Mitchell Agentic Sprint

AI-led 6-step sprint that walks an AI builder through validating one idea — vertical, buyers, framework, positioning — to a sales deck, outreach plan, and pre-seed investor deck. Adversarial by default. Output is artifacts, not product.

View project on GitHub

Recommendation — Build Approach for Mitchell Agentic Sprint

Phase F output of the research plan. The deliverable. Selects one of the three approaches in build-options.md, drawing on the gap analysis.

TL;DR

Build the Moderate option as a Claude Code plugin that depends on campaign-mode and six-animals. Ship a Minimum-shaped walking skeleton in week 1 and put it in front of three friendly testers before committing to the remaining 6–7 weeks.

Estimated total: 7–8 weeks for one engineer, with a hard go/no-go decision after week 1.

Add the NotebookLM MCP integration as an optional bolt-on (recommended) — adds ~1.5 weeks to v1, replaces ~3 weeks of Extensive-tier bespoke synthesis tooling, and aligns with MAS’s “verbatim quotes only” rule. Decision can be deferred to week 5 (Steps 3–4 implementation) without restructuring earlier work. See build-options.md §4.

The choice

Recommended: Option 2 — Moderate.

Option Verdict Reason
Minimum Reject as final shape Cannot deliver the concept’s headline promise (anti-sycophancy + investor-conversation-ready artifacts). Council voices will feel interchangeable; Mom Test discipline cannot be enforced; investor decks will lack defensible evidence. However, Minimum’s first-week deliverables are reused as Moderate’s walking skeleton — see “Phasing” below.
Moderate Select Hits the headline promise (real domain-knowledgeable Council voices, enforced Mom Test, working loopback) without front-loading tooling that may not be the bottleneck. Closes Q1–Q4 from the gap analysis with high confidence. Q5 (anti-sycophancy verification via eval) is deferred to Extensive — accepted risk.
Extensive Defer Premature. The concept itself flags core uncertainties (linear flow, named-expert assumption, whether AI can replace a coach). Building tooling for 12–16 weeks before validating those assumptions inverts cost-of-no. Tooling layer becomes justified only if Moderate ships and tester demand pulls it forward.

Phasing — week 1 walking skeleton

The first week of Moderate is shaped to be Minimum-equivalent: a runnable Step 1 + a single Council invocation, end-to-end. The point is to get something in front of testers fast.

Week Output Decision gate
1 Plugin scaffolds. /sprint-start runs Step 1 with Gandalf-as-skin. /sprint-council invokes one Mini Council voice (Seed VC, drafted as the canonical voice). State written to .sprint/sprint.md. Three friendly testers run through Step 1 + Council. Go/no-go on Moderate. If testers report the Seed VC voice has real domain depth and the Step 1 framing is useful, continue. If not, revisit.
2–3 Remaining 4 Council voices drafted. /sprint-council orchestrates all five. Step 2 (Vertical) command. Loopback command. Continue if Council voices are differentiable.
4–5 Steps 3 (Buyer Interviews) + 4 (Framework). Banned-question linter. Saturation detector. Continue if linter false-positive rate is acceptable.
6–7 Steps 5 (Competitor) + 6 (Synthesis). Three artifact templates. Three worked examples. Tester debriefs. Ship.
8 Buffer / polish / docs. Tag v0.1.0.

If the week-1 gate fails, fallback options: (a) prompt-iterate on the Seed VC voice for another week before continuing; (b) downgrade scope to Minimum and accept the headline-promise gap; (c) escalate findings to Scott and reframe the product.

Why Moderate is the right pick — five points

  1. The headline product promise is anti-sycophancy (concept §6, §10). That requires the Council voices to actually push back with domain credibility. Minimum can’t get there with profile-pack skinning alone (gap analysis §2). Moderate’s 5 new SKILL.md files can.
  2. Mom Test discipline is the difference between “buyer interviews” and “founder reassurance theatre” (concept §4.3). Minimum has no enforcement layer. Moderate’s banned-question linter + saturation detector make the discipline operational.
  3. Loopback is the cost-of-no mechanism that prevents the concept’s worst failure mode — builders shipping doomed products because admitting wrongness is too expensive (concept §7). Minimum has no loopback handler. Moderate ships one.
  4. Extensive’s tooling layer is solving downstream problems (slide rendering, theme clustering, evals) that aren’t on the critical path until users specifically demand them. We add them when the demand signal arrives, not speculatively.
  5. Campaign-mode is genuinely well-shaped to host this, per the gap analysis. ~70% of the infrastructure is reusable as-is (transcripts, progress logging, AskUserQuestion conventions, profile system, six-animal advisory layer). We’re extending, not fighting.

Answers to the seven research questions

Q Question Answer Where to read
Q1 Sequence fit Achievable through new state schema + new gate command, not pure rewrite. Moderate scope. gap-analysis §1
Q2 Mini Council depth Profile packs insufficient. Need 5 new SKILL.md files with embedded domain heuristics. gap-analysis §2
Q3 Artifact reach Markdown-first closes 80% of value. Slide rendering is optional polish, defer to Extensive. gap-analysis §7
Q4 Loopback semantics New schema + 1 new command. Achievable in Moderate. gap-analysis §4–5
Q5 Anti-sycophancy enforcement Tone via prompt preambles (free, lifted from Dragon). Verification via eval set is Extensive-tier — accepted as residual risk for Moderate v1. gap-analysis §6
Q6 Distribution Open — assumes Claude Code plugin for Moderate. If standalone web/desktop is required, Extensive is the floor and recommendation changes. open question
Q7 Ownership Open — assumes one engineer, presumed to be Chris. Affects time budget but not architecture. open question

NotebookLM addendum

Per the issue #1 follow-up, Google NotebookLM is a serious option for synthesis work. As of May 2026, viable access paths exist via official Google Cloud Enterprise API, unofficial CLI (tmc/nlm), and several MCP servers actively maintained for Claude Code. See build-options.md §4 for the full evaluation.

