🧢 Mitchell Agentic Sprint

A Claude Code plugin that walks an AI-strong builder through six sequential steps to take one of their many ideas and prove out whether it can become a real business — adversarial by default, refusing flattery, demanding verbatim evidence.

By the end of the Sprint, the builder has a locked Vertical Stack, 15–30 buyer interviews captured verbatim, a Mom Test–disciplined theme map, an adapted expert framework, a competitor cluster map with positioning whitespace, and three markdown artefacts ready to take into customer conversations and pre-seed pitches: a sales deck, a customer outreach plan, and a pre-seed investor deck.

The Sprint does not build the product or the website. It gets the builder to the point where they know whether what they’re building will sell, and they have the materials to prove it.

Concept

Authored by Scott Mitchell (Mitchell Agentic Partners). Working name.

→ Full concept: docs/concept/concept.md

Status

v0.3.2 — Claude Code + Cowork plugin testing. Ironed out a few bugs, trying out the workflow to see if it works end to end!

Design principles

Adversarial by default — every prompt refuses flattery and pushes back on vague answers (concept §6)
Behavioural over hypothetical — interview scripts pass through a Mom Test linter that bans “would you…” questions before the user ever sees them (SPEC-008-A)
Saturation over headcount — interviews stop when the last 3 surface no new themes, not at an arbitrary count (helpers/saturation.py)
Loopback on validation failure — failed verticals retain Builder Profile, themes, and framework into the next attempt; Vertical Stack and downstream are cleared (concept §7, SPEC-005-A)
Mini Council — five skeptical NPC voices (Seed VC, Churned Customer, Competitor Founder, Future-Self, Grumpy PM) tear up the idea between Steps 1→2 and 4→5
Output is artefacts, not product — three markdown artefacts populated from sprint state, no padding when evidence is missing (ADR-009)

Quick start

Claude Cowork (recommended for MAS)

The Sprint is well-suited to Cowork. A six-step validation pipeline runs across multiple sessions over days or weeks, with persistent state at each gate and asynchronous artefact production — that’s the shape Cowork is built for. Use Cowork unless you have a reason to prefer the CLI.

From the Cowork tab in Claude Desktop, select + → Plugins → Add plugin
Add marketplace — Select the By Anthropic dropdown, then Add marketplace from GitHub and enter:
```
https://github.com/cgbarlow/skills
```
Install three plugins from the marketplace: Six Animals, Campaign Mode, and Mitchell Agentic Sprint (the Sprint depends on the other two; per ADR-003)
Begin your Sprint — run:
```
/sprint-start
```
On the first run in a project, /sprint-start auto-invokes setup: prerequisites are verified and the 🧢 Mitchell profile is offered (default install). The Sprint runs in Claude-only synthesis mode — fully featured for Cowork users. (The optional NotebookLM bolt-on for source-grounded synthesis is Claude Code CLI only; see below for that path.) Then Step 1 (Discovery) opens with Mitchell framing your unfair advantages, runway, and rough idea.

To re-configure later (install Mitchell after declining, onboard NotebookLM after the fact), call /sprint-setup directly.

Claude Code CLI

For terminal-first users:

Install Claude Code (full guide):

curl -fsSL https://claude.ai/install.sh | bash

Add the marketplace — in a Claude Code session:
```
/plugin marketplace add cgbarlow/skills
```

Install the three plugins:

/plugin install six-animals@cgbarlow-skills
/plugin install campaign-mode@cgbarlow-skills
/plugin install mitchell-agentic-sprint@cgbarlow-skills
/reload-plugins

Begin your Sprint:
```
/sprint-start
```
First run auto-invokes setup (prerequisites, Mitchell profile, NotebookLM offer). Re-configure later via /sprint-setup.

Pinned compatibility: campaign-mode@0.4.8, six-animals@0.1.2 (verified 2026-05-08). Or clone the repos into your .claude/plugins/ directory if you’d rather skip the marketplace entirely.

Optional (Claude Code CLI only): NotebookLM MCP bolt-on

The NotebookLM bolt-on works in Claude Code CLI only — it does NOT work in Claude Cowork. This is a platform constraint, not a Sprint limitation: nlm setup add claude-code writes to Claude Code (CLI)’s user-scope config (~/.claude/...), not Claude Desktop’s claude_desktop_config.json. And per Anthropic’s Cowork + local MCP support article, Cowork only bridges remote MCP servers (custom connectors); it doesn’t pick up local stdio MCP servers from claude_desktop_config.json either. So even manual config-editing doesn’t help Cowork users.

Cowork users: skip this entirely. Claude-only synthesis works for the entire Sprint. You miss the NotebookLM audio overview in Step 6, but the three written artefacts (sales deck, outreach plan, investor deck) are produced equally well.

Claude Code CLI users: opt in if you want it. Setup runs on your local machine:

uv tool install "notebooklm-mcp-cli>=0.6.6"   # uv handles Python ≥3.11 automatically
# or: pip install "notebooklm-mcp-cli>=0.6.6" (needs Python ≥3.11 already)
# or: pipx install notebooklm-mcp-cli

nlm login                      # browser-based; one-time
nlm setup add claude-code      # wires up MCP for Claude Code (CLI only)
nlm doctor                     # verify

After install, restart Claude Code. Subsequent sessions pick up the saved login. The Sprint reads .sprint/setup.md to decide whether to route synthesis through NotebookLM or fall back to Claude-only.

See ADR-006 and SPEC-006-A for the full integration contract.

