Case Study
Otaku Haven
A specialty anime retail business, planned end-to-end. 49 documents. 7 intents. One workday.
Otaku Haven
Otaku Haven is a fictional anime retail store based in Austin, designed as a thorough stress test for the Trail Framework. It was never intended as a real business, but serves to evaluate whether Trail's default method of ensuring consistency across multiple documents can handle an extensive, multidisciplinary set that could uncover any inconsistencies. This case study does not involve actual Texas LLC registration, licensing, lease negotiations, or anime IP licensing. The test focused on the coordination layer rather than on determining the legal feasibility of an anime retail business.
The deliverable included a full pre-launch business package comprising formation and brand documents; a detailed strategic package with a functioning multi-tab financial model featuring live formulas; a Texas-compliant employee handbook; eleven retail SOPs; finance and compliance policies; vendor and supply chain documents; and customer-facing legal and marketing materials. This totaled 49 documents across three formats (Word, Excel, PowerPoint), with 7 Intents and 9 Runs.
The Architect/Product Owner was a single individual who also acted as the Reviewer. The Manager was Codex, and the Developer was Claude Code. The total effort from the Architect/PO was about 12.25 hours, divided into three work categories: 7.5 hours for planning, which included identifying documents for a viable business plan, creating the meta documents, and seven intents; 3 hours and 16 minutes of build time spent orchestrating sessions, invoking the Manager for each Intent, and the Developer for each Run; and 1.5 hours for Checkpoint A and Checkpoint B reviews conducted between sessions during the build.
No Intent was modified after execution began. No Run bundle required revision at Checkpoint A. No output required revision at Checkpoint B. Across nine runs and 16 review checkpoints, the Architect's role was validation, not correction.
* Review was done during the build between and during intents, plus an additional 45 minutes of validation at the end. Actual review time is approximately 1.5 hours, as the Architect's role is to validate after run creation and after each run completes.
** See Why this matters below.
How Trail coordinated the build
This was not a custom coordination layer. Trail's default policy did the work.
The canonical reference data file, Otaku_Haven_Reference_Data.md, is stored in trail/meta/files/. This single file contains unchanging facts such as legal entity details, ownership structure, address, store hours, compensation tables, technology stack, fictional timeline, and financial assumption ranges. According to Trail's standard file-access protocols, files located in trail/meta/files/ can be used as Developer inputs when explicitly declared. Every Dependency & Files section for each Intent declares this file, and the Developer accesses it directly during each Run — not through Manager translation, but as a specific input declared in the process.
Trail's global-operating-instructions.md contains a Cross-Document Coherence section that mandates exact-value reads when referencing content from prior deliverables. The rule states: "The Developer must read the source document and use the exact values. Do not paraphrase numbers or approximate figures." When applied consistently across nine iterations, this policy achieved the referential accuracy needed to ensure that the Business Plan's $54,733.23 break-even point precisely matched the Financial Model's output, as the Developer was required to read the model and quote the cell directly, avoiding approximation.
Architect-declared cross-document dependencies. Each intent's Dependencies & Files section listed the prior deliverables the Developer needed to read for that intent. Intent 7's Return Policy declared SOP-005 from Intent 4 as a required input. Intent 5's Cash Management Policy declared SOP-003. Intent 6's Receiving Guidelines declared SOP-007. The Manager included these in each run bundle. The Developer read the named files directly.
Forward-reference documentation: When an Intent produces decisions needed by a later Intent — such as cash-handling thresholds, return windows, or revenue splits — the intent's operating-instructions-override.md requires the Developer to record those decisions in results.md. Managers of subsequent Intents review those results.md files when planning the next run.
The Architect did not engineer custom mechanisms for this process. The reference data file, declared dependencies, and override-enforced forward references are all standard Trail patterns. The Architect's preparatory work involved the essential planning that any business plan effort requires: selecting which documents would form a viable plan, identifying their dependencies, and recording key facts in the reference data. Trail's role was to create a structure that transformed that planning into executable artifacts. A consultant performing the same planning would have created a project brief, but the Architect developed seven intents and a meta package.
