/gtm-audit

/gtm-audit

The flagship orchestration command for this plugin. Pipelines eleven analysis stages into a unified GTM intelligence brief — strategy through execution layer through PDF artifact.

Argument

$ARGUMENTS — a website URL (preferred) or brand name. If neither is provided, ask the user before proceeding.

Pipeline

Run these stages in order. Each stage consumes the prior stage's output. Use the Task tool with subagent_type to invoke each agent so they run in isolated context windows.

Stage 1 — Crawl + brand context (preflight)

Fetch the brand's site (homepage, /about, /pricing, /features, /customers if present). Extract:

• Brand name + tagline
• One-sentence value proposition
• Specific products / services
• Pricing model + price tier (premium / mid-market / budget)
• Target customer signals (case-study logos, ICP language)
• Industry + sub-vertical

Persist this as the "brand context" — every downstream stage references it.

Stage 2 — Competitor mapping

Invoke competitor-mapper-agent with brand context. Uses the competitor-discovery-cot skill (A3) to apply the cross-shop disqualifier and produce 3-6 defensible competitors, rejected candidates, and whitespace gaps. Block on output.

Stage 3 — Brand strategy (TAM/SAM/SOM + horizons)

Invoke brand-strategist-agent with brand context + competitor set. Runs the 6-dimension analysis using the tam-sam-som-horizons skill (C2): market sizing, niche, products, opportunities (quick wins / medium / long), customer journey, brand & messaging.

Stage 4 — Strategic SWOT

Apply the swot-analysis skill (B1) to brand using Stage 1+2+3. Produce S/W/O/T quadrants + cross-quadrant priorities with stopDoing fields. Strategic moves emitted here feed Stage 8.

Stage 5 — Market structure (Porter's Five Forces)

Apply the porters-five-forces skill (C1) to the sub-vertical. Produce five-force scoring + overall attractiveness + strategic implications.

Stage 6 — Channel scoring

Apply the marketing-channel-scoring skill (A1) against the 28-channel taxonomy in data/channel-taxonomy.json. Produce per-channel scores, rollout plan (now / next quarter / year two), and portfolio check. Active publishing channels feed Stage 10.

Stage 7 — Positioning whitespace (NEW — A2)

Apply /positioning-pass logic with Stage 1+2+3 as input. Score brand + competitors on the 5-dimension positioning space (price_tier, audience_sophistication, feature_depth, channel_fit, defensibility_commitment), set the founder's defensibility envelope, then run the max-min-distance Z3 template to find the provably-optimal positioning vector. Emit:

• Optimal positioning vector + current vector + min-distance to nearest 3 competitors
• Repositioning effort signal (dimensions where current is >2.0 from optimal)
• 3 sharper positioning alternatives (anchor / mechanism / outcome)
• Content pillars (4-6) derived from the chosen positioning — these feed Stage 10
• MaxSAT-ready claims per the gtm-output-schemas atomic-claim contract

Skip if no clear competitor positioning vectors emerged from Stage 2 — output would be hand-wavy.

Stage 8 — Competitor war-gaming (NEW — E1)

Apply /war-game logic with Stage 1+2 + candidate moves drawn from Stage 4 SWOT (O and S quadrants) and Stage 7 positioning alternatives. Auto-infer predicate from move language (pricing → 4a / feature → 4b / channel → 4c); default 4b. Run the quantifier-alternation Z3 template with manual skolemization, 30s timeout, combo-space precheck. Emit:

• Durable moves table (each with proof: kill scenarios it survives)
• Kill scenarios for non-durable moves (the specific response combo that defeats each)
• Predicate used + founder data gaps
• MaxSAT-ready durability claims

Skip if Stage 4 produced fewer than 3 candidate strategic moves — nothing to war-game.

Stage 9 — MaxSAT claim synthesis (renumbered — D1)

After Stages 2-8 complete and before writing the synthesis document, run the MaxSAT step. Replaces naïve recommendation concatenation with a provably-optimal consistent subset.

Claim collection. Scan recommendations from Stages 2-8 (now includes positioning whitespace claims from Stage 7 and war-game durability claims from Stage 8). Accept a claim only if all four MaxSAT fields are present: claimId, atomicClaim, weight (1-10 int), confidence (0.0-1.0 float). Skip incomplete recommendations — do not default missing fields. Collect incompatibleWithClaimIds edges and build incompatible_pairs by mapping claim IDs to their 0-based index.

If fewer than 3 claims pass collection: skip the solver and fall back to prose synthesis. Note "insufficient MaxSAT-ready claims for solver synthesis" at the top.

Solver call (uses solver-maxsat — NOT solver-z3). Consult solver-patterns skill Template 6:

1. mcp__solver-maxsat__clear_model
2. mcp__solver-maxsat__add_item(1, ...) — claim data (list of dicts, incompatible_pairs)
3. mcp__solver-maxsat__add_item(2, ...) — WCNF build + RC2 solve + export_solution(result_dict)
4. mcp__solver-maxsat__solve_model(timeout=10000)

Parse the result:

• selected_claim_ids — render as the synthesis recommendations
• dropped_claim_ids — list in the "Filtered claims" appendix with one-line reason (identify the conflicting selected claim if determinable)
• status == "unsatisfiable" → note "no synthesis-ready claims available; all recommendations require manual review"
• Timeout/error → fall back to prose synthesis with a visible warning

Stage 10 — 14-day content calendar (NEW — E2)

Apply /content-calendar logic with Stage 7 pillars + Stage 6 active publishing channels as input. If Stage 7 produced fewer than 3 pillars, fall back to brand-context-derived pillars per content-calendar.md's own fallback logic. Run solver-mzn with the assignment-with-diversity template, 30s timeout, feasibility precheck.

