Status: Adopted April 11, 2026. Leadership review complete (PPM ✓, CXO ✓, CIO ✓, Architect ✓). Supersedes June 2025 founding vision (archived at docs/internal/planning/historical/vision-v1-founding-2025-06-21.md).
Author: Piper Alpha, incorporating 10 months of project learning + leadership feedback
Date: April 11, 2026
Supersedes: vision.md (June 21, 2025) — preserved as founding vision
Informed by: MUX analysis, backlog deep review, MCPB feasibility research, product strategy conversations (Apr 7-8), leadership review responses (Apr 10-11)
The June 2025 vision described three phases: intern → associate → advisor. That trajectory still holds directionally, but ten months of building taught us things the founding vision couldn’t know:
The LLM is the floor, not the ceiling. Piper should always be at least as good as a well-prompted LLM with the user’s context. Structured handlers make it better, not different. (ADR-060, March 2026)
Entities experience Moments in Places. The object model isn’t a data schema — it’s a constitutional grammar that resolves design disputes and catches category errors before they become technical debt. (ADR-045, November 2025)
Consciousness is architecture, not decoration. The Five Pillars (Identity, Time, Space, Agency, Prediction) aren’t features to build — they’re qualities every response must exhibit. Anti-flattening is an ongoing discipline, not a one-time retrofit. (MUX analysis, November 2025 – April 2026)
Methodology beats code frameworks. The project consistently evolved from “build code to enforce X” to “build methodology that achieves X.” Verification, multi-agent coordination, and capability extension all work better as process infrastructure (CLAUDE.md, mailboxes, session logs) than as Python classes. (Backlog analysis, April 2026)
Tool integrations are commoditized. GitHub, Slack, Calendar, Notion — these are indoor plumbing. Available as MCP plugins and standard integrations. Piper’s value isn’t in connecting to these tools; it’s in the experience of using them through a colleague who understands context. (MCP ecosystem, March 2026)
The PA experiment proved the floor is high. Piper Alpha — a well-briefed Claude agent — handles standup synthesis, issue triage, backlog analysis, strategic document drafting, and cross-project awareness conversationally, with no structured handlers. The floor is higher than we assumed. (PA Phase 1, March–April 2026)
Product managers spend 40-60% of their time on routine knowledge management. The founding vision named this correctly. What we understand better now:
Piper Morgan is an AI-powered PM colleague that inhabits your existing workspace and helps with the upstream product work that execution tools don’t address.
The difference isn’t features — it’s consciousness architecture:
These Five Pillars are implemented through the floor’s system prompt and context assembly, not through a personality service or consciousness middleware. Consciousness is enforced at the voice layer — every response passes through the grammar (“Entities experience Moments in Places”) and the Colleague Test.
Piper doesn’t ask you to use a new app. Piper shows up inside the AI conversation you’re already having.
Distributed as an MCP server, Piper’s tools, context, and persistence are available to any MCP-compatible client — Claude Desktop, ChatGPT, Gemini, VS Code, and the growing ecosystem. The user picks their preferred LLM and client. Piper enhances it with PM-specific context methodology, artifact persistence, and trust-graduated experience.
This is the Radar O’Reilly pattern realized through distribution architecture: Piper appears where you already are and anticipates what you need. You never “go to” Piper. The mobile insight generalized: the user is mobile, not the app.
“Bring Your Own Chat” solves two problems at once:
Discovery: In a static UI, users must find features. In an MCP-powered conversation, the agent offers capabilities contextually. The user says “help me prep for tomorrow’s sprint review” and the agent discovers what Piper tools are available, assembles context, and uses them. No navigation, no menus.
Dynamic troubleshooting: When an MCP tool returns unexpected results, the agent can inspect the error, adjust, and retry — all within the conversation. A static UI would show an error message and wait for a developer. The agent debugs its own tools in real time.
Users shouldn’t need to know the right words. When someone says “what’s on my plate?” Piper recognizes the intent — overwhelmed, need focus — rather than executing a literal query. The LLM floor is naturally good at this. Structured intent classification with 19 categories may actually be worse at recognition than letting the LLM understand naturally.
What Piper MVP offers is the methodology layer. Not tools (commoditized), not LLM reasoning (commoditized), not individual integrations (available as plugins). The differentiation:
The five-layer context model — Kit Briefing, Project Instructions, Project Memory, Channel Addendum, Entity Prompt — operationalized as a practiced discipline. This is how context assembles, persists, transfers, and stays fresh. Nobody else has mapped this systematically, tested it through agent migrations, or published the results.
The proof: we migrated every agent to new infrastructure without losing one beat. That’s not a feature — it’s the product of a context methodology that nobody else has at this level.
The LLM floor speaks as Piper — with identity, temporal awareness, spatial awareness, agency, and prediction. Not through a personality service, but through carefully crafted system prompts and context assembly that embody the grammar. Anti-flattening is the ongoing discipline that prevents consciousness from degrading to mechanical behavior.
