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

Gap Analysis — Mitchell Agentic Sprint vs Campaign-Mode + Six Animals

Phase C output of the research plan. Maps MAS concept requirements against existing capability in cgbarlow/campaign-mode and SimonMcCallum/six-animals and identifies what must be built new.

Method

Each MAS requirement is sourced from docs/concept/concept.md. For each, three columns:

  • Existing capability — what campaign-mode or Six Animals gives us today, with file references
  • Gap — what’s missing or insufficient
  • How to close — the concrete work, classified as prompt-only (re-use existing agents with new copy), new SKILL.md (new agent file), or new code/state (changes to commands or state schema)

1. Linear sequence enforcement (Q1)

Aspect Existing capability Gap How to close
Phase counter .campaign/quest.md frontmatter has phase: 1-6; updated by command logic Phase 1–6 in campaign-mode are quest-themed (Quest Definition, Character Setup, Execution, Guardian, Dragon, Debrief) and not strictly ordered (Council is optional and can fire mid-quest) Replace phase model with step: 1-6 named steps (Discovery, Vertical, Buyer Interviews, Framework, Competitor, Synthesis). New code/state.
Step completion gates Guardian agent (skills/guardian-agent/SKILL.md) provides “approve / not yet / conditional” verdicts at checkpoints Guardian’s checkpoints are user-invoked at any time, not enforced as a precondition for advancing. MAS needs hard gates per step. Wrap step transitions in a command that runs Guardian as a precondition. New code (one new command + state field last_approved_step).
Skip protection None — campaign-mode trusts the user to advance phases sensibly A user could run /sprint-step-3 without completing 1 and 2 Command guards check last_approved_step ≥ N-1 before entering step N. New code.
Anchor to artifacts quest.md has free-form Progress Log MAS needs each step to declare what artifact must exist before next step (e.g., Step 2 outputs a Vertical Stack; Step 3 needs ≥15 interviews on file) Per-step required-artifact list in state schema. New state.

Verdict on Q1: Cannot be done by command rewrites alone. Requires new state fields and a new gate-checking command. ≈ 1 week of work.

2. Mini Council personas (Q2)

The five MAS Mini Council voices and their nearest existing analogues:

MAS voice Closest existing Why partial What’s missing
Seed VC Cat (risk) + Bear (strategy) Cat flags risk; Bear articulates competitive position Unit-economics heuristics, traction expectations, moat taxonomy (“workflow lock-in vs data flywheel vs proprietary AI”), 2026 AI-era investor lens (working demo, evals, inference economics — all called out in concept §6)
Churned customer Cat (caution) + Puppy inverse Cat warns of failure modes Retrospective regret framing — “I paid and got burned” — is post-purchase narrative, not pre-purchase risk
Competitor founder Cat + Bear Cat maps threats; Bear positions Market knowledge, competitive intelligence, knowledge of gaps in current offerings
Future-self-12-months Owl (process) + Cat (unmitigated risks) Owl tracks what was skipped Temporal hindsight — “what did you regret not doing?” is a different orientation than present-tense risk assessment
Grumpy PM Owl + Rabbit Owl tracks timeline; Rabbit maps resources Pessimistic execution voice — Owl/Rabbit are optimistic about clearing barriers; the Grumpy PM assumes barriers multiply

Verdict on Q2: Profile packs cannot close this gap — profile packs only re-skin name/voice/emoji and cannot add domain heuristics or change evaluation logic (see campaign-mode/docs/3_specs/SPEC-CM-009-A). Each Council voice needs its own SKILL.md with embedded heuristics. 5 new SKILL.md files. Each ≈ 1–2 days to draft + iterate against worked examples.

