/product-template.md (map, narrative, 3–5 incidents, 2–3 numbered key judgments with prediction bands). The Global Finance section swaps the map for a chart suite per references/visual-spec.md.

Length budget: 8–12 pages. If you blow past 12, you cut; you do not ship a 16-page "daily" brief.

Workflow (orchestrator's view)

You are the orchestrator. Run the steps below in order. Steps 2–4 can run their three subagents in parallel — the data hand-off is filesystem-based and they don't block each other once collection is done.

Step 0 — Frame the day

Read the brain's last 48h of episodic memory:

python3 brain/scripts/brain.py recall "yesterday's brief and overnight events"

Then create today's working directory:

WD=~/trevor-briefings/$(date -u +%F)
mkdir -p "$WD"/{raw,analysis,visuals,final}

Record 00-framing.md in $WD per analyst/playbooks/analytic-workflow.md Phase 0. The framing for the daily product is constant: "Provide Trevor's principal with the most decision-relevant security and financial picture from the last 24h, structured for ~5 minutes of read time."

Step 1 — Spawn the collector subagent

Read agents/collector.md. Spawn a subagent with that prompt. Hand it:

$WD/raw/ (output dir)
Today's UTC date
analyst/meta/sources.json (the durable source registry)

It produces $WD/raw/incidents.json — a normalised list of security-incident items tagged by region, each with a NATO Admiralty reliability/credibility code attached at point of capture (per skills/source-evaluation).

Wait for completion before Step 2 — the analyst needs the full incident set before scoring.

Step 2 — Spawn the analyst subagent

Read agents/analyst.md. Spawn a subagent with that prompt. Hand it:

$WD/raw/incidents.json
$WD/analysis/ (output dir)
The model identifier: deepseek/deepseek-v4-pro (escalation tier per ORCHESTRATION.md — the daily product is a high-stakes principal product and qualifies for V4 Pro).

The analyst calls the DeepSeek Direct API once per region (six calls) plus once for the executive summary, producing $WD/analysis/<region>.json and $WD/analysis/exec_summary.json. Each file contains the regional narrative, 2–3 key judgments with Sherman Kent bands, and per-judgment prediction percentages over a 7-day horizon.

Step 3 — Spawn the visuals subagent

Spawn this subagent in parallel with Step 2 (it doesn't need analyst output for the maps; it only needs incident locations from $WD/raw/incidents.json).

Read agents/visuals.md. Hand it:

$WD/raw/incidents.json
$WD/visuals/ (output dir)

It produces:

visuals/map_<region>.png × 5 (the five geographic regions; finance uses charts instead).
visuals/finance_charts.png — composed chart panel for the Global Finance section (FX, oil, equities, sovereign yields).
visuals/relationships_<region>.png — Mermaid-rendered actor/event graph for the most active region of the day.

Visual generation routes through:

skills/geospatial-osint for regional maps.
skills/chartgen (ChartGen API) for the finance suite.
skills/mermaid for the relationships diagram.

Step 4 — Quality gates

Run analyst/playbooks/quality-gates.md against the assembled analysis. If a gate fails, fix it before Step 5 — do not ship and footnote. Particular attention:

Gate 2 (calibration): every key judgment has both a Sherman Kent band and a numeric prediction band. The two must agree (verbal anchor matching numeric range per skills/source-evaluation).
Gate 3 (source hygiene): no single-source key judgment unless flagged as such in the JSON.
Gate 5 (red-team pass): at least one regional KJ has been red-teamed; record the challenge in $WD/analysis/red_team.md.

Step 5 — Assemble

Run the assembler:

python3 skills/daily-intel-brief/scripts/build_pdf.py \
  --working-dir "$WD" \
  --template skills/daily-intel-brief/templates/daily-product.html \
  --out-pdf "$WD/final/brief-$(date -u +%F).pdf"

The assembler builds the structured JSON expected by skills/pdf-report and invokes its renderer. The DOCX equivalent is produced in the same step via python-docx and saved alongside.