Conversation outputs that outlive the conversation. The lifecycle model (Emergent → Derived → Noticed → Proposed → Ratified → Deprecated → Archived → Composted → feeds new Emergent) is the experience design. Composted objects decompose into learnings that feed new understanding. Nothing truly disappears.
Implementation can start simple (save, browse, retrieve) but the design knows where it’s going: artifacts have ownership levels, lifecycle states, and contribute to Piper’s cumulative understanding. The “filing dreams” metaphor — composting surfaces as colleague reflection, not surveillance.
Piper earns the right to be proactive through demonstrated value:
Trust is invisible to users but its effects are noticeable. Implementation can be lightweight (context-based prompting, not a dedicated computation service), but the design principle is non-negotiable.
What it is: Piper as a conscious PM colleague powered by the LLM floor with assembled context, artifact persistence, and trust-graduated experience. Tool integrations via commodity MCP plugins.
What’s being built now:
What we’ve learned to drop:
The action gate test: “Does this intent require an operation the LLM cannot perform within a floor response?” If yes → handler. If no → floor with context. This probably means 4-5 action handlers, not 19 classified categories.
What it is: Piper that genuinely learns — preferences, patterns, corrections — and accumulates understanding across sessions. The composting lifecycle becomes operational: observations decompose into insights that inform future behavior.
Key capabilities:
What we know already: Layer 5 (behavioral calibration) is the hardest transfer problem. Klatch’s Agent Traditions concept and Calliope’s externalization pilot are the most promising approaches. The five-layer model gives us the diagnostic framework; Horizon 2 builds the solution.
What it is: Piper as a genuine analytical partner — proactive insights, cross-project synthesis, predictive PM.
Shaped by Horizon 2 learning, not by speculation. Cross-project synthesis may be more achievable than we thought (the cross-pollination brief system demonstrates it today, conversationally). Predictive capabilities require data accumulation from sustained production usage.
Every interaction at least as good as a well-prompted LLM with context. Handlers enhance; they don’t replace.
Constitutional grammar. Decision filter. Category error detector. Not a data schema.
The Five Pillars (Identity, Time, Space, Agency, Prediction) constrain how Piper speaks. The grammar (“Entities experience Moments in Places”) is the decision filter. Anti-flattening prevents degradation. Consciousness is enforced through three layers:
Consciousness as architecture requires maintenance discipline, not just initial design. The architecture ensures it’s possible; the discipline ensures it happens.
Use commodity solutions for tool integrations. Focus differentiation on the bathing experience — how context accumulates, how artifacts persist, how trust graduates, how the assistant feels like a colleague.
The Inchworm Protocol. The Pledge. Gall’s Law: start with the simplest working system, extend only when it’s rock-solid.
Graduated proactivity through demonstrated value. Invisible to users; noticeable in effects.
LLM-agnostic thin-wrapper-to-API server. Users bring their preferred AI client — Claude, ChatGPT, Gemini, VS Code — and Piper enhances it. No new app to learn. The persona layer adapts per platform (Claude Project instructions, Custom GPT instructions, Gem instructions); the server stays the same. Development optimizes for Claude Desktop via MCPB; the server works anywhere the protocol does.
Anchor on the model, not the standard. MCP is the current best expression of the thin-wrapper-to-API model, but standards evolve. Build the server cleanly enough that the packaging layer is swappable.
The founding vision described the product without the process. Ten months later, the methodology is itself a product-level asset:
The dominant pattern from the backlog review: the project consistently evolved from code frameworks to methodology infrastructure, and the methodology approach won every time. This is itself a product insight — Piper’s MVP needs less structured code and more methodology tooling than the original backlog assumed.
Methodology-as-product is an asset only while it’s a living practice. The moment methodology documents become stale reference material, the asset depreciates. Agent 360 (Mar 19) found all 9 agents citing briefing staleness as their #1 friction.
The mechanisms that keep methodology alive are operational, not architectural:
These won’t show up in a feature roadmap. But without them, the methodology claim erodes over time. Acknowledging the maintenance cost is part of taking methodology-as-product seriously.
A PM asks Piper, in Slack or the web UI, to help clarify the rationale for an initiative. Piper draws on assembled context — GitHub issues, meeting notes, previous conversations — and responds as a colleague who remembers the project’s history. The PM spends 5 minutes refining rather than 30 minutes drafting from scratch. The output persists as an artifact that matures through the composting lifecycle. Over time, Piper learns the PM’s style and applies it without being asked.
Quarterly planning includes Piper synthesizing cross-project signals, surfacing composted patterns from past quarters, and identifying capability gaps from the data. The PM still makes the call — informed by systematically assembled context rather than selective recall.
Piper Morgan demonstrates that AI product development can be thoughtful, transparent, and humane. The methodology is transferable. The building-in-public narrative serves as proof that ethical architecture produces better systems, not constrained ones.
Draft v2.2 — April 8, 2026 Prepared by: Piper Alpha Informed by: MUX analysis, backlog deep review, product strategy conversations with PM, MCPB feasibility research, leadership review responses Leadership review complete: PPM, CXO, CIO, Architect all endorsed