01For the outspent and outgunned
€0 beat €4 billion. Hungary, April 2026. A grassroots opposition with near-zero paid media took a parliamentary supermajority — against a twenty-year incumbent wired to a state machine and backed by every major global power that wanted it to stay.
A Claude Code plugin that turns a brief into an insurgent campaign plan. It encodes the asymmetric playbook that won that fight, and runs it against any incumbent with infinite budget and the platforms on their side.
Add the marketplace from GitHub, then install
$/plugin marketplace add https://github.com/bencium/bencium-marketplace
$/plugin install insurgent-campaign@bencium-marketplace
Watch the playbook
02What it refuses
Paid spend is amplification. Never the plan.
A normal marketing assistant reaches for the establishment playbook by default: content calendar, ad spend, influencer outreach. An insurgent cannot win that game — it is the game the incumbent already owns. So the skill refuses it. It treats credibility and earned trust as the real bottleneck, paid media as a boost gate on top of content that has already proven it works, and incremental lift as the only number that counts. Platform-reported ROAS describes activity, not cause.
Most campaign-AI tools sell reach. This one sells the discipline that wins when reach is the rigged game.
03What makes it different
Four claims most marketing-AI tools cannot make.
01. Anti-fabrication is a discipline, not a hope.
Explicit rules cover unnamed competitors, industry peers, SEO keyword examples, and earned-media targets. When the brief says “the dominant incumbent” without naming anyone, the skill refuses to quietly fill in Coca-Cola, Nike, or McDonald’s. It stays at the level of abstraction the brief gave it. This is the sharpest part of the skill — and the loop in section 06 is what forged it.
02. A real refusal mechanism, not just “best practices.”
The message-market-fit gate is a hard stop. If the offer is not validated — no paid commitments, no audience language, no close signal — the skill refuses to plan, names the missing evidence, and routes you to a discovery cycle instead. The intent is not to always hand back a campaign. The intent is to keep you from running the wrong one.
03. Modular references, loaded on demand.
Archetypes, sector riders, asymmetry tables, channel tiers, the authenticity playbook, lift-test templates, and the Hungary case study live in separate files. The workflow pulls each one only when a stage actually needs it, instead of cramming every decision into one monolithic prompt. The skill stays legible; the agent stays accurate.
04. Orchestration-agnostic. Runs in any agent harness.
The skill is portable by design. It ships as a Claude Code plugin, and the same files run unchanged inside Codex, Pidev, Opencode, or any agent environment that can read a prompt. Where the workflow says “ask the user,” it uses whatever structured-question facility your harness provides, or falls back to plain text. You are not locked to one orchestrator.
04The kinds of fights
Anywhere the spending math is rigged and the trust math is not.
- 01Indie SaaS foundervsBig 4 AI-advisory arm
- 02Open-source maintainervsVC-funded commercial fork
- 03Single-product B2B startupvsenterprise-suite incumbent
- 04Niche analytics toolvscategory-king platform
- 05Privacy-first consumer appvsad-funded incumbent
- 06Local newsroomvsprivate-equity media rollup
- 07Independent bookstorevsglobal logistics moat
- 08Indie game studiovsAAA publisher marketing
- 09DTC skincare brandvslegacy beauty conglomerate
- 10Specialty coffee roastervsmultinational chain
- 11Boutique hotelvsglobal hospitality chain
- 12Solo educator, cohort coursevsbootcamp marketing budget
05How it works
Three moves.
- You paste a campaign brief. Any length. Even one sentence.
- The skill runs five stages behind a message-market-fit gate. If the offer is not validated, the gate refuses the plan and routes you to a discovery cycle first.
- You pick a campaign shape. The skill returns first-thirty-days actions, a lift test, and an anti-vanity dashboard — in full, not in fragments.
See a real run: enterprise AI adoption, categorical asymmetry, €0 budget
A solo educator teaching enterprise AI adoption, outspent by the named consultancies’ AI-advisory arms. Excerpt from a live run.