Step 6 — Archive and deliver

Append a one-paragraph entry to memory/$(date -u +%F).md per AGENTS.md Memory section. Capture: top 5 KJs, any I&W indicators that fired, any new sources promoted to durable list.
Refresh the brain index:
```
python3 brain/scripts/brain.py reindex
```
Deliver to the principal:
- Default channel: AgentMail (skills/agentmail) to the address in ~/.openclaw/openclaw.json → trevor.principal_email.
- Confirm send before pushing — agentmail is an external action.

Subagent boundaries

The point of the three subagents is parallelism + isolation, not just delegation. Keep each subagent's job tight:

Subagent	Reads	Writes	Does NOT
Collector	RSS, web, sources.json	raw/incidents.json	analyse, judge, predict
Analyst	raw/incidents.json	analysis/.json	re-collect, draw maps
Visuals	raw/incidents.json, analysis/	visuals/*.png	re-analyse, write narrative

When a subagent breaks discipline (analyst tries to recollect, visuals tries to write narrative), the orchestrator calls it back. This isn't bureaucratic — overlap is the failure mode that turns a 2-hour daily into a 6-hour one.

Operating guidance

On the prediction percentages. The analyst is instructed to return calibrated 7-day probabilities, not theatrical ones. A prediction of "85%" means the analyst would take a bet at 85:15. If V4 Pro returns suspiciously round, high-confidence numbers across the board (e.g. everything in 80–95%), that's a calibration smell — surface it in the hand-off to the user. Over-confidence is the most common failure mode of LLM analysis and the daily product is the worst place to let it quietly compound.

On the regional taxonomy. The collector and analyst share references/regions.json. Don't ad-hoc reassign countries — Türkiye is in Middle East here (analyst convention), Russia is in Europe, Mexico is in North America, Cuba/DR/Haiti are in South & Central America. If the principal wants a reclassification, edit regions.json; don't argue at runtime.

On reruns. If the principal says "rerun the brief" later in the day, do not delete the working directory. Spawn the collector again into raw/incidents-<HHMM>.json and re-run the analyst on the union of all incidents. Yesterday's brief stays in ~/trevor-briefings/$(date -u -d yesterday +%F)/ for retrospective calibration.

On API keys. The analyst expects DEEPSEEK_API_KEY. The visuals subagent expects MAPBOX_TOKEN (optional — falls back to OSM tiles) and CHARTGEN_API_KEY (optional — falls back to matplotlib). If required keys are missing, fail loudly at Step 0 rather than producing a degraded product silently. Tell the principal exactly which env var is missing.

On vocabulary. TREVOR brand voice (ORCHESTRATION.md): methodology, disclosure, candor, completeness, confidence, indicator, posture, finding, assessment. Avoid: solution, leverage, unlock, empower, disrupt, game-changing.

Files in this skill

skills/daily-intel-brief/
├── SKILL.md                              # this file
├── _meta.json
├── README.md
├── agents/
│   ├── collector.md                      # subagent prompt: news harvest
│   ├── analyst.md                        # subagent prompt: DeepSeek V4 Pro analysis
│   └── visuals.md                        # subagent prompt: maps / charts / diagrams
├── playbooks/
│   ├── daily-cadence.md                  # the 90-minute daily run
│   └── escalation-criteria.md            # when the daily flips into a special
├── references/
│   ├── regions.json                      # country → region taxonomy
│   ├── product-template.md               # the daily product structure (per section)
│   ├── deepseek-prompts.md               # per-region prompt templates for V4 Pro
│   └── visual-spec.md                    # what each visual must show
├── scripts/
│   ├── orchestrate.py                    # entrypoint that walks Steps 0–6
│   ├── collect.py                        # collector implementation
│   ├── analyze.py                        # DeepSeek V4 Pro client + prompt assembly
│   ├── build_visuals.py                  # invokes geospatial-osint, mermaid, chartgen
│   └── build_pdf.py                      # assembles the structured JSON for pdf-report
└── templates/
    ├── daily-product.html                # Jinja2 template for pdf-report
    └── daily-product.md                  # markdown mirror (working draft + DOCX source)