00 Gtm Router

GTM Router - the orchestrator (00)

Every other skill in this chain wraps a tool. This one wraps the decision: given an ICP, which chain do you run, in what order, at what depth, and what will it cost. Tool-routing ("use Firecrawl for scraping") is table stakes; this skill does methodology-routing - the shape of the ICP picks the discovery path, volume picks the extraction depth, evidence type picks the signal source, every discovered list passes the qualification gate (01-icp-qualify) before credits are spent on it, and everything funnels through judgment (05) before anyone writes an email. Two model-judgment layers bracket the paid steps: the gate decides who is worth paying to know more about, the judge decides what to say to them.

The router itself spends nothing. All credits are spent inside component skills, each of which keeps its own spend gates. The router's job is to make the spend deliberate: plan first, estimate honestly, execute only after a go-ahead, and checkpoint between steps.

Inputs

• icp_description (required) - who the client is targeting, in plain English.
• icp_context - what the client sells, to whom, and what pain it removes (2-4 sentences). 05-signal-builder cannot judge without it, so collect it during planning, not four steps in. If it's missing, ask once.
• exclusions (optional, ask once) - competitors, current customers, do-not-contact domains. The gate (01-icp-qualify) enforces them for free; discovery skills also apply them as filters where supported.
• budget_ceiling, target_volume, urgency (optional) - shape the plan. Missing values don't block planning: assume, label the assumption, and show the math so the user can correct it.

Start where the data runs out

Discovery is only step one if the user has nothing. Route to the first step whose input is missing:

A domain list is ambiguous input - resolve it before routing. The same CSV routes two opposite ways: a target list gets enriched (03/04), a seed list gets expanded (01-prospeo-lookalike). "Here are our 40 closed-won accounts" or "our competitor's customers" is almost always seeds; a conference attendee export or a purchased list is almost always targets. When the framing is genuinely unclear, ask - routing seeds into 03 scrapes the customers they already have.

Inherited data passes through untouched - the chain contract is additive (headless-gtm-shared/CONVENTIONS.md): every record keeps its upstream fields, keyed by normalized domain.

Step 1 - classify the ICP shape

If the shape is genuinely ambiguous, ask one question - "could you find these companies in a B2B database by industry and size, are they local businesses you'd find on Google Maps, or are they defined by something that just happened?" - then commit. Don't run a full interview; one answer plus labeled assumptions beats five questions.

Step 2 - assemble the chain

Default chain per shape. Deviate when the modifiers below say to, and say why in the plan.

The calls that make these chains work:

• The gate runs before any paid step - and twice. 01-icp-qualify is free (in-session judgment), so it sits between discovery and the first credit-spending step in every chain: pass 1 judges fit on discovery fields, and only the survivors in its records.jsonl get enriched, scraped, or resolved. After 03/04 add evidence, its pass 2 re-judges the uncertain rows and can demote qualified ones - schedule it before 05 ranks anything. The plan hands the gate what it needs: the client profile (icp_context), the exclusion list, and the per-company downstream cost from unit-costs.md, so its spend-avoided line is a real number instead of a question back to the user. Discovery filters match labels; the gate judges fit - don't loosen the gate to keep volume up, fix the discovery filters instead.
• 04 is a layer, not one skill. 04-crustdata-signals covers what already happened (funding, joins, headcount); 04-theirstack-jobs covers what's open right now (job reqs). Both write the same additive records - 05 merges by domain - so run both when the play needs both kinds of evidence. The theirstack sizing count (~1 credit against a whole domain list) prices the open-req overlay before any commitment, so adding it to any chain is cheap to check.
• 01 is a layer too - filters vs seeds. 01-prospeo-discover builds a list from firmographic criteria; 01-prospeo-lookalike builds one from example companies. Route on what the client can actually articulate: a described ICP ("Series B fintech, 50-200, US") goes to discover, and example logos ("more like these 30 closed-won accounts") go to lookalike, which is the better starting point when the client's real pattern is tacit - they recognize a good fit but can't name the filters. Lookalike's Mode 2 closes the loop: it reads the 25 closest matches for their shared pattern, builds an ICP from it, and hands that to discover for the broad search. Use Mode 1 when the matches themselves are the deliverable, Mode 2 when the seeds are a means to a bigger list. One caution worth planning around: a niche seed set often classifies under a huge parent industry, so a Mode 2 ICP of industry + size + geo can balloon into six figures - the specificity lived in the keywords. Sanity-check the count and re-add the recurring keyword before exporting.
• b2b-saas scrapes selectively. 04-crustdata's structured signals (funding, headcount, hires) cover database-tracked companies well. Scrape (03) only the slice that needs page-level evidence - the top decile by 04 signal strength, or accounts where 04 came back empty. Scraping all of a 2,000-domain list burns credits on companies judgment will score 3/10 anyway.
• vertical-smb skips 04-crustdata. Funding/headcount vendors barely cover owner-operated businesses; the website and Maps listing are the signal source, so the scrape is not optional - it's most of the evidence 05 will get. If the client's angle is hiring, a 04-theirstack sizing count is the cheap way to test whether postings coverage exists before assuming it doesn't.
• hiring-defined routes by evidence type. Open reqs: 04-theirstack-jobs discover mode searches "companies with an open X req" directly - its free blurred count returns the job AND company yield before any spend (then 1 credit per job returned), so no superset is needed. Past joins ("just hired X"): nothing searches that directly, so build a bounded firmographic superset with 01, then filter by 04-crustdata-signals recent_hires. Enriching an unbounded superset is the dominant cost in this shape - cap the superset in the plan, and run the cheaper filter first.
• web-scattered starts from a proposed source of record. Propose the 2-3 places these companies are already listed - a licensing registry, professional college, trade association directory, or marketplace, with Maps (02) as the general fallback - and confirm with the user before extracting. Don't open by asking them to supply URLs: proposing the source is the plan's job, confirming it is the user's. 03's "Directory and registry extraction" turns the confirmed source into a company list, and the chain proceeds normally from there.
• Judge before you resolve - by default. 05's ranking decides which accounts deserve per-contact resolution spend, and the signal work is what makes the email worth sending. Flip to resolve-first (06 straight after discovery, signals only on the emailable set) when the client will email every resolvable account regardless of angle, or when per-domain signal spend clearly exceeds per-contact resolution spend at the expected coverage. Either order is defensible - the plan states which one it picked and why.
• Free gates shrink paid steps. 01-icp-qualify is the institutional one - always on, right after discovery. Beyond it, look for any other free call that disqualifies domains before a paid pass: a free employee scan that shows who even has the target function, a count-only size check, a cached prior run. Halving a per-domain credit step beats optimizing anything downstream of it.