3. Step-specific intelligence (not in any existing framework)

Each of the 6 steps requires methodology that neither campaign-mode nor Six Animals encodes:

Step Methodology required Source Existing?
1. Discovery Founder profiling — unfair advantages, runway, vertical fit Concept §4.1 Generic mentor (Gandalf) provides framing only — no founder-fit heuristics
2. Vertical Vertical x sub-vertical x ICP x workflow x AI leverage stack Concept §4.2 None
3. Buyer interviews Mom Test discipline (banned hypothetical questions, costly-action signal detection, saturation rule) Concept §4.3, §6 None — new validator needed
4. Framework Named-expert discovery, framework adaptation (3 keep verbatim / 3 adapt) Concept §4.4 None — needs WebSearch + synthesis
5. Competitor Top-10 competitor analysis, positioning whitespace map Concept §4.5 None — needs WebSearch + scraping or user-provided URLs
6. Synthesis Three-artifact generation (sales deck, outreach plan, investor deck) Concept §4.6 None

Verdict: This is the bulk of the new work. Roughly:

  • Steps 1, 2 — prompt-only (new commands + Gandalf consultation pattern)
  • Step 3 — needs a behavioral-interview validator (banned-question linter); prompt-only for the script generator
  • Steps 4, 5 — need either WebSearch tooling or structured user-provided inputs; can start prompt-only with user-provided inputs and add tooling later
  • Step 6 — prompt-only for markdown drafts; new code for slide rendering (PPTX/PDF)

4. State schema for named-step artifacts (Q4)

Current .campaign/quest.md schema:

---
campaign-mode: Grow | Ship | Grow & Ship
phase: 1-6
created: YYYY-MM-DD
---
## Quest Narrative
## Success Criteria
## Anticipated Dragon
## Progress Log

MAS needs a multi-section schema that survives loopback (see §5). Proposed:

---
sprint-mode: Solo | Coached
step: 1-6
last_approved_step: 0-6
attempt: 1-N        # increments on loopback
created: YYYY-MM-DD
---
## Builder Profile           # Step 1, persists across attempts
## Vertical Stack            # Step 2, replaced on loopback
### Backup Verticals
## Interview Log             # Step 3, transcripts at .sprint/interviews/
### Theme Map
### Costly-Action Signals
## Expert Framework          # Step 4, persists across attempts
## Competitor Map            # Step 5
### Positioning Whitespace
## Artifacts                 # Step 6
### Sales Deck path
### Outreach Plan path
### Investor Deck path
## Loopback Log              # records each attempt's vertical + reason for loopback
## Progress Log              # reused from campaign-mode

Verdict on Q4: Schema must be rewritten — quest.md is too thin. Loopback is supportable if and only if Builder Profile and Expert Framework persist across attempts (the concept explicitly says these are reusable; only Vertical Stack and downstream is replaced). New state schema doc + a small migration helper. ≈ 2 days.

5. Loopback semantics (Q4 cont’d)

Concept requirement (§7): on validation failure (no costly action found, scorecard < 65, no whitespace), loop back to Step 1 retaining user profile, interview themes, and framework knowledge — only the disqualified vertical and downstream is discarded.

Aspect Existing capability Gap How to close
“Not yet” handling Guardian’s “not yet” verdict comes with feedback and lets user iterate within the same step MAS loopback is cross-step — fail Step 3, restart Step 2 with new vertical New command /sprint-loopback that resets last_approved_step to 1 and clears Step 2+ sections while preserving Step 1 + accumulated themes
Three loopback flavours None Concept names three: different ICP same vertical / different vertical adjacent skills / different workflow same vertical Each is a different reset depth — encode in the loopback command
Cost-of-no Concept emphasizes “make a no cheap” Campaign-mode’s Dragon Prevails verdict is final-stage only Per-step early-no handling — interview Saturation detection in Step 3 should fire loopback automatically when no costly-action signals after N interviews

Verdict on Q4: Loopback is genuinely new logic. Roughly 1 week of work including the saturation detector for Step 3.

6. Anti-sycophancy enforcement (Q5)

Concept §6 makes this the headline product promise. Two questions: (a) where does it live, (b) how do we prove it’s working.