File context equaled execution context
Every Manager and Developer session began without any prior context.
The Architect ended each AI session after finishing its task and started a new one for the next step, with no continuity or inherited conversation history. Over nine runs and the Manager's planning that preceded them, no AI session had access to anything beyond the contents of the files.
This indicates that coherence across the document set could not be achieved by a single AI maintaining context across multiple runs. Codex did not "remember" Intent 2 when planning for Intent 5. Similarly, Claude Code did not recall the Financial Model from a previous session while writing the Business Plan. Each session started without any prior context, only reading the Run bundle and declared inputs, then generating output. The artifacts were the only elements shared between sessions.
Trail's invariant that "file context equals execution context" is validated through the actual workflow, not just policies. The files contained all the necessary information for the Developers and Managers. The receipts confirm that this invariant was maintained.
The Architect/Product Owner's role
The Architect/Product Owner is human only in Trail. The role exists at the two checkpoints in the loop where intent is at risk of drift.
Each intent runs through a five-step loop:
- Architect/PO creates the intent package: Three files: intent.md, manager-instructions.md, operating-instructions-override.md.
- Manager creates the run bundle from the intent package: Five files (six with an optional workplan): the Manager translates intent into a runnable form, declaring all permitted inputs.
- Checkpoint A — Architect/PO reviews the run bundle: Verifies the Manager's translation is faithful before the Developer reads anything. Approval invokes the Developer; rejection returns it to the Architect for a new intent.
- Developer executes the run bundle: Reads only the run files plus declared inputs. Produces deliverables and populates results.md.
- Checkpoint B — Architect/PO reviews the output: Validates emergent outputs and decisions before downstream intents consume them. Outcomes: accept and close, request fixes (new run, same intent), or re-scope (new intent).
Across 7 intents and 9 runs, there were 16 human checkpoints distributed across the build, totaling about 1.5 hours of Architect/PO review time.
The Architect/PO does not draft or edit documents. The Architect/PO does the irreducible human work: defining what is wanted, verifying that Intent was correctly translated, and validating that output matches Intent. These are Product Owner concerns. At each checkpoint, the question is not "is this prose good?" but "is this what we wanted?"
Zero revisions across the build
The two-checkpoint review structure is intended to identify errors early and prevent their spread. In this build, no errors were found.
Of the nine run bundles handled by the Manager, none required revision at Checkpoint A because their translations were sufficiently accurate, and each bundle proceeded directly to Developer execution. Similarly, among the nine sets of Developer output, none required revision at Checkpoint B, and all executions met the acceptance criteria on the first attempt. This does not mean they are production-ready; it means they are first-draft ready for the Architect/PO to work from.
This does not claim that AI is flawless. Instead, it asserts that Trail's framework — characterized by transparent objectives, faithful Manager translation, clearly declared inputs, strict read policies, and role-specific execution — established an environment where AI contributors could operate reliably. All actions are auditable: each Run bundle stored in the run folders, every results.md file filled out by the Developer, and each Checkpoint A and B decision, inferred from the lack of a subsequent run within the same intent.
Trail's role separation didn't just work; it worked. It worked without any of the structural failure modes that the framework is designed to prevent.
Why it matters for the financial figures
The Financial Model includes nine interconnected Excel tabs with live formulas.
Key Assumptions feed Monthly Revenue for Year 1. Revenue flows into the Income Statement. Operating Expenses feed the Break-Even Analysis. The break-even point is calculated with the formula =D2/'Key Assumptions'!$B$8, resulting in $54,733.23. The Business Plan references this precise figure.
When Intent 2 was written, the figure wasn't available. The Architect outlined inputs such as a 45–55% gross margin range, $3,500 monthly NNN rent, eight employees at the standard compensation rates, and a $250,000 startup budget. They also set a constraint that break-even should be achieved between months 10 and 14. However, the specific break-even threshold was not specified, as it only emerged after building the Financial Model and running the calculations in Excel.
