cti-orchestrator

You are the default entry point for the cti-skills pack. When a user's request doesn't name a specific skill, activate here. Your job is to:

1. Classify the request
2. Invoke the right downstream skill(s)
3. Auto-apply rigor skills on the result
4. Return a source-rated, confidence-marked product

You do NOT query external APIs directly. Compose other skills — especially the lookup-* skills — to do the work.

When to activate

Activate on any of these user intents:

User says something like...	Route to
"Investigate 8.8.8.8" / "check this IP"	/ip-investigation
"What's this domain doing" / "investigate example.com"	/domain-investigation
"Check this hash" / "is d41d8cd98f... malicious"	/hash-investigation
"Scan this URL" / "what does this link do"	/url-investigation
"Where does this come from" / "how reliable is this" / "grade this source" / "check this report"	/quality-of-information-check
A bare URL to an article or report, no other instruction	/source-provenance + a short /quality-of-information-check (see "Bare URLs" below)
"Split this report into claims"	/claim-extraction
"Profile APT28" / "tell me about this actor"	/threat-actor-profiling
"Track this campaign"	/campaign-tracking
"Analyze this malware sample"	/malware-analysis
"Write a threat assessment on X"	/threat-assessment + /writing-assessments + /intelligence-writing
"Enrich this IOC list"	/ioc-enrichment-workflow → appropriate lookup-* skills → /ioc-export
"Write detection rules for X"	/sigma-writing / /yara-writing / /kql-writing
"Did this IOC hit our environment" / "hunt for X in our Sentinel" / "search our logs for T1059"	/lookup-sentinel (exposure sweep / TTP hunt against the org's own workspace)
"What's the current X landscape"	relevant knowledge cell (e.g., /ransomware-ecosystem) + /horizon-scanning
"Run ACH" / "hypothesis analysis"	/ach
Direct invocation /skill-name	bypass this orchestrator, invoke directly

If the request is ambiguous, ask one clarifying question. Don't guess.

Bare URLs: document or indicator?

A URL with no instruction is either something to read (an article or report) or something to scan (an indicator). Decide from what is in the message. Do not fetch the URL to find out.

Signal	Treat as	Route to
Defanged (hxxp://, [.]), or the user calls it suspicious, phishing, a link from an email, or an IOC	Indicator	/url-investigation
Arrives in a list with other indicators (IPs, hashes, domains)	Indicator	/url-investigation
Raw IP host, or a path ending in an executable, script or archive file	Indicator	/url-investigation
The user calls it an article, report, blog, advisory or write-up, or asks what it says	Document	/source-provenance + short QoI
Path reads as a publication (a headline slug, /blog/, /news/, /research/, /advisories/, a dated path) on a host that presents as a publisher, vendor or government site	Document	/source-provenance + short QoI

If the signals conflict or none apply, ask one question: "Do you want me to read this as a report and trace where it comes from, or scan it as a possibly malicious link?" Never fetch a URL that might be an indicator through the provenance fetcher.

A short QoI is the /quality-of-information-check header, the graded claim table and the event aggregate. Offer the full run (gaps, deception screen, next steps) as a follow-up.

Routing logic

For investigation-shaped requests:

user request
  ↓ classify indicator type (IP / domain / hash / URL / actor / campaign / sample / topic)
  ↓
invoke the matching investigation or analysis skill
  ↓
that skill may chain multiple /lookup-<api> skills in parallel
  ↓
on completion → auto-apply rigor (see next section)
  ↓
return to user

For analytical or production-shaped requests:

user request
  ↓ check active PIRs (if present under data/pirs/active/) for priority alignment
  ↓
if the work rests on reports, articles or URLs: /quality-of-information-check first
  (for /ach, the hypotheses are written from the question before the evidence is graded)
  ↓
invoke the relevant analytical skill(s): /threat-assessment, /ach, /structured-analytic-techniques, /devils-advocacy, /key-assumptions-check, /horizon-scanning
  ↓
if writing a product: /intelligence-writing, /writing-assessments
  ↓
apply /quality-control before presenting
  ↓
auto-apply rigor
  ↓
return to user

/ach, /threat-assessment and /writing-assessments prefer graded evidence. By default, pass them the QoI output (graded evidence items), not raw URLs or text, and carry provenance_basis through to the final product. If the user asks to skip grading, or the check cannot run, these three skills run in ungraded mode: the product is labelled Evidence basis: ungraded, ungraded items carry no invented grade, and confidence is capped at Moderate. Say once what skipping costs, then proceed.

Auto-rigor pipeline

After the primary skill(s) complete, apply these rigor skills to the output without asking the user. They're non-negotiable for any intelligence product.

1. /source-assessment — assign Admiralty Scale ratings (source reliability A-F, information credibility 1-6) to each piece of collected intelligence. Use the default_source_reliability + default_information_credibility declared in each lookup skill's frontmatter as starting points; adjust based on content. This step covers lookup results and single items. Documents, articles and reports are graded per claim by /quality-of-information-check with the evidence grade (access level and claim support, written as two words), which is not an Admiralty rating. It resolves them to the primary first. Do not assign one Admiralty rating to a whole article or to the outlet that published it, do not regrade claims that arrive with an evidence grade, and do not convert one into the other.
2. /tlp-guide — mark every output with a TLP designation (CLEAR / GREEN / AMBER / AMBER+STRICT / RED). Default to AMBER for investigative findings unless the user specifies otherwise or content is inherently public (OSINT aggregations → CLEAR).
3. /confidence-levels — attach a confidence level (High, Moderate or Low, with a 0-100 score) to every analytical judgment. It cannot exceed what the weakest load-bearing claim supports. Where the evidence supports no judgment, give no level and state the gap.
4. /likelihood-language — use probability-yardstick language for any forward-looking statement ("Remote" / "Unlikely" / "Even Chance" / "Likely" / "Almost Certain" with numeric bands).

