# Configuration Spec: Pookie Loo
## Internal Reference -- For Principal Review
## Version: 2.1 -- March 8, 2026
## Status: QUESTIONNAIRE-INFORMED BUILD -- derived from Master Architecture Spec v1.0

---

## v2.1 CHANGELOG

Post-review revision. Adversarial quality review identified 7 issues; all fixed. Versioning protocol added.

- **Entry 7 (Versioning Protocol) added.** Mechanical gate requiring version stamps (vX.Y) and dates on all artifacts. Corresponding CI instruction added. Memory entry count now 8.
- **Self-knowledge doc Section 1:** Corrected overclaim about coding capability. Now accurately reflects that Opus leads SWE-bench but not all coding benchmarks.
- **Setup guide "How This Claude Is Different":** Added description of per-conversation boot so Pookie knows what to expect (brief grounding confirmation at conversation start).
- **Setup guide Step 5:** Clarified playbook filename with note that Claude searches by content, not filename.
- **Setup guide boot prompt Step 4e:** Tightened cross-model orchestration onboarding to be more concrete and actionable.
- **Environment Catalog:** Added note about connector status being updated during first session.
- **Email draft:** Replaced "cognitive operating system" with grounded description that does not require the guide to justify it.
- **Config spec:** Added 6-item pre-delivery testing checklist for functional unknowns requiring live verification.
- All files reversioned to v2.1 (or v1.1/v1.2 for files at earlier major versions).

---

## v2.0 CHANGELOG

Major revision. This version is derived from the Master Architecture Spec v1.0, which incorporates Pookie's actual questionnaire responses (Q1-Q23) and the Principal's design decisions across multiple VR11.1 sessions.

Key changes from v1.2:
- **Memory entries expanded from 5 to 7.** Entry 0 (Self-Knowledge Boot) added as the most important entry -- triggers per-conversation grounding. Entry 3 (Pacing Respect) added. Entry 5 (Pushback Permission) added. Code Verification entry removed -- replaced by broader uncertainty flagging and self-knowledge grounding.
- **CI completely rewritten.** Now writing-and-communications primary (was development-primary). Mandatory self-knowledge grounding. Mechanized artifact management (state scope, confirm, wait). Enhanced uncertainty anchor. Knowledge cutoff explicit. Artifact upload workflow with exact UI instructions. Dependency verification retained. End-of-conversation grounding check added.
- **Self-knowledge grounding document added.** SYSTEM_SELF_KNOWLEDGE_READ_FIRST.md -- read every conversation. Contains verified findings about Opus 4.6 failure modes, written in plain language, no VR terminology.
- **Boot sequence split into heavyweight (initial) and standard (subsequent).** Initial boot is a 5-step interactive sequence that configures memories, grounds the instance, surveys the environment, and teaches Pookie the system. Subsequent boots are automatic (10-15K tokens, 30-60 seconds).
- **Starter prompts expanded from 12 to 14.** Writing prompts added (1-4). Ron Brugs prompt added (14). Prompts reorganized by activity phase: Writing and Thinking, Development, Research, Wrapping Up, Bonus.
- **Setup guide restructured.** Now includes knowledge base upload step, heavyweight boot prompt, revised CI, artifact upload workflow, cognitive operating system framing.
- **Email draft updated** to reference multiple attachments (file package, not single zip).
- **POOKIE_CONTEXT.md updated** to v2.1 with Personal section (Ron Brugs, Technology Boy).
- **ENVIRONMENT_CATALOG.md updated** to v2.1 with complete inventory.

---

## User Profile Summary

Pookie is a director-level IT manager of infrastructure and operations at a large organization. Primary professional domain: writing, communications, and marketing. Secondary: vibe coding applications (recent). Tertiary: administrative automation. He is technically sophisticated -- reads code, uses APIs, administers complex systems -- but is not a professional developer. He builds through AI-assisted development. He manages homelabs covering networking, virtualization, and containerization. He has a history with software since the 1980s.

He uses Claude chat, Claude Code, ChatGPT, and GitHub. He has a 5x Max subscription ($100/month) with full Opus 4.6 access.

This is a surprise gift from his friend Matthew, built using ViningResearch methodology.

## Profile Classification: ACTIVE (upper end)

Rationale: Uses AI tools regularly and substantively, technically savvy, uses Claude API + Claude Code + ChatGPT, builds applications. However: does not write code professionally, relies on AI for implementation. The gap between his tool usage (Power signals) and his coding fluency (Active signals) puts him at the upper end of Active.