Recommended adoption pattern:

  • Make NotebookLM optional, declared as an MCP dependency in plugin.json. MAS works without it; works better with it.
  • Defer the integration decision to week 5 (Steps 3–4 implementation) — the architecture supports adding it without restructuring earlier weeks.
  • Pick the unified nlm + notebooklm-mcp package as primary path (CLI + MCP merged Jan 2026). Maintain a Claude-only fallback for users who can’t or won’t auth to Google.
  • Use NotebookLM for: theme map + saturation analysis (Step 3), expert framework synthesis (Step 4), positioning whitespace (Step 5), audio overview as a v2 artifact (Step 6).
  • Do not use NotebookLM for: Mini Council adversarial logic, Mom Test enforcement, slide rendering — these remain MAS’s job.

Cost: ~1.5 weeks added to Moderate (or ~3 weeks saved vs full bespoke Extensive tooling).

Dependencies: Google account per user; MCP server install at user’s machine; initial cookie/OAuth handshake during /sprint-setup.

Risk to track: if Google blocks unofficial automation, fall back to Enterprise API for org-deployed users and Claude-only synthesis for individual users. Don’t make NotebookLM load-bearing.

Next actions (concrete)

  1. User answers the seven open questions (below) — blocks everything downstream. ETA: 1 conversation.
  2. Spin up plugin scaffolding in this repo: .claude-plugin/plugin.json, commands/, skills/, docs/templates/, .sprint/ ignored. ETA: half a day.
  3. Draft seed-vc-agent/SKILL.md as the canonical Council voice. Includes: 2026-AI-era investor checklist, moat taxonomy, traction expectations, voice/tone borrowed from Dragon. This file becomes the template for the other four. ETA: 2 days.
  4. Build the Step 1 + Council walking skeleton end-to-end. Tester-facing. ETA: 4 days, ending by end of week 1.
  5. Recruit three friendly testers from Chris’s immediate network. Send them through Step 1 + Council. Run debrief interviews using Mom Test discipline. ETA: ongoing through week 2.
  6. NotebookLM MCP spike (week 5) — add nlm/notebooklm-mcp as optional dependency; wire Step 3 close-out to invoke notebook creation + theme-map query. ETA: 1–2 days. Skippable if user declines NotebookLM bolt-on (Q8).

Residual open questions for the user

These are open in the research plan and remain unanswered. Each shapes execution:

  1. Distribution channel (Q6) — Claude Code plugin? Claude Desktop plugin? Cursor? Standalone? Recommendation assumes Claude Code plugin.
  2. Spike or paper-only? — Recommendation collapses this question by making the spike the week-1 walking skeleton. The week-1 deliverable is the spike; if it satisfies, we continue to full Moderate.
  3. Time budget — Is 7–8 weeks acceptable? Is there a hard date? (Add ~1.5 weeks if NotebookLM bolt-on selected.)
  4. Coupling tolerance — Hard campaign-mode dependency OK, or does MAS need to stand alone? Recommendation assumes hard dependency.
  5. v1 user persona — Immediate AI-builder network or general-public AI-builders? Recommendation assumes immediate network for v1, raises bar for v2.
  6. Scott’s involvement — High-fidelity loop (Scott reviews artifacts as we go) or low-fidelity (Scott waits for v1)? Recommendation assumes high-fidelity — Scott reviews after week 1 walking skeleton, before committing remaining 6 weeks.
  7. NotebookLM bolt-on? — Recommended yes (optional MCP dependency, +1.5 weeks). Are testers willing to authenticate to Google? Is vendor lock-in acceptable for v1?

What success looks like 90 days from now

  • v0.1.0 tagged on mitchell-agentic-sprint repo
  • Three completed Sprints by friendly testers, with all six steps and at least one loopback observed
  • Three artifact sets (sales deck + outreach plan + investor deck) shared with Scott for review
  • A documented assessment of whether the Mini Council voices feel adversarial-with-credibility or just adversarial
  • A short list of Extensive-tier capabilities that user feedback has pulled forward (e.g., “everyone asked for PPTX → add slide rendering”). Without that signal, no Extensive work happens.

What success does not look like

  • A polished plugin that no one has used end-to-end
  • Slide rendering before three testers have completed Step 6
  • An anti-sycophancy eval set built before we know what real users find sycophantic
  • A second product extension (e.g., a CRM integration, a buyer-database lookup) before v0.1.0 ships

Confidence and what could change this recommendation

High confidence:

  • Campaign-mode + Six Animals are well-shaped hosts (Q1–Q4)
  • Markdown-first artifacts cover the headline promise (Q3)
  • Moderate is the right scope for v1 (Q1–Q4)

Medium confidence:

  • The 5 Council voices will feel sufficiently differentiated after the first prompt-iteration round (Q2). Tested in week 1.
  • Mom Test linter false-positive rate will be acceptable (gap analysis §3). Tested in week 4–5.

Low confidence:

  • Anti-sycophancy holds under sustained pressure from real users without an eval set (Q5). Accepted risk for v1; revisit for v2.
  • Distribution as Claude Code plugin is the right channel for the v1 audience (Q6). Open question.
  • NotebookLM MCP servers remain stable across Google’s UI updates. Mitigated by the maintainer track record (April 2026 UI refresh shipped fixes within 24–72 hours per BrightCoding writeup) but residual risk remains.

This recommendation is reversible until week-1 gate. After that, the architectural commitments (state schema, command structure) are difficult to back out.