Skip to content

Repository files navigation

Insight

(formerly Founder Mode)

An AI advisory board for Claude Code. Before Claude Code starts building your idea, it challenges it: 15 independent expert personas, a research engine, a creative-pivot engine, and a deterministic build-readiness score — then hands off to implementation.

Status: functional (M1–M6), published (M7). Core board, extended board, research/creative engines, the deterministic scoring server, and the Idea Evolution Engine (M9) are all implemented and exercised live (see Example sessions below).

Install

/plugin marketplace add bhouvana/Insight
/plugin install insight@insight-marketplace

Restart Claude Code when prompted to activate the bundled MCP server.

What it does

Say "I want to build X" and Insight decides whether the idea warrants a board review. If it does, it dispatches relevant personas in parallel, runs research and creative-pivot passes where useful, synthesizes everything into strengths / weaknesses / unknowns / risks / opportunities / next experiments, and produces a Build Readiness Score with a Go / Pivot / Pause recommendation — before any code gets written. A small, bounded request with no new business/market stakes ("add a dark mode toggle") skips the board and goes straight to implementation.

See PROJECT_SPEC.md for the full engine design and architecture, and .claude/skills/founder-mode/protocol.md for the exact orchestration steps.

Skills

Five skills, installed under the insight namespace. Only founder-mode triggers automatically on trigger phrases like "I want to build X" — the rest are standalone, user- or Claude-invocable, and never duplicate reasoning that belongs in their underlying agent (see ENGINEERING_PRINCIPLES.md §2.3).

Skill What it does
/insight:founder-mode Convenes the board (personas, research, scoring) to challenge an idea before implementation begins. Auto-triggers on "I want to build X" and similar — the only skill here that's Claude-invocable by default
/insight:research-engine Gathers and synthesizes competitor, market, and sentiment research for an idea via web search — never a raw search-result dump
/insight:creative-brainstorm Generates lateral pivots, cross-domain analogues, and business-model variants when an idea is cornered or underdeveloped
/insight:build-readiness-score Computes and renders the 10-dimension Build Readiness Score with a Go/Pivot/Pause call, via the real MCP server — never an LLM-guessed number
/insight:idea-evolution Refines a low-scoring idea against the board's own stated weaknesses, then re-judges it with the same unmodified board

Persona roster

Core board (always dispatched when a board is convened):

Persona Reasoning frame
founder Why-now, unfair advantage, venture-scale vs. good-business
investor Market size, moat, competition, monetization
customer Answers as the target user — pain, alternatives, willingness to switch
product-manager Scope, MVP definition, success metrics
cto-staff-engineer Technical feasibility, architecture, build-vs-buy
devils-advocate Assumes the idea is wrong, finds the weakest link, names the dead end
future-self How this looks three years out — victory-lap and eulogy retrospectives

Extended board (adaptively selected — only when a specific trigger signal is present, see protocol.md Step 1b):

Persona Trigger
ux-designer Consumer/prosumer-facing where interaction design itself drives adoption
growth-hacker A plausible viral or network-effect growth angle
marketing-lead Positioning is make-or-break (crowded category or new-category creation)
sales-lead A B2B/enterprise sales motion is implied
security-engineer Handles auth, personal data, health data, or payments
legal-compliance A regulated domain (fintech, health, privacy, minors, regulated physical goods)
operations A physical, logistics-heavy, or high-touch support business model
ai-researcher The core value proposition itself is an AI/ML capability

Engines (dispatched by trigger, not persona selection):

Engine When
research-analyst Every board session, in parallel — competitor/market/sentiment synthesis, never a raw search dump
lateral-thinker Reactively, only when devils-advocate names a real dead end (not speculatively)
idea-incubator Only via the standalone idea-evolution skill, opt-in after a Pivot/Pause score — never dispatched by founder-mode itself

Idea Evolution Engine

Given a Pivot/Pause session, idea-evolution proposes one concrete revision targeting the board's own stated weaknesses, then re-judges it with the same unmodified board — only re-dispatching the personas whose dimensions the revision actually touches (plus devils-advocate, unconditionally, every iteration). The board's calibration never loosens for a re-run: in a real test, the board found a genuinely new flaw in the "improved" idea rather than rubber-stamping it. Opt-in only (mentioned once by founder-mode on a Pivot/Pause result, never forced), checkpointed after every iteration. See .claude/skills/idea-evolution/protocol.md for the full loop.

MCP server

founder-mode-mcp-server is a local stdio MCP server (TypeScript, @modelcontextprotocol/sdk) that turns the board's qualitative synthesis into a deterministic score — no dimension is ever guessed by an LLM. It exposes four tools:

  • score_readiness — pure function: 10 dimensions (0–10 each) → overall score (0–100) + Go/Pivot/Pause band. Identical input always yields identical output.
  • save_session — scores and persists a run under a stable idea slug, appending to that idea's history.
  • load_session / list_sessions — retrieve prior runs, so re-scoring the same idea shows a delta instead of a cold start.