Layer Existing capability Gap How to close
Voice/tone Dragon already enforces “no encouragement, no warmth, no hedging” (skills/dragon-agent/SKILL.md lines 88–96) Dragon is one agent at the end of a quest. MAS needs adversarial posture in every step prompt Reuse Dragon’s tone guidance verbatim in each new SKILL.md and step command preamble. Prompt-only.
Behavioural enforcement None Mom Test discipline (banned hypothetical questions in interview script) Banned-question linter — regex/LLM check on generated interview scripts. Small new code.
Verification None “Does the model actually push back, or does it flatter under pressure?” Eval set: 20–30 prompts where the user is wrong, and the expected response is pushback. Run before each release. New work, classified as test infrastructure.
Council adversarial fidelity Dragon evaluates work product; Guardian gates progression Mini Council needs to attack the user’s reasoning at Step 1→2 and Step 4→5 boundaries New Council orchestration: invoke each of the 5 voices in sequence, each receiving the same context, each producing one strongest critique. Aggregate, present to user. New code (one new command).

Verdict on Q5: Tone is prompt-only via Dragon-style preambles. Verification requires an eval set — that’s an Extensive-tier deliverable, not Moderate-tier. The Mini Council orchestration command is moderate-tier, ≈ 3 days of work after the 5 voices exist.

7. Final artifact generation (Q3)

Three Step 6 outputs:

Artifact Prompt-only feasibility When tooling becomes necessary
Sales deck (6–8 slides) High — markdown can carry slide-per-section structure with headlines + speaker notes; user pastes into their own slide tool If user wants a .pptx or .pdf directly. PPTX generation via python-pptx is well-trodden.
Customer outreach plan High — markdown sections (re-engagement messages, target list, sequence). The hard part is retrieving who to re-engage; that’s solved by the Interview Log persisting across attempts None until we want CRM integration
Pre-seed investor deck (12 slides) Medium — the structure is prompt-only; the evidence (interview tape, evals, inference economics) requires real inputs from prior steps. If those inputs are missing, the prompt should refuse to pad rather than fabricate Slide rendering as above. Also: VC-grade decks usually need a designer pass — out of scope.

Verdict on Q3: Markdown-first artifacts close 80% of the value. Slide rendering is genuinely optional for v1 — it’s polish, not core. Tooling is not on the critical path until users specifically demand it.

7a. NotebookLM as a synthesis layer (Q3 augmented)

Per Issue #1 comment, evaluating Google NotebookLM as an artifact / synthesis layer. Findings as of May 2026:

Access path Status Fit for MAS
Google Cloud NotebookLM Enterprise API (released Sep 2025) Official, supports notebook CRUD, source management, audio overview generation, queries. Workspace/Enterprise customers only — not consumer. Viable for organisations on Google Workspace Enterprise. Adds Google Cloud auth surface to MAS.
Unofficial Python API + CLI (tmc/nlm, notebooklm-cli, notebooklm-py) Active community projects; cookie-based auth via dedicated browser profile Lowest-friction for individual builders. Per-user authentication; rate-limit and ToS risk.
MCP servers (PleasePrompto/notebooklm-mcp, julianoczkowski/notebooklm-mcp-2026, roomi-fields/notebooklm-mcp) Multiple actively maintained; designed for Claude Code, Claude Desktop, Cursor, VS Code. As of Jan 2026 the CLI + MCP merged into a single nlm package Strongest fit — MAS as a Claude Code plugin can declare an optional MCP dependency. User installs the MCP server; MAS skills invoke NotebookLM tools through it.

What NotebookLM can do that’s directly relevant to MAS:

  • Step 3 (Buyer interviews) — Drop interview transcripts into a notebook; ask for theme map, saturation analysis, costly-action signal extraction. NotebookLM’s source-grounding posture aligns with the “verbatim quotes only” rule (concept §6).
  • Step 4 (Framework) — Drop the chosen expert’s books/courses/interview transcripts in; ask for “summarise their framework, identify 3 to keep verbatim and 3 to adapt for our ICP”.
  • Step 5 (Competitor) — Drop top-10 competitor websites/case studies in; ask for clustering and positioning whitespace map.
  • Step 6 (Synthesis) — Drop everything in; generate audio overview (10-min podcast briefing for the founder pre-VC meeting), mind map of the validated thesis, briefing doc that complements the three written artifacts. Not a slide renderer — NotebookLM does not output PPTX. It outputs prose, audio, and visual maps.

What NotebookLM cannot do for MAS:

  • Replace the Mini Council adversarial logic (NotebookLM is grounded but not adversarial-by-default)
  • Replace the Mom Test linter (NotebookLM doesn’t enforce interview methodology)
  • Render .pptx slides — slide rendering remains a separate question
  • Run inside a fully self-contained Claude Code plugin without external auth (cookie or OAuth)

Auth/coupling friction:

  • Cookie-based auth (unofficial CLI/MCP): user logs into Google in a dedicated browser profile, cookies are extracted, login persists. Friction at first run, smooth thereafter. ToS risk if Google blocks unofficial automation.
  • Enterprise API: requires Google Workspace Enterprise + GCP project + service account. High friction for individual builders; appropriate for organisational deployments.

Verdict on Q3 augmented: NotebookLM is a serious option for Steps 3–6 synthesis, particularly via MCP. It does not eliminate the slide-rendering question (NotebookLM doesn’t produce PPTX) but it does replace much of the Extensive-tier theme clustering + web research tooling with a single integration. Treated as a tier-bridging option below.

Cross-cutting: what we get for free

These campaign-mode capabilities are reusable as-is, no work needed:

  • Conversation transcripts (.campaign/conversations/{date}-{agent}.md verbatim record) — directly serves the verbatim-quote requirement in Concept §6
  • Progress Log format and triggers
  • AskUserQuestion presentation rules (one question, context summary)
  • Profile system (light skinning of NPCs even if we don’t use it for the Council)
  • Six Animals as advisory layer for the user’s team thinking (orthogonal to the Mini Council)

Summary — what must be built

Classified by type of work:

New SKILL.md files (5):

  • seed-vc-agent/SKILL.md
  • churned-customer-agent/SKILL.md
  • competitor-founder-agent/SKILL.md
  • future-self-agent/SKILL.md
  • grumpy-pm-agent/SKILL.md

New commands (≈ 8):

  • /sprint-setup (bootstraps .sprint/ directory)
  • /sprint-start (Step 1 entry)
  • /sprint-step-N (one each for steps 2–6, gated on prior-step approval)
  • /sprint-council (Mini Council orchestration; runs at Step 1→2 and Step 4→5 transitions)
  • /sprint-loopback (resets to Step 1 with retained context)
  • /sprint-continue (re-entry)

New state schema:

  • .sprint/sprint.md (replaces .campaign/quest.md shape)
  • .sprint/interviews/{N}-{interviewee-pseudonym}.md per buyer interview

New code (lightweight):

  • Banned-hypothetical-question linter for interview scripts
  • Saturation detector for Step 3 (count themes-since-last-new-theme)

New code (heavy, only in Extensive tier):

  • Slide rendering (PPTX/PDF)
  • WebSearch + scrape for expert and competitor discovery
  • Theme clustering for the saturation detector beyond simple count
  • Anti-sycophancy regression eval

Optional integration (tier-bridging, see §7a):

  • NotebookLM via MCP — replaces much of the heavy tooling above for Steps 3–6 synthesis, at the cost of an external auth surface and Google dependency

Pure prompt engineering (no code):

  • Step 1–6 command prompts
  • Mini Council voice prompts (within their SKILL.md files)
  • Anti-sycophancy preambles (lifted from Dragon)
  • Artifact templates (sales deck, outreach plan, investor deck)

Map to research questions

Q Question Answer Confidence
Q1 Sequence fit Requires schema + new gate command — not pure rewrite, but not a fork either High
Q2 Mini Council depth Profile packs insufficient; need 5 new SKILL.md files High
Q3 Artifact reach Markdown-first closes 80%; slide rendering is optional High
Q4 Loopback semantics Schema rewrite + 1 new command High
Q5 Anti-sycophancy Tone prompt-only; verification needs eval set (Extensive tier only) Medium — verification claim untested
Q6 Distribution Open — needs user input n/a
Q7 Ownership Open — needs user input n/a

Q6 and Q7 are addressed in the recommendation where build-tier choice depends on user answers.