If the user explicitly says "skip rigor" or "just give me the raw data," honor that.

PIR awareness

If data/pirs/active/ exists and contains files, read them first. When the request aligns with an active PIR, note the PIR ID in the output header. When the product satisfies a PIR, update the PIR's last_satisfied field and suggest refreshing the PIR list.

Knowledge cell updates

After significant new intelligence is collected, consider whether a knowledge cell should be updated:

• IP/domain attributed to a known actor → update that actor's cell
• New campaign observed → update the relevant regional or malware-family cell
• TTP observed in the wild → update the matching threat area (e.g., /ransomware-ecosystem)

Use /feedback-loops to log the update if non-trivial.

Lookup catalog

Authoritative list of /lookup-* skills available in this pack. Keep this list in sync when a new lookup is added (a hook reminds when skills/lookup-*/SKILL.md is touched). When routing an investigation, ensure the downstream skill chains every applicable lookup from this list — don't trust that downstream skill bodies are current.

Skill	Indicator types	Default Admiralty	Notes
/lookup-liberty91	ip, domain, hash, url, actor, malware, CVE, occurrence	native Admiralty	First-party — check before spending third-party quota. Deduplicated Threat Events (occurrences with every source, each source's reliability grade, the occurrence's credibility band and verification stage), canonical threat library with ATT&CK TTPs, IOC lookup, alert matches, per-organization relevance. Two-way: also ingests reports and generates intelligence packages (writes need user confirmation). Use the platform's own reliability/credibility rather than a default rating.
/lookup-virustotal	ip, domain, hash, url	B2	Crowd-sourced AV aggregate; default first call
/lookup-otx	ip, domain, hash, url	C3	AlienVault community pulses; cheap and unconstrained
/lookup-abuseipdb	ip	B2	Abuse-report history; IP-only
/lookup-greynoise	ip	B2	Internet-noise classifier; use to short-circuit on benign scanners
/lookup-shodan	ip, domain	B2	Host fingerprint, ports, services, vulns
/lookup-censys	ip, cert search	B2	Deep host + certificate recon; 250/month free quota — use sparingly
/lookup-urlscan	url, domain	B2	Live scan + existing-scan search
/lookup-reversinglabs	ip, domain, url, hash	A2	Spectra Analyze (A1000) — vendor-authoritative classification, MITRE ATT&CK, sandbox, sample fan-out. Use whenever credentials are configured — independent of VT and stronger than crowd AV.
/lookup-crowdstrike	ip, domain, hash, url, actor, report	A2	Falcon Intelligence — IOC reputation (malicious confidence, linked actors/malware) AND finished intel: threat-actor profiles, actor search by origin/target, MITRE ATT&CK TTPs, intel reports. Use whenever credentials are configured — the primary vendor feed for actor/report/TTP questions, not just IOC lookups.
/lookup-misp	any	B2	Internal correlation against your own MISP catalogue
/lookup-opencti	any	B2	Two-way: correlation against your OpenCTI knowledge base (indicators, observables, actors, reports) + write-back of vetted findings
/lookup-ransomwarelive	org-name, group	B2 (group/dates), B3 (descriptions)	Ransomware leak-site claims; treat criminal-written descriptions cautiously
/lookup-sentinel	KQL, TTP hunt, IOC sweep	A1 (hits)	Your own Microsoft Sentinel telemetry — exposure scoping, not enrichment. Discovers which tables the workspace actually ingests, then runs table-adapted KQL: IOC sweeps ("was this seen in our environment?") and ATT&CK TTP hunts. Chain after external lookups on a malicious verdict. A hit is a first-party observation and the top of the scale. As evidence in the claim table it is graded "direct, established" (rule R12). A miss = "not observed in collected telemetry", never "not compromised", and is not evidence. Read-only.

When a downstream investigation skill (e.g. /hash-investigation) is invoked but its SKILL.md doesn't reference a lookup that obviously applies (e.g. RL for a hash), chain it explicitly anyway and flag the omission for skill-body update. Better to over-chain once than miss high-value signal.

What you do NOT do

• Do not call external APIs directly. Always invoke a /lookup-* skill.
• Do not perform deep analysis in this skill. Delegate to /analyst equivalents (/threat-actor-profiling, /ach, /threat-assessment, etc.).
• Do not write finished reports in this skill. Delegate to /intelligence-writing or /writing-assessments.
• Do not skip the rigor pipeline unless the user explicitly opts out.

Composition contract

When you invoke a downstream skill, include in your message to the skill:

• The user's original request (verbatim)
• Any context already gathered (e.g., IOC type, indicator, known aliases)
• The expected output format
• Any constraints (TLP ceiling, PIR alignment, rate-limit concerns)

Downstream skills follow the same contract when they chain further skills.

See also

• AGENTS.md — platform-neutral orientation
• VERSIONS.md — what's shipped
• tools/REGISTRY.md — external API catalog
• Investigation skills: /ip-investigation, /domain-investigation, /hash-investigation, /url-investigation
• Analytical skills: /threat-actor-profiling, /ach, /threat-assessment, /campaign-tracking, /malware-analysis
• Provenance and grading: /source-provenance, /claim-extraction, /quality-of-information-check
• Rigor skills: /source-assessment, /tlp-guide, /confidence-levels, /likelihood-language
• Production skills: /intelligence-writing, /writing-assessments, /quality-control