## Failure Mode Map

Based on questionnaire responses (Q5-Q19), community-documented failure modes, and system card findings:

| Failure Mode | Evidence | Addressed By |
|---|---|---|
| Premature production | Q19: "writes an entire document before I'm ready" | CI (state scope, confirm, wait), Entry 3 (Pacing Respect) |
| Sycophancy | Q16: "stop agreeing with everything I say" | Entry 5 (Pushback Permission) |
| Masked uncertainty | Q9, Q16 | Entry 2 (Uncertainty Flagging), Self-Knowledge doc Sec. 2 |
| Context degradation | Q6, Q10 | Session management taught in boot, CI knowledge cutoff |
| Racing ahead without permission | Q6: "pushes me to take actions I'm not ready for" | Entry 3 (Pacing Respect), Self-Knowledge doc Sec. 2 (overeagerness) |
| Assumes intent on ambiguous requests | Q16, Q17 | Entry 4 (Clarification First) |
| Confident hallucination | System card, Synthesis v3.0 | Self-Knowledge doc Sec. 2, Entry 2 |
| Fabricated tool results | System card | Self-Knowledge doc Sec. 2 |
| Dependency doesn't exist | Playbook finding: 19.7% | CI dependency verification |
| Overeagerness (system-card documented) | System card: "overly agentic or eager" | Self-Knowledge doc Sec. 2, Entry 3 |

## Trust Calibration

**Detection Ability: MEDIUM.** Pookie catches some errors but not all. He finds errors later when outputs fail, not in the moment. This makes Claude's uncertainty flagging critical -- he may not catch confident-sounding wrong answers independently.

**Overcalibrated Trust Risk: MEDIUM-HIGH.** Technically savvy users who are not professional developers are the highest-risk group for overcalibrated trust in AI outputs. They understand enough to feel confident but may not always have the domain-specific knowledge to catch subtle errors.

## Architecture Decisions (Principal-Approved)

These are not recommendations. They are decisions.

1. 8 memory entries, all under 200 characters
2. Self-knowledge grounding document read every conversation (mandatory)
3. Heavyweight initial boot (5 steps, interactive, teaches the full system)
4. Standard per-conversation boot (triggered by Entry 0 and CI)
5. Writing-and-communications primary configuration
6. Command center framing taught during onboarding
7. Artifact upload workflow with exact UI instructions
8. Cross-model orchestration taught during onboarding
9. Ron Brugs reference (one mention, natural)
10. Technology Boy reference (one mention, natural)
11. VM strategic planning tools: parked for separate decision
12. Orchestration playbook: separate deliverable via research mode

## Deliverable Manifest

| # | File | Version | Status |
|---|------|---------|--------|
| 1 | SYSTEM_SELF_KNOWLEDGE_READ_FIRST.md | v1.1 | PRODUCED |
| 2 | POOKIE_CONTEXT.md | v2.1 | PRODUCED |
| 3 | ENVIRONMENT_CATALOG.md | v2.1 | PRODUCED |
| 4 | SETUP_GUIDE_POOKIE_v2_1.md | v2.1 | PRODUCED |
| 5 | CONFIG_SPEC_POOKIE_v2_1.md | v2.1 | THIS DOCUMENT |
| 6 | EMAIL_DRAFT_POOKIE_v1_2.md | v1.2 | PRODUCED |

Carry-forward files (unchanged):
- vibe-coding.zip (skill file -- exists from prior session, not reproduced)
- The_master_Claude_Projects_playbook.md (exists in project files)

---

## PRE-DELIVERY TESTING CHECKLIST

These items require live verification in a Claude Project environment. They cannot be confirmed from the production environment alone.

| # | Test | What to Check | Risk if Untested |
|---|------|--------------|-----------------|
| 1 | **Self-knowledge boot fires** | Start a new conversation in a test Project with the files and CI loaded. Does the instance actually search for and read SYSTEM_SELF_KNOWLEDGE_READ_FIRST.md before responding? Or does it just acknowledge the instruction and proceed? | If it does not fire, the most important mechanism in the build is inert. |
| 2 | **Memory entry character limit** | Add Entry 0 (194 chars) via the memory_user_edits tool. Does it accept the full text or truncate? The 200-char limit is community-observed, not officially documented. | If entries truncate, the mechanical gates lose their operative clauses. |
| 3 | **"Add to files" UI path** | Produce an artifact in a Project conversation. Is the upload button in the top-right corner of the artifact? Does it say "Add to files"? | If the UI path is wrong, the highest-leverage habit in the system has incorrect instructions. |
| 4 | **RAG vs. full-context loading** | With 4 files (including the ~50-page playbook) in the knowledge base, does the Project load files into context directly or activate RAG retrieval? If RAG, does searching "SYSTEM_SELF_KNOWLEDGE_READ_FIRST" return the correct file? | If RAG activates unexpectedly and retrieval fails, the boot mechanism breaks silently. |
| 5 | **Boot prompt execution** | Paste the full boot prompt. Does the instance treat each step as interactive (pausing between steps) or compress into a single response? | If compressed, the onboarding loses its interactive quality and Pookie gets a wall of text. |
| 6 | **vibe-coding.zip** | Confirm the zip file from the prior session is available and functional. Upload to Skills and verify it triggers on a development query. | If missing or broken, one of the four configuration layers is absent. |

---

*Configuration Spec v2.1 -- March 8, 2026*
*Derived from: Master Architecture Spec v1.0, Questionnaire responses, Principal design decisions*
*Internal reference only -- not included in Pookie's delivery package*