Session history is plain JSON under ${CLAUDE_PLUGIN_DATA}/sessions/ — no database. Scoring uses equal-weighted dimensions by default (config.ts), with go/pivot band thresholds at 75/50. Corrupt or missing session files degrade to null/empty rather than throwing (see session-store.ts). vitest run covers boundary conditions (score exactly 75, exactly 50) and corrupt-file handling; tsc --noEmit and eslint . both pass clean.

Example sessions

Real transcripts, captured during this milestone — not idealized write-ups.

1. Skip path — "add a dark mode toggle to the settings screen"

Triage (Step 1) recognizes this as a small, bounded addition to an already-established project with no new business/market stakes. No board is convened; the request goes straight to implementation. This is the fast path — most day-to-day feature requests take it.

2. Full board — "a simpler, cleaner habit-tracking app"

Idea as given: "I want to build a mobile app that helps people build daily habits — you set a habit, log it each day, and see a streak counter and simple progress charts. Think a simpler, cleaner alternative to Habitica or Streaks, aimed at people who found existing habit trackers too gamified or too plain."

Triage convened the board. Extended-board selection (Step 1b) included ux-designer (consumer app, UX-differentiated) and marketing-lead (crowded existing category); it correctly excluded growth-hacker, sales-lead, security-engineer, legal-compliance, operations, and ai-researcher — none of their trigger signals were present in the idea as stated. research-analyst ran in parallel with the board, not before it.

Build Readiness Score: 37/100 → Pause (first run on this idea slug)

Dimension Score
problemClarity 4
marketNeed 3
originality 2
differentiation 3
technicalFeasibility 8
businessPotential 3
executionComplexity 7
risk 4
defensibility 2
aiLeverage 1

Key strengths: a narrow, genuinely buildable v1 scope — converged on independently by founder, product-manager, cto-staff-engineer, and ux-designer. Legible positioning that names two specific competitors and passes the "one-sentence test" (investor, marketing-lead). Low technical risk: a local-only MVP is a complete, shippable product with no scaling wall (cto-staff-engineer).

Critical weaknesses: no moat — "simpler/cleaner" is a design-taste head start, trivially cloneable in a sprint (unanimous across founder, investor, devils-advocate, marketing-lead). The category is already saturated with near-identical "clean middle ground" positioning — Way of Life, Loop, HabitKit, Habitify, Productive, and Done were independently named by three personas and confirmed by research-analyst. Removing gamification removes one of the category's few proven retention levers with no stated replacement. The single most emotionally important UX moment — what happens when a user misses a day — is completely undefined (ux-designer, customer, independently).

Unknowns: whether the "too gamified / too plain" segment is real and findable at scale, or founder-taste generalized into a market. Platform, monetization model, and backend/sync architecture are all unstated — cto-staff-engineer flags the local-only-vs-sync decision as the single fork that determines whether this is a weekend project or a multi-month one.

Risks: Apple could absorb this exact feature set natively into Health/Reminders (investor). Category-wide churn is structural — research-analyst sourced a 92%-of-attempts-fail-within-60-days stat — not fixable by a nicer UI. Low switching cost cuts both ways: cheap to acquire from competitors, equally cheap to lose to the next clone.

Opportunities: future-self's concrete mechanism — a streak that dims but doesn't reset, with a 48-hour repair window — as the actual differentiator instead of the color palette. Reactively dispatched after devils-advocate named a real dead end (not "none found"), lateral-thinker (Creative Engine) proposed: re-target to a narrow real population (sobriety, medication adherence, PT recovery) instead of an aesthetic segment; remove the logging step via passive/sensor-based detection; add commitment-contract stakes instead of polish; license the engine B2B2C to employers/insurers instead of competing for App Store search. The one idea worth taking seriously even though it sounds crazy: pivot from a consumer app to a verification/trust layer — passive-sensor-confirmed adherence sold to institutions that need to trust the data, which self-reported competitors structurally can't offer.

Recommended next experiments: mine App Store reviews of Streaks/Habitica/Loop for the literal "too gamified"/"too plain" complaint pattern before writing code (afternoon) → 10–15 mom-test interviews with current/recently-churned users (a week) → prototype the streak-repair mechanic against a hard reset in a 2-week logging cohort (a week) → if pursuing the vertical pivot, validate willingness-to-use in one narrow community before building general-purpose (a month).

Confidence: Low — the entire premise rests on an unvalidated assumption (a real, sizable "too gamified/too plain" segment) that multiple independent personas flagged as founder-taste rather than evidence, and research found a disconfirming stat, not a confirming one.

Update

/plugin update insight@insight-marketplace

Contributing

See CLAUDE.md, PROJECT_SPEC.md, ENGINEERING_PRINCIPLES.md, and IMPLEMENTATION_PLAN.md.

About

An AI advisory board for Claude Code that challenges, researches, and strengthens ideas before a single line of code is written.

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages