View on GitHub

campaign-mode

Work with a party of AI advisors with genuinely different perspectives.

Council Report

Animal Perspectives

Bear — Vision & Direction

  1. The vision is well-articulated and architecturally sound. The three-NPC decomposition (Gandalf/Dragon/Guardian) maps cleanly to mentorship, adversarial testing, and quality gates. The context isolation model (advisory/independent/maximum) is a genuinely original contribution that prevents the AI from gaming its own evaluations — the project’s most defensible innovation.

  2. The “User as Protagonist” principle is the strategic anchor. Every design choice — mode selection, agent invocation, proactive elicitation — flows from SDT’s autonomy construct. This isn’t flavour; it’s architectural backbone.

  3. The v0.2 scope is disciplined. Council, quest lifecycle, three NPCs, plugin distribution. Open questions (Simon’s dual role, multi-user campaigns, cultural adaptability) are properly deferred. The roadmap shows restraint.

  4. The gap between vision and validation is the strategic risk. 11 ADRs, 12 specs, 51KB of skill definitions — but no evidence of user campaigns having been run. The architecture is well-designed in theory; the question is whether the quest flow works in practice, particularly whether Dragon’s maximum isolation produces genuinely independent evaluation.

Cat — Risk & Scope

  1. Context isolation is enforced by instruction, not architecture. ADR-CM-003 acknowledges this — the Dragon receives “only success criteria + work product” because the SKILL.md says so, not because there’s a technical boundary. When Claude plays all roles in the same session, the model has seen everything. “Maximum isolation” is aspirational. This is the single biggest integrity risk.

  2. The documentation-to-implementation ratio is inverted. 11 ADRs, 12 specs, a north-star, and a detailed README — for three SKILL.md files, four command files, and no runtime code. Design documentation outweighs implementation by roughly 4:1. This signals over-planning relative to what’s been built and tested.

  3. Several ADRs are still “Proposed” status. The master ADR (001) and others haven’t been formally accepted. Fifteen open questions remain in the north-star, some resolved but not all cleaned up. This creates ambiguity about which decisions are final.

  4. The Tolkien naming carries IP risk. “Gandalf” is a trademarked name. Using it in a distributed marketplace plugin is a non-trivial legal exposure that should be resolved before wider distribution.

  5. No tests, no validation harness. There’s no way to verify that agents behave as specified. When a SKILL.md is 21KB of behavioural instructions, how do you catch regressions?

Owl — Process & Structure

  1. File organisation is clean and well-documented. Canonical skills in skills/, auto-discovery in .claude/skills/, commands in commands/, docs with numbered subdirectories. The README’s project structure section matches reality. Well-organised.

  2. Duplication between skills/ and .claude/skills/ creates maintenance burden. Any change must be made in both locations with no automated sync. Plugin distribution may reduce relevance, but for clone-path users, drift is a process risk.

  3. Campaign lifecycle has clear phase transitions but no enforced ordering. Six phases are documented, but nothing prevents invoking the Dragon before the Guardian or skipping quest definition. Guardrails are advisory.

  4. Version management is informal. Plugin.json says v0.2.0, README has milestones, but there are no git tags for releases. Changelog is embedded in README. For marketplace distribution, formalising versioning would help.

Puppy — Opportunities & Momentum

  1. The quest metaphor is genuinely powerful. Framing work as a quest with mentors, adversaries, and checkpoints creates emotional engagement. The three campaign modes let users choose their relationship with the work. “Dragon Confrontation” is memorable and motivating in a way “acceptance testing” never will be.

  2. Proactive elicitation is a major UX win. Agents offer next steps at every phase transition — users never remember slash commands. “Ending a phase without a next-step question is a bug” shows real commitment to user experience.

  3. Character profiles are an untapped opportunity. The generation system (flavour + modifier, theme-agnostic) is built but not exercised. Example profiles in the README would show people what’s possible — a quick win for adoption.

  4. The Council feature is already valuable. Multi-perspective project diagnostic with synthesis — a powerful tool even outside quest context. Could be the gateway feature that gets people to try Campaign Mode.

Rabbit — Resources & Stakeholders

  1. Dependency chain is minimal but critical. Claude Code (platform) and Six Animals (recommended). Six Animals detection via five paths is robust engineering. Both are external dependencies — Claude Code’s skill system could change, Six Animals has a separate maintainer.

  2. Distribution has four paths but uneven maturity. Plugin (recommended), clone, copy-to-project, and personal skills. Different installation methods create different user experiences. Pragmatic but creates support surface area.

  3. Bus factor of 1. Two listed authors but single-author git history. For a project with this scope of ambition, the sole-implementer risk is worth noting.

  4. No community infrastructure. No CONTRIBUTING.md, issue templates, discussion forum, or examples directory. For marketplace distribution, community scaffolding would help adoption.

Wolf — Cohesion & Balance

  1. Design documentation dominates; practical guidance is thin. Extensive ADRs and specs — but no example campaigns, sample profiles, or end-to-end walkthrough. A new user understands the concept but doesn’t see what it feels like. The intellectual framework is strong; experiential onboarding needs warmth.

  2. NPC agents are thoroughly specified; animal integration is assumed. Gandalf, Dragon, and Guardian have 14-21KB SKILL.md files. But animal agents’ campaign-aware behaviour is listed as “future.” The animals don’t know they’re in a campaign. Council bridges this partially.

  3. Progress tracking asks all agents to log, but animals aren’t campaign-aware. CLAUDE.md says “all agents — including animal agents — must track meaningful progress,” but animal SKILL.md files (in Six Animals) contain no campaign-mode instructions. Specification-reality mismatch.

  4. Grow mode is richer than Ship mode in design attention. Character profiles, transformation criteria, deep debrief, and ZPD assessment all lean toward Grow. Ship mode is “streamlined” but has fewer unique features. For broad appeal, delivery-focused mode should feel equally compelling, not reduced.

Consensus

Common themes: The design is mature but validation is not — the project has been deeply thought and now needs to be deeply used. Context isolation is both the core innovation and the core risk (conceptually strong, instruction-enforced). The animal-NPC integration gap creates a specification-reality mismatch around campaign-aware behaviour.

Key tensions: Depth vs. breadth (Grow mode richer than Ship, but Ship users likely broader audience). Documentation vs. validation (4:1 spec-to-implementation ratio). Independence vs. integration (designed as complementary but Council assumes animals are present).

Overall assessment: Campaign Mode is a well-architected quest framework at the transition point from design to adoption. The NPC decomposition, context isolation model, and user-as-protagonist principle are genuine contributions. The project now needs the shift from designing the system to experiencing the system — running campaigns, collecting feedback, and letting practice inform the next iteration.

  1. Run a real campaign end-to-end. Complete a full quest lifecycle (Council through Debrief) on a real project challenge. This will surface whether phase transitions, proactive elicitation, and context isolation work in practice.
  2. Create example content. Sample character profiles and a campaign walkthrough would transform the README from “how it works” to “what it feels like” — a quick win for adoption.
  3. Address the Gandalf naming risk. Resolve the Tolkien IP question before wider marketplace distribution. Decide whether to keep, rename, or disclaim.
  4. Add git tags for releases. Tag v0.1.0 and v0.2.0 in git to establish release points — a 2-minute task that improves project professionalism.
  5. Bridge the animal integration gap. Either add campaign-awareness to Six Animals skill definitions (upstream contribution) or adjust CLAUDE.md’s progress-tracking expectation to match what animals can actually do.