When the Manager planned subsequent runs in Intent 2 (including the Business Plan, Pitch Deck, and Executive Summary), the run bundles specified Financial-Model.xlsx as a necessary input. According to global policy, the Developer was required to read the model and reproduce its outputs exactly as computed. The $54,733.23 break-even point in the Business Plan is not a figure that was shared via chat or stored in the Manager's memory. Instead, it is a number read by Claude Code directly from cell B18 on the Break-Even Analysis tab and quoted verbatim.
This feature makes the Financial Model a functional tool with meaningful formulas. When a key assumption input is changed, the model recomputes accordingly. While the Business Plan might then differ from the model's new outputs, the model remains internally consistent with the updated figures. To reflect these changes throughout the documentation, the Architect would create a new Intent to update all impacted downstream documents. The original Intent 2 stays unchanged. The new Intent's Run bundles would specify the updated Financial Model as a required input, and the Developer would access its current values during execution. These dependencies are maintained through structural file references mandated by global policy, not pasted values, making the update process manageable: since nothing is hardcoded, creating a new Intent with the updated model results in an updated document set.
What Trail produced
Seven intents, forty-nine deliverables.
- Intent 1 — Formation and Brand: Brand Style Guide, Cap Table Summary, Operating Agreement.
- Intent 2 — Strategic and Investor-Facing Documents: Executive Summary, Business Plan, Pitch Deck, Financial Model, Risk Assessment, Competitive Analysis, Marketing Plan, KPIs, and Reporting Framework. The Manager correctly decomposed this intent into three runs because the Financial Model had to exist before any document referencing it could be built.
- Intent 3 — HR, Employment, and People: Texas-compliant Employee Handbook, Anime Product Knowledge Guide, four role-specific job descriptions, two offer letter templates, and New Hire Onboarding Checklist.
- Intent 4 — Operations SOPs: Eleven standard operating procedures covering store opening, closing, cash handling, POS usage, returns, customer service, inventory receiving, visual merchandising, loss prevention, safety, and vendor consignment.
- Intent 5 — Finance and Compliance: Chart of Accounts, Accounting Policies, Inventory Valuation Policy, Cash Management Policy, Record Retention Policy, Data Security Policy, and Annual Compliance Checklist.
- Intent 6 — Vendor and Supply Chain: Master Vendor List, PO Template and Policy, Receiving and QC Guidelines, Consignment Agreement Template.
- Intent 7 — Customer, Marketing, and Community: Customer Privacy Policy, Loyalty Program Terms, Return and Exchange Policy, Website Terms of Use, Social Media Community Guidelines, Community and Events Plan.
What Trail demonstrated
Coordination relies on explicitly declared inputs and global policy rather than the Manager translating facts.
The Developer never reviewed intent.md or meta-level policy files. They only considered the inputs declared by the Architect and those surfaced by the Manager in the run bundle. Coherence emerged because Trail's standard policy on Cross-Document Coherence mandated exact-value reads from declared source documents, rather than requiring any contributor to maintain context between runs.
Emergent figures, structural references. Financial figures, operational thresholds, and policy details that did not exist at the time of Intent definition emerged during execution. Downstream documents read those emergent outputs via declared file-to-file references rather than pasted values. The chain is structural and re-runnable.
The immutability invariant remained valid during realistic load scenarios. Six intents finished in one cycle. Intent 2, aimed at producing eight key deliverables — such as a financial model and an investor pitch deck — was appropriately split by the Manager into three separate runs because the Financial Model had to be developed first, before the related documents. The Architect had foreseen multiple executions when defining the intent. All three runs were completed within the original scope. No Intents were altered after execution began.