Modifiers, applied in this order:

1. Volume sets extraction depth: under ~200 domains, 03 standard mode; 200-2K, standard on the signal-bearing slice and minimal elsewhere; over 2K, minimal mode or top-decile-only scraping. Deep mode is for short high-value lists.
2. Budget ceiling caps volume: compute max affordable records from references/unit-costs.md and say plainly if the target volume doesn't fit.
3. Urgency trades depth for speed: minimal scrape mode, default enrichment windows, skip nice-to-have overlays. It never skips 05 - unjudged lists produce the generic copy this whole chain exists to avoid.

Pilot first, always. Propose a 100-200 record pilot end-to-end before full volume (smaller for small targets). Experiments beat opinions: a pilot validates per-unit cost, the qualification pass rate, hit rates, and signal quality before the budget is committed. Full volume on the first pass needs an explicit reason. The pilot also carries the gate's calibration rounds, so the full run inherits an already-corrected qualification brief.

Step 3 - estimate cost and time

Read references/unit-costs.md - don't estimate from memory - and build per-step line items: volume x unit cost, as a range. Every step after the gate prices on post-qualify volume (discovered x pass rate); the pass rate is a labeled assumption until the pilot trues it. Flag which numbers are plan-dependent (credit prices vary by vendor plan) and which steps dominate the total. Estimate wall-clock time from the rate limits in the same file; enrichment RPM, not scraping, is usually the bottleneck on large runs.

Step 4 - present the plan, then stop

## Campaign plan: {one-line ICP}
**Shape:** {shape} - {one sentence why}
**Entry point:** {where and why, if not discovery}
**Volume:** {target} · **Budget:** {ceiling or "none given"} · **Pilot:** {size}

| # | Step | Skill | In -> out | Est. cost | Est. time |
|---|---|---|---|---|---|

**Total:** {range} · **Dominant cost:** {step and why}
**Why this chain:** {2-4 bullets - only the non-obvious calls}
**Alternate:** {1-2 lines - what changes if budget or volume moves}
**Assumptions / gaps:** {labeled assumptions; anything to confirm}

Run the pilot ({n} records, ~{cost})?

Present the plan and wait. Don't start executing in the same turn - the gap between plan and execution is where the user corrects the ICP, the budget, or the angle, and a correction after step 1 has already spent credits is a refund nobody issues.

Step 5 - execute, checkpointing between steps

On go-ahead:

1. Create a run folder at ./runs/<run-id>/ in the working directory (timestamp+slug, e.g. 2026-07-21-00-router-yoga-gta): the approved plan.md plus manifest.json tracking {step, skill, run_dir, status, est_cost, actual_cost} per step.
2. Run each step by its own SKILL.md. Every component skill writes into that same ./runs/ root under its own step-prefixed run-id; record each path in the manifest. Hand the previous step's records.jsonl to the next step by explicit path - never a reformatted copy, never a guessed sibling folder.
3. Plan approval covers the spends listed in the plan. When a component skill asks to confirm a spend within its plan line, proceed; anything above the line, or any step not in the plan, stops and asks. 01-icp-qualify's own user gates stay live either way - the qualification brief gets approved before judging on a first client run, and uncertain rows are never forwarded silently. Those are judgment gates, not spend gates; plan approval doesn't answer them.
4. Checkpoint after each step: actual cost and record count vs estimate. Stop and re-plan if actual unit cost reaches 2x the estimate, output count lands under half of expected, or a step returns empty. A qualification pass rate far below the assumption is a discovery-filter problem - tighten the filters and re-run discovery rather than loosening the gate. Report the numbers either way - the estimate-vs-actual gap is how the unit-costs file gets better.
5. After 06, close out: final records.jsonl path, total spent vs estimated, count at each stage (discovered -> qualified -> judged -> resolved), the gate's spend-avoided total, and the handoff - top signals per account plus fallback go to email-writer, which needs the icp_context collected at the start.

When a step can't run

A missing skill or API key doesn't invalidate the route. Keep the step in the plan, mark it manual, name the tool and what to do by hand (e.g. "no CRUSTDATA_API_KEY - pull funding/headcount manually for the top 50, or add the key and run 04"). Degrade the plan, don't silently reroute around the methodology.

Scope boundaries

References

• references/decision-tree.md - the full routing logic as a readable doc: shapes, worked examples, the resolution waterfall. Shareable on its own.
• references/unit-costs.md - per-step unit costs, rate limits, estimation formulas, worked examples. Read before any estimate; update after any run where actuals diverged.
• ../headless-gtm-shared/CONVENTIONS.md - the chain contract: JSONL records, runs/ folders, additive fields, domain as key.