Skip to content
Back to skills

Pricing Research

ASecurity

Process reproductible de recherche pricing pour tout produit/service ElPi Corp. 4 axes parallèles avec gate approval competitors. Subagent B = 2 phases (candidates list → approval Laurent → deep scrape Firecrawl). Output multi-format pour comparison table landing + articles blog "Notre app vs competitor".

  • 3 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 29, 2026
ai-agents

Works with

  • mcp

Security analysis

A100/100

Scanned September 29, 2026

npx -y skills add vantageos-agency/vantage-peers --skill pricing-research --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Pricing Research?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Pricing Research
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vantageos-agency-pricing-research/badge)](https://www.skillsdirectory.com/skills/vantageos-agency-pricing-research)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: pricing-research
description: Process reproductible de recherche pricing pour tout produit/service ElPi Corp. 4 axes parallèles avec gate approval competitors. Subagent B = 2 phases (candidates list → approval Laurent → deep scrape Firecrawl). Output multi-format pour comparison table landing + articles blog "Notre app vs competitor".
metadata:
  version: "1.1.0"
license: Proprietary
---

# pricing-research v1.1

Process reproductible recherche pricing avec 4 axes parallèles + gate approval competitors.

## WORKFLOW

### Step 1 — Capture contexte produit
Créer `decisions/pricing-research/<product-slug>/_context.md`.

### Step 2 — Dispatch 4 subagents parallèle (strategy-researcher Sonnet bg)

**Subagent A — Personas** : 3-5 personas réelles avec douleurs + budget. Sources web obligatoires.

**Subagent B — Competitive (2 phases avec gate)** :
- Phase B1 : 8-15 candidats (nom + URL + positionnement). Output `competitive-candidates.md`.
- GATE APPROVAL Laurent : send_message Pi+Laurent avec liste, demande APPROVE/REVISE.
- Phase B2 (post-approval) : Firecrawl scrape pricing/features/comparison pages. Extraire features list + tiers + pricing per tier + matrice features×tier. Output multi-format : `competitive.md` (full) + `competitive-comparison-table.md` (landing) + `competitive-blog-data.json` (articles "Notre app vs X").

**Subagent C — Psychological pricing** : seuils B2B SaaS marché cible 2026, études récentes citées.

**Subagent D — Anchor pricing** : 8-12 produits SaaS comparables avec pricing 2026 exact.

### Step 3 — Gate competitors + audit
3a. Phase B1 termine en premier, bloque deep scrape jusqu'à approval Laurent.
3b. Orchestrateur send_message avec candidats list pour approval.
3c. Post-approval, dispatch Phase B2.
3d. A/C/D continuent en parallèle pendant le gate.
3e. Audit : tous les fichiers existent et substantiels (sources URLs réelles).

### Step 4 — Synthèse orchestrateur
`synthesis.md` avec personas retenues + concurrence positioning + seuils psy + anchors mentaux + recommandation pricing sourcée + plan B.

### Step 5 — Briefing note VantagePeers
`mcp__vantage-peers__create_briefing_note` topic=product avec content synthesis + decisions list.

### Step 6 — Output user-facing
Recommandation finale + plan B + artifacts paths + briefing-id, décision attendue APPROVE/REVISE/ARCHIVE.

## RULES

- 4 subagents toujours en parallèle
- Subagent B = 2 phases avec gate approval Laurent (pas de deep scrape sans approval)
- Sources obligatoires (pas de chiffres inventés)
- Firecrawl pour deep scrape concurrents (mcp__firecrawl__firecrawl_scrape format markdown)
- Output competitive multi-format : full data + comparison table + blog data JSON
- Artifacts persistants commit/push
- 1 briefing note finale systématique
- Recommandation = 1 principal + 1 plan B
- Process identique tous produits

## EXAMPLES

### Example 1 — VantagePeers Cloud pricing
/pricing-research vantagepeers-cloud → 4 subagents → gate B candidates → Laurent approve → Phase B2 deep scrape → synthesis → briefing → recommandation finale

### Example 2 — GPTPowerUps Cloud pricing
/pricing-research gptpowerups-cloud → même process, dossier dédié, briefing distinct

## SELLABLE AS

Skill core du plugin perello-pricing-research — process reproductible recherche pricing pour tout produit B2B SaaS / formations / prestations. v1.1 ajoute gate approval competitors + Firecrawl deep scrape + multi-format output (landing comparison + blog articles data).

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.