The Standard Trail policy was sufficient. The Architect did not create custom rules for cross-document consistency. Trail's default global-operating-instructions.md already mandates exact-value reads from source documents. The Architect did not develop a fact-propagation mechanism. The reference data file located in trail/meta/files/ follows Trail's standard shared input pattern. The 7.5 hours of intent preparation went into correctly applying these patterns to a complex problem, rather than engineering custom infrastructure to make Trail work.
What this is not
The prepared documents are ready for review but not yet ready for shipment.
A genuine business launch would require legal and financial evaluations of the model's assumptions, as well as editorial reviews of customer-facing materials. Trail does not replace human judgment during the review process. Instead, it shifts production costs away from the review phase, allowing humans to focus more on their expert judgment.
The core idea is that judgment should be made at the input stage. The reference data indicates a gross margin range of 45–55%. If a real anime retail business operates at 38–45%, the Financial Model remains internally consistent with the incorrect margin, and all subsequent documents would remain internally coherent. However, the entire set of documents would still be incorrect. Human review at the input stage is where domain expertise is essential. After inputs are accurate, Trail's framework ensures proper propagation.
Why this matters
According to an independent benchmark from Perplexity, based on consultant pricing and document-drafting labor data, the realistic cost to produce a scope of 49 documents is roughly 100–150 hours, ranging from $15,600 to $38,500, with a median of $25,200.
The most significant expense items are the financial model, full business plan, pitch deck, employee handbook, and eleven SOPs — key deliverables that usually involve discovery interviews, outlining, drafting, revisions, and ensuring consistency across multiple specialists.
This illustrates how role separation, immutability, and artifact-focused handoffs shape multi-document knowledge work. Speed emerges as a byproduct. The core idea is that judgment should occur at the input stage, while coordination is handled through default policies. Enforcing these principles makes multi-document knowledge work more manageable compared to chat workflows or specialist coordination.
What Trail is and is not
Trail is not an SDLC. It does not substitute for Agile, Scrum, Shape Up, waterfall, or any other methodology used for prioritization, roadmap planning, or release management.
Additionally, Trail does not replace project management tools, source control workflows, or existing coordination systems.
Trail's opinions are limited to artifacts, roles, and handoffs during work execution. The underlying methodology that determines what is built and when remains unchanged. Trail focuses on ensuring the reliable execution of a specific work unit once the team has decided to proceed, regardless of the decision-making process used.
Production summary
Complete breakdown of effort, output, and role assignments for this build.
| Architect/PO planning, meta documents, and intent preparation | 7.5 hours |
| Architect orchestration + AI execution | 3 hours 16 minutes |
| Architect/PO Checkpoint A + B reviews | ~1.5 hours |
| Total Architect/PO effort | ~12.25 hours |
| Third-party benchmark for equivalent consultant work | 100–150 hours ($15,600–$38,500) |
| Approximate labor compression | ~10x at midpoint |
| Documents produced | 49 |
| Intents | 7 |
| Runs | 9 |
| Run bundles requiring revision at Checkpoint A | 0 |
| Outputs requiring revision at Checkpoint B | 0 |
| Intents revised after execution | 0 |
| Manager | Codex |
| Developer | Claude Code |
| Architect/PO / Reviewer | Single human |
| AI session continuity between runs | None |
Perplexity sources
Consultant pricing and document-drafting labor data used to calculate the equivalent third-party cost benchmark.
- Consulting Fees and Pricing in 2026: Hourly, Retainer, Fixed Fee
- Hourly Consulting Rate: 2026 Benchmarks & Calculator
- How to Set a Consulting Hourly Rate (And What To Do Instead)
- How Much Should I Charge as a Consultant?
- How Much to Charge as a Consultant (2026 Rate Guide)
- Business Consulting Firm Pricing Guide April 2026
- LLC Formation Timeline: State-by-State Guide
- Average Legal Expense for Small Businesses: A Guide
- C Corporation Startup Legal Fees
- Everything you need for a complete retail store cost breakdown
- How Much It Costs to Start a Retail Store in 2025
- How long does the process of forming a business in Texas usually take?