Commands

Command	Step	What it does
`/sprint-start`	1	The single entry point. First run auto-invokes setup (verify prereqs → install 🧢 Mitchell profile → offer NotebookLM bolt-on); subsequent runs go straight to Step 1 Discovery framed by Mitchell.
`/sprint-setup`	re-config	Re-run only when you want to change setup choices (e.g. install Mitchell after declining, onboard NotebookLM after the fact). Auto-invoked by `/sprint-start` on first run.
`/sprint-step-2`	2	Vertical Stack — narrow to vertical × sub-vertical × ICP × workflow × AI leverage
`/sprint-step-3`	3	Buyer Interviews — Mom Test–linted scripts; saturation detection
`/sprint-step-4`	4	Framework — find the named expert, adapt their framework
`/sprint-step-5`	5	Competitor + Positioning Whitespace
`/sprint-step-6`	6	Synthesis — populate the three markdown artefacts
`/sprint-council`	between steps	Convene the five Mini Council voices in sequence
`/sprint-loopback`	any time	Reset to Step 2 retaining profile + themes + framework
`/sprint-continue`	re-entry	Read sprint state, route to next action

Mini Council voices

Five new SKILL.md NPC agents (per ADR-004 and SPEC-004-A) — each with embedded domain heuristics, Dragon-style adversarial tone, and a three-verdict output (Pass / Pass for now / In play):

Voice	Lens
💼 Seed VC	2026-AI-era pre-seed: real moats, inference economics, working demo + evals + buyer-tape as table stakes
💔 Churned Customer	post-purchase regret — onboarding-promise audit, “I’d cancel because…”
⚔️ Competitor Founder	rival operator — where I’d attack you; what I’d copy / what I’d dismiss
⏳ Future-Self (12 mo)	regret enumeration; “the version of you that succeeded did this one thing”
📋 Grumpy PM	operational reality — realistic timeline; “what gets dropped on Tuesday”

The Council fires meaningfully at the Step 1→2 and Step 4→5 transitions.

Mentor profile — 🧢 Mitchell

The Sprint’s framing mentor is campaign-mode’s Gandalf NPC — the voice that frames Discovery (Step 1), the Vertical Stack (Step 2), and the expert-framework search (Step 4). MAS ships a profile pack that re-skins Gandalf as 🧢 Mitchell — Scott Mitchell as the founder-mentor persona, asking the questions he’d ask in a real conversation.

/sprint-setup offers to install the profile by default. Decline if you want the default 🧙 Gandalf appearance, or remove .campaign/profiles/gandalf.md later. The profile pack itself ships at profile-packs/mitchell-agentic-sprint/gandalf.md and follows campaign-mode’s SPEC-CM-006-A profile format.

State

Sprint state lives at .sprint/sprint.md per SPEC-005-A — frontmatter (sprint mode, current step, last approved step, attempt counter) plus named markdown sections per step. The schema accommodates loopback semantics: Builder Profile / Backup Verticals / Interview Log / Theme Map / Costly-Action Signals / Expert Framework / Loopback Log / Progress Log are retained on loopback; Vertical Stack / Competitor Map / Positioning Whitespace / Artifacts are cleared.

Interview transcripts are stored verbatim per concept §6 at .sprint/interviews/{NN}-{pseudonym}.md.

Generated artefacts land at .sprint/artifacts/:

sales-deck.md
outreach-plan.md
investor-deck.md
audio-overview.mp3 (NotebookLM bolt-on only)

Architecture decisions

All twelve ADRs follow the enhanced WH(Y) format (per ADR-001, ported from cgbarlow/iris):

ADR	Title
001	Adopt Enhanced WH(Y) ADR Format
002	Build Tier — Moderate + NotebookLM bolt-on
003	Sibling-Plugin Shape with hard `campaign-mode` + `six-animals` deps
004	Mini Council as five new SKILL.md NPC agents
005	`.sprint/sprint.md` State Schema with Loopback
006	NotebookLM via MCP — non-Enterprise, optional, Claude-only fallback
007	Per-Step Gate Enforcement via `last_approved_step`
008	Anti-Sycophancy via Dragon-style preambles + behavioural-interview linter
009	Markdown-First Artifacts; defer slide rendering to v0.2.0
010	TDD Scope — Code Only; Council Voices via Worked-Example Review
011	Single `feature/v0.1.0` Branch with Gated PR Review
012	Python (with pytest) for Helpers

Specs sit alongside in docs/adrs/specs/.

Worked examples

Three reference Sprints in docs/examples/:

Worked example 1 — Vertical SaaS — Step 1 + Mini Council voice differentiation reference
Worked example 2 — Prosumer with loopback — Step 3 saturation with zero costly-action signals → loopback to a different ICP
Worked example 3 — Full Sprint with prior loopback — full Step 1→6 walkthrough including all three populated artefacts

Development

Helpers are Python ≥3.11 with pytest. Per ADR-010 every helper module has unit tests; Mini Council voice fidelity is verified by manual worked-example review at gate reviews.

python -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest

Helpers live in helpers/:

sprint_state.py — read/write .sprint/sprint.md, gate enforcement, loopback semantics, atomic writes
interview_validator.py — Mom Test linter (12 regex rules across 7 categories) + python -m CLI
saturation.py — interview-theme saturation detector + python -m CLI
notebooklm.py — pure-function prompt builders for the four NotebookLM integration points + is_mcp_available()

Development protocols are vendored from cgbarlow/protocols at docs/protocols.md. Protocol 13 ({@html} security) does not apply — MAS is a Claude Code plugin, not a Svelte project.

Research

The build was preceded by structured research published under docs/research/:

Research plan
Gap analysis — what MAS needs that neither campaign-mode nor Six Animals provides
Build options — Minimum / Moderate / Extensive trade-offs
Recommendation — the chosen path (Moderate + NotebookLM bolt-on)

License

CC-BY-SA-4.0

Authors

Scott Mitchell — concept and product direction
Chris Barlow — implementation