Section 0. Assumptions
| Outcome | Cohort and workshop signups |
| Audience | VPs of Engineering at mid-market |
| Sector | Cohort education, personal-brand overlay |
| Budget yours | €0 / month |
| Budget incumbent | Categorical (effectively unbounded) |
| Bottleneck | Trust |
Section 3. Spend-asymmetry verdict
Your asymmetry is categorical. You cannot be in a spending race. This is the good news: the recommended playbook, Tier 1 to Tier 2 only, refuse broad cold-paid, counter-position against incumbent saturation, is the right answer at this ratio, not a compromise.
Preconditions score: 5 / 6. Credible insider, accumulated grievance, felt pain, threshold-rewarding system, and overplayed incumbent all present. Consolidated challenger field is partial; differentiation matters more than usual.
Section 6. Competitor saturation (excerpt)
What they saturate: polished maturity-assessment frameworks, multi-author PDFs, op-eds, conference circuit presence, corporate-voiced thought leadership authored by rotating partners.
The absence that becomes your signal: your named individual voice, with real engagement detail, published consistently on one platform. They predict; you report.
The incumbents publish AI roadmaps for the board. I help you ship the one that actually runs.
Section 7. Three shapes
- Founder-arc plus curriculum. Weekly post on one platform for twelve months, with the cohort curriculum published openly. Slowest to revenue, highest compounding.
- Earned-media-first. One signature artifact, then a six-to-eight week earned-media push. Fastest measurable spike.
- Community-first. Peer-validated admissions for one in-person workshop. Fastest to revenue, no public artifact.
Full output runs roughly seven hundred lines including the week-by-week plan, the lift test, and the anti-vanity dashboard.
06The engineer
I build skills that hold their edge as models change.
An agentic engineer’s claim is worth what their loop produces. So I built a self-grading loop that runs on every skill I ship. Here is what it did to this one.
The problem.
The skill looked solid. It had stages, refusals, case studies, all the structure you would want. Then I tested it.
Give it a brief that says “we compete with a well-known incumbent” without naming anyone, and the AI quietly fills in the blank. It picks a real company from its training data: Coca-Cola, Nike, whichever felt right. The whole campaign anchors on that guess. Concepts, search keywords, earned-media pitches, all built on a name the user never gave.
The user takes that plan to a client meeting. Weeks later: “actually our competitor isn’t Coca-Cola.” Too late.
This is the AI failure mode that does not show up in benchmarks. It looks right. It passes a quick eyeball review. The damage surfaces after it ships.
The fix.
- 01Six yes-or-no questions that define “good output.”
- 02An AI judge grades every generation against those questions.
- 03An AI optimizer proposes surgical edits to the skill prompt.
- 04Keep only edits that strictly improve the aggregate score.
- 05Log every attempt: what was kept, what was reverted, why.
I walked away. The loop ran.
The numbers.
| Iter | Brief | Mutation | Score | Verdict |
|---|---|---|---|---|
| 00 | SaaS | baseline | 16 / 18 | · |
| 01 | SaaS | anti-fabrication rule, stage 2 | 17 / 18 | keep |
| 02 | NGO climate | anti-fab carry-through, stage 4 | 18 / 18 | keep |
| 03 | legal-tech terse | industry-peer rule, stage 2 | 18 / 18 | ship |
107 / 108 = 99.1% across six sector briefs: SaaS, NGO climate, political underdog, solo consumer brand, cohort course, legal-tech terse. Forty-one minutes wall-clock. Roughly fifty cents of Haiku compute.
Why the log matters.
The improvement is not the asset. The log is. Every edit the loop tried, every score, every kept-or-reverted verdict.
When I edit this skill in six months, the log is my regression test. When a stronger model arrives next year, the loop picks up where this run stopped. Research data compounds.
Most teams treat AI as a tool they use. The teams building a real advantage treat AI as a system they teach. The teaching is what compounds.