Emit the D×P schedule grid, pillar coverage summary, cadence summary, and sensitivity notes. Write JSON sidecar to content-calendar-{brand-slug}-{YYYY-MM-DD}.json.

Skip if the brand explicitly opts out of execution-layer planning (strategic-only audit).

Stage 11 — PDF render (NEW — best-effort)

After the synthesis markdown is written, render a PDF artifact alongside it. From PowerShell:

node scripts/report-bundler/bundle.mjs <path-to-synthesis-md>

Stage 11 must:

1. Check if scripts/report-bundler/bundle.mjs exists. If not, log a warning and report the markdown path only.
2. Check if node is available on PATH.
3. Invoke the bundler with the markdown path.
4. Report the PDF path to the user on success.
5. On any failure (missing bundler, missing npm dep, Puppeteer can't launch headless Chromium, etc.): log the error, surface the markdown path only, do NOT block the audit.

The bundler outputs gtm-audit-{brand-slug}-{YYYY-MM-DD}.pdf next to the markdown.

Solver serialization (critical)

The solver-z3 MCP server holds shared session state. Stages 2, 3, 4, 5, 6, 7, 8 must all run sequentially — never invoke them from parallel Task sub-agents. Stage 9 (solver-maxsat) and Stage 10 (solver-mzn) each use a different MCP server but the same serial discipline applies: one solver invocation at a time.

Synthesis output

After Stage 10 completes, produce a single synthesis document. Only render recommendations that appear in selected_claim_ids from Stage 9. Dropped claims go to the appendix. Section order:

1. Executive summary (5-7 bullets, every bullet cites a stage's evidence; only solver-selected claims from Stage 9)
2. The market and the niche (Stages 3 + 5 condensed)
3. The competitive set (Stage 2 — top 3-5 competitors + whitespace gaps)
4. Positioning whitespace (Stage 7 — proposed positioning vector, min-distance to nearest 3 competitors, which dimensions provide most separation, primary recommended option)
5. Strategic priorities (Stage 4 + 3 — only MaxSAT-selected, ranked by weight × confidence, with explicit stop-doings)
6. Competitor war-gaming (Stage 8 — durable moves table with proof, kill scenarios for non-durable moves, predicate used)
7. Channel mix and rollout (Stage 6 — phase 1 / phase 2 / phase 3 — only selected claims)
8. 14-day content calendar (Stage 10 — calendar grid day × platform, pillar coverage summary)
9. Open questions and confidence (where data quality was thin, what to investigate before acting)
10. Filtered claims appendix (Stage 9 dropped claims — one line each: claim ID, atomic claim text, reason dropped)

Write to gtm-audit-{brand-slug}-{YYYY-MM-DD}.md in the user's current working directory unless specified otherwise. Stage 11 produces gtm-audit-{brand-slug}-{YYYY-MM-DD}.pdf alongside it (when possible).

Quality bar

• Don't skip stages without saying so. If a stage fails or produces low-confidence output, surface that — don't paper over it.
• All references tie to canonical schemas. Synthesis isn't creative writing; every section ties back to structured agent output.
• Confidence is honest. Thin competitor data or missing pricing → audit confidence drops. Say so.
• Every recommendation is actionable. "Improve positioning" is not actionable. "Tighten positioning from 'AI marketing platform' to 'AI-powered demand-gen co-pilot for B2B SaaS founders post-PMF' to escape the HubSpot comparison" is.
• MaxSAT synthesis is not optional. If sub-agents emit MaxSAT-ready claims, Stage 9 must run before synthesis. Naïve concatenation is a regression, not a fallback.
• All 8 strategic stages contribute claims to MaxSAT. Positioning (Stage 7) and war-game (Stage 8) outputs MUST emit MaxSAT-ready claims per the gtm-output-schemas atomic-claim contract.
• Dropped claims get an explanation. Name which selected claim conflicted, or flag "no incompatible pair found — dropped by solver weight optimization."
• PDF is a convenience artifact, not the source of truth. The markdown stays authoritative. PDF failure does not block the audit.

When to skip stages

• Skip Stage 5 (Porter's) for very early-stage / pre-PMF brands — market structure analysis is premature.
• Skip Stage 6 (channel scoring) if the brand has explicitly asked for a strategic-only audit (no execution layer).
• Skip Stage 7 (positioning) if no clear competitor positioning vectors emerged from Stage 2 — output would be hand-wavy.
• Skip Stage 8 (war-game) if Stage 4 SWOT produced fewer than 3 candidate strategic moves — needs candidate moves to war-game.
• Skip Stage 9 (MaxSAT) only if fewer than 3 synthesis-ready claims exist across Stages 2-8, or if solver-maxsat is unavailable. Note the skip reason in the synthesis header.
• Skip Stage 10 (content-calendar) if the brand explicitly opts out of execution-layer planning (strategic-only audit).
• Skip Stage 11 (PDF) if scripts/report-bundler/ is missing — markdown stays the primary artifact regardless.
• Always run Stages 1-4 — they're the foundation.

Solver dependency

The full audit requires three MCP servers and one Node toolchain:

• solver-z3 — Stages 2-8 (registered in ~/.claude.json as mcp-solver-z3.exe). Serial: one stage at a time.
• solver-maxsat — Stage 9 (mcp-solver-maxsat.exe). Without it, Stage 9 falls back to prose with a prominent warning.
• solver-mzn — Stage 10 (mcp-solver-mzn.exe, MiniZinc on PATH). Without it, Stage 10 is skipped with a note.
• Node.js + npm + bundler deps — Stage 11. Requires scripts/report-bundler/bundle.mjs plus its installed Puppeteer dependency. Failure here is non-blocking.

Never silently degrade — always state the solver/bundler status at the top of the synthesis document.