pm-llm-wiki boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run covers the shipped PM Wiki as a prompt-defined product-management knowledge system: its reusable schema, setup and operating instructions, content-model rationale, and complete fictional FitLoop pilot. The intended use is to retain source-linked product knowledge and reuse it for decisions, discovery plans and generated deliverables. It is in scope because an external model-backed coding agent performs the prescribed ingestion, synthesis, retrieval and maintenance; the repository is not itself an enclosing agent runtime (SRC-1, README.md, template/SCHEMA.md).

You (the LLM agent) are the maintainer of this product-management knowledge wiki. The human PM curates sources, directs analysis, and makes the calls; you do all filing, cross-referencing, synthesis, and bookkeeping. This file defines the structure, conventions, and workflows. Follow it exactly; propose changes to it rather than silently deviating. --- template/SCHEMA.md

The boundary is whole-system for that shipped knowledge system, not for its external agent host. Inspection traced the README setup into the schema, its ingest/query/lint instructions, the pilot's instantiated instructions, raw fictional inputs, retained pages, index, operation journal and filed-back PRD. The prescribed loop includes admission by the PM, agent-maintained persistence, index-first later consumption, proposal gates and recurring maintenance. The pilot is included as shipped state illustrating this loop, not excluded as an application caller. Its PRD is a retained downstream consumer of that state (SRC-1, README.md, template/SCHEMA.md, pilot/CLAUDE.md, pilot/index.md, pilot/log.md, pilot/wiki/prd-adaptive-plans-v1.md).

1. Read index.md, then only the relevant pages. 2. Synthesize with evidence chains. 3. Offer to file valuable answers back as a wiki page (analyses, comparisons, prioritizations compound too — file under wiki/ and log as exploration). --- template/SCHEMA.md

The Git checkout's HEAD was verified against the acquired full commit. Only this repository at 64fb748ac9d187be4c51636aa56bfee1d9e5b116 supplies evidence; the applicability cutoff is 2026-11-09. The evidence tier is doc-grounded: shipped instructions and example Markdown state support intended operation, not executable enforcement. The tracked tree contains no executable host, provider adapter, automated evaluator or test suite. The contribution guide prescribes fresh-agent testing, but supplies no inspectable transcript or evaluation results from which to establish reliability (SRC-1, CONTRIBUTING.md, README.md).

Test schema changes the way this repo was tested: give the changed schema to a fresh agent session with a minimal prompt and check it behaves. Say in your PR what you observed. --- CONTRIBUTING.md

Tracked top-level coverage

All tracked files were inspected. “Included” means inside the functional or contextual analysis boundary; it does not promote prose to implementation evidence. The table accounts for every top-level tracked file or directory (SRC-1).

External dependencies and prevented conclusions

The PM curates inputs, starts operations and approves gated edits. A separately supplied agentic tool with local-file access executes the model calls and file operations; the README names Claude Code as reference tooling and Codex or similar agents as alternatives. A folder is required; Obsidian and Git are optional user tools (SRC-1, README.md). Their implementations, permission systems, context loading and model behavior are outside this repository boundary. Consequently this run cannot establish runtime scheduling, context-window policy, approval enforcement, atomicity, concurrency safety or provider-specific behavior from this source.

The source explicitly requires the pilot's product evidence to remain fictional (SRC-1, CONTRIBUTING.md; also pilot/CLAUDE.md):

Everything in pilot/sources/ and pilot/wiki/ is fictional and must stay clearly marked as such. Never add real customer data to this repo. --- CONTRIBUTING.md

Its populated pages and journal permit analysis of example structure and declared operation, not real customer outcomes, observed model compliance or causal benefits. No independent execution trace or causal experiment was supplied. This prevents claims of measured effectiveness, real-world maintenance cost or reliable gate adherence. Proposed later search, scheduled automation and graph tooling are not treated as deployed dependencies or implemented features (SRC-1, template/SCHEMA.md). These limits are evidence limits, not uninspected material shipped paths; no material portion of the repository's own prompt-and-state loop is deliberately excluded.

Source register