The $274 CI Stack

The problem

Competitive intelligence breaks at the moment a page change turns into a claim. Monitoring tools catch the changes, and an AI model can summarize each one. The hard step is the claim a product marketer hands to sales: what changed, when, and what the evidence supports.

A live link can’t prove a competitor claim, because the page can change again. In the workshop’s main example, a competitor’s pricing page dropped its money-back guarantee on Aug 8, 2026. On Aug 14, a second monitor on the same page caught the guarantee coming back. Anyone holding only the link lost the proof six days later. The dated captures kept it.

The system

The CI stack is three paid tools and eight open-source skills. Visualping Business 20K ($140 a month, 200 pages, 5 users) watches competitor pages and keeps each dated before-and-after capture. Claude Max ($100 a month) runs the skills in Claude Code; they also run in Codex. tl;dv Pro ($29 a month) records and transcribes sales calls. At month-to-month list prices before tax, checked Aug 6, 2026, the three tools add up to $269 a month. The title rounds that up, so the stack runs for under $274.

The CI stack’s automation layer adds nothing to the bill. Run a skill by hand on a pasted diff, let the agent pull change history through Visualping’s MCP server, or schedule it in n8n or GitHub Actions. The repo ships no runner, so each team wires its own.

The eight ci-stack skills share one contract. A new CI program usually loops through three of them:

1. ci-portfolio builds a watchlist tied to a decision, with ignore rules and alert prompts.
2. ci-triage turns one change into a draft decision brief with the evidence, the uncertainty and one recommended action. It can also build a dated timeline from supplied change history.
3. ci-weekly rolls the period’s briefs and decisions into an internal update. The reporting period can be weekly, twice-monthly, monthly or a custom interval.

Five more skills branch off the loop. ci-corroborate tests a claim against independent sources. battlecard-patch drafts a battlecard diff, or an affected-deal brief from approved guidance. battleagent answers a rep’s question from shared guidance and dated evidence, with citations, gaps and the next discovery question. ci-pattern-check counts whether one deal’s story recurs, or drafts interview questions from supplied reviews. ci-call-mentions turns Gong or tl;dv transcripts into Slack drafts or a proposed CRM tag.

The original seven skills also cover eight webinar workflows through optional modes: a call-to-competitor-tag, a call briefing, a change-to-battlecard, a competitor timeline, an internal newsletter, an affected-deal briefing, a research brief and reviews-to-discovery. battleagent adds a sourced answer to a rep’s question over the same evidence.

Visualping, where I lead marketing, is optional: every skill accepts pasted evidence.

The repo is at https://github.com/edoc33/ci-stack-skills. I shared it publicly in a webinar.

Install ci-stack-skills in Claude Code with two commands:

/plugin marketplace add edoc33/ci-stack-skills
/plugin install ci-stack@ci-stack

From a terminal, the same install is claude plugin marketplace add edoc33/ci-stack-skills, then claude plugin install ci-stack@ci-stack. For Codex, clone the repo and run python3 scripts/install_codex.py --dry-run, then python3 scripts/install_codex.py. Already installed? The repo’s first-run guide covers the 0.5.0 upgrade, including backups of local customizations.

The two keys

The first key is the agent’s draft. In the repo’s first-run example, a fictional competitor’s pricing page changes from “Team: SSO available on Enterprise only” to “Team: SSO included.” ci-triage states the exact proposition, cites the dated before and after text, and rates the evidence on three scales: evidence basis (observed, field report or inferred), verification (not tested, single-source, corroborated, contradicted or unresolved) and handling (public, internal or restricted). The brief recommends exactly one of seven responses: ignore, continue monitoring, validate, reframe, respond, match or diverge. In the 0.3.0 test run, ci-triage recommended validating the SSO claim first.

The second key is a person. Every brief starts at “Human decision: pending”, “Review status: draft” and “External-use approval: not approved”, and the shared contract forbids a skill from moving those fields or calling its own output approved. A product marketer reads the evidence, accepts or rejects the recommendation, and the brief records who decided and when. Battlecard diffs wait for a person’s review. battleagent answers stay drafts. ci-call-mentions posts to Slack only with recorded authorization.

The ci-stack contract’s evidence rules hold the agent to what the evidence shows. A skill marks a claim missing from a public page “unresolved” and may not call it “contradicted”: private pricing and negotiated terms never appear in a page diff. A dated capture proves what a page displayed and when; it says nothing about whether the feature works. One sales call counts as one person’s field report until an independent source agrees.

I built the ci-stack skills under the same two keys. Before the first six skills shipped on Aug 6, 2026, a second agent, OpenAI’s Codex, reviewed them in three rounds. One fix stopped the skills from granting their own approval.

What it produced

The $274 CI Stack was a 40-minute workshop at Product Marketing Summit, San Francisco, on Sep 10, 2026. It drew on Visualping’s H1 2026 Competitor Monitoring Report, published Jul 3, 2026, a study of 14,678 competitor monitors.

The workshop’s take-home is ci-stack-skills, a personal, MIT-licensed repo on GitHub: https://github.com/edoc33/ci-stack-skills. The seventh skill, ci-call-mentions, landed on Sep 10, 2026, the day of the workshop, in version 0.3.0. The repo’s pma-workshop-2026 branch holds the workshop pack: a fictional practice case, exercises for those seven skills, worked outputs and a visual guide to 27 source types. The pack downloads with no email or form.

Main is now version 0.5.0, reviewed October 21, 2026. It adds battleagent and keeps the original seven skills. Version 0.4.0 added the optional webinar workflow modes.

The 0.5.0 validation report records 32 package and installer tests, all passed. All eight skill entrypoints passed the skill-creator structural validator. An independent Codex agent ran the three new battleagent requests using only the skill and raw fictional inputs. The 0.3.0 report, dated Sep 10, 2026, still stands as dated evidence for the original seven skills: ten package checks passed, and independent Codex agents ran nine fictional scenarios blind to the expected answers. Both reports list what stays untested, including live connections, live delivery and model costs.

What I’d change

I’d test the CI stack on one live path before the next edition: a real Visualping monitor feeding ci-triage through the MCP server, a person approving the brief, and a delivery record for the Slack post. The 0.5.0 checks used supplied evidence only.

I’d move the CI stack’s cost table into the repo. The three list prices sit on a workshop slide today. A table with the date I checked each price would let any reader re-add the $269 and catch a price change.