Task Router¶

How to use this file (agents): find your job below, read the Read first items in order (they're short concepts), open the Deep dive source only if the task needs procedure-level detail, respect the Gates, and log where Log says. Do not load the whole KB. Paths are repo-relative; on the site, concepts live under /compiled/concepts/, sources under /compiled/sources/.

Two rules that apply to every job:
1. First-party data only — every number in outward-facing output traces to a real log/source, or it doesn't ship (human-review-gate enforces; see the agentisable fact-check gate).
2. Channel policy — owned channels may be full-auto after gates; X/LinkedIn are drafted-human-posts; HN/Reddit/Discord are human-only, always.

Read first: content-machine · experiment-loop · brand-voice-drift
Deep dive: marketing-automation-what-works (sequences & what converts), larry-agent-tiktok-growth (hook iteration, format economics)
Gates: voice linter + fact-check (agentisable marketing-factory/gates/); frame the experiment (hypothesis/metric/deadline) before drafting
Log: experiment file per piece; results per the experiments workflow

2. Write or update a GEO / LLM-search article (owned site)¶

Read first: geo-prompt-research → geo-content-pipeline → content-authority
Then: llms-txt · schema-markup · robots-txt (the machine-readable surface)
Deep dive: geo-factory-operations and llm-search-visibility-and-content-metrics (Princeton-validated tactics, +115% citation lift via external citations)
Gates: first-party-data rule; publish full-auto on owned site only after voice+fact gates
Log: citation-based experiment metric; measurement per job 7

3. Write cold outbound or a warm pitch¶

Read first: ideal-customer-profile · cold-email-sequence · email-deliverability · marketing-compliance
Deep dive: outbound-playbook (705 lines: deliverability, sequence design, LinkedIn mechanics, benchmarks)
Gates: drafts only — a human sends everything; compliance section of ai-marketing-risks applies (CAN-SPAM/GDPR)
Log: prospect tracker (agentisable outreach/); replies feed lead-scoring-model

4. Choose channels / plan the quarter¶

Read first: channel-selection · gtm-archetype · marketing-budget-allocation · marketing-benchmark-prior
Deep dive: channel-selection-planning (decision tree by ACV/cycle/ICP density), growth-models (flywheel vs bowtie vs ABM fit), benchmarks-as-priors
Gates: benchmarks are priors, not targets — update with own data per job 5
Log: decision + rationale in the project's plan file; revisit dates explicit

5. Design or conclude an experiment (small-N reality)¶

Read first: experiment-loop · minimum-detectable-effect · sequential-testing · bayesian-decision-rule
Deep dive: small-n-experiment-design (peeking trap, stopping rules, worked examples at 2–3 posts/week)
Gates: every experiment pre-registers hypothesis/metric/deadline; metrics trace to leads/signups/citations/revenue — never impressions
Log: experiments workflow — concluded experiments are KB knowledge

6. Pre-test variants with an SSR synthetic panel¶

Read first: synthetic-consumer-panels · semantic-similarity-rating · anchor-statements · calibration-loop · domain-transfer-risk
Runbook + artifacts: playbooks/ssr/ — personas, anchor sets, procedure, cost (~$0.12–0.35/test)
Deep dive: ssr-synthetic-panels; primary paper ssr-paper-arxiv-2510-08338
Gates: comparative use ONLY; advisory-only until 6–10 weeks of calibration against real post metrics
Log: every SSR call + eventual real outcome into the calibration log (see runbook)

7. Produce the weekly measurement report¶

Read first: closed-loop-attribution · geo-citation-measurement · incrementality-testing (when correlation isn't enough)
Deep dive: ai-marketing-measurement · incrementality-and-geo-experiments · marketing-data-layer (if tracking itself is the problem)
Gates: rank experiments by business outcomes, not reach; flag experiments past deadline with no result
Log: weekly report; update experiment result/learning fields; feed the KB per the experiments workflow

8. Research pricing / packaging¶

Read first: willingness-to-pay · pricing-packaging · synthetic-pricing-research
Deep dive: pricing-packaging-and-wtp (Van Westendorp, SSR tie-in)
Gates: synthetic WTP data is directional only — verify with real conversations before pricing decisions
Log: decision + evidence in the project's plan file

9. Risk & compliance review before anything ships¶

Read first: agentic-failure-modes · brand-voice-drift · prompt-injection · marketing-compliance
Deep dive: ai-marketing-risks (every risk has a mitigation; compliance cites actual regulations)
Gates: this job IS the gate — a human-review-gate instance
Log: blocked items with reason; recurring failure patterns become new KB concepts

10. Orchestrate a multi-agent campaign¶

Read first: workflow-vs-agent (start here — most "campaigns" are workflows) · agent-orchestration · shared-agent-memory · agent-ownership-boundary
Deep dive: marketing-agent-orchestration (five canonical patterns), agent-workflow-templates (per-channel trigger→input→action→output)
Gates: simplest coordination that works; human gates at hand-offs per channel policy
Log: orchestration decisions + failure escalations

11. Ingest new knowledge into this KB (meta)¶

Read: AGENTS.md (rules: provenance, naming, lint, logging) — that file, not this one, governs ingest
Gates: lint PASS; every load-bearing claim spot-checked at primary source; log entry in compiled/log.md or the run didn't happen

If your job isn't here: check All Pages, then add the job to this router once you've solved it — the router is maintained, not generated.