Website of Peter Westerman

This series has documented a case study, extracted principles, and described a framework. All of that is useful context. None of it produces leverage until you do something with it. This post is about the doing — a practical 90-day roadmap for building a systematic AI collaboration practice from scratch, with specific milestones for each phase and the disciplines that produce compounding returns over time.

The roadmap does not require technical background. It does not require building software. It requires deliberate practice on a consistent schedule, willingness to invest in context infrastructure before the return is visible, and the discipline to complete the methodology fully rather than stopping at the point where you have a draft that seems acceptable.

Days 1–30: Foundation

Week 1: The Work Audit. Before applying AI collaboration effectively, you need a clear picture of what you actually do and which parts are most amenable to systematic AI assistance. Keep a daily log for one week. At the end of each day, note what you produced and categorize each output: production work (drafting, compiling, formatting, repeating a known pattern) or judgment work (deciding, evaluating, synthesizing, directing). Be honest — many tasks that feel like judgment work are actually production work with a thin layer of judgment applied at the end.

Most professionals who complete this audit honestly find that 30–50 percent of their time goes to production tasks that AI could handle at a draft level. That fraction is your leverage opportunity. It is also the work that, if AI can handle it at 80 percent of human quality in 10 percent of the time, creates a large and immediately applicable leverage ratio.

Week 2: One Task Type, Done Right. Select one recurring production task from your audit — ideally something you do weekly or more frequently. Spend this week using AI for that task systematically, following the five-phase methodology every time without shortcuts. Define specifically. Load context. Direct with bounded scope. Evaluate against the definition. Iterate with specific corrections. Do this for every instance of the task type this week.

By the end of the week, you should be completing this task type in 30–50 percent of the previous time with comparable or better output quality. If you are not, the issue is almost certainly in Phase 1 — the definition is not specific enough for the AI to produce what you actually need on the first pass. Sharpen the definition.

Weeks 3–4: Build the Context File. Using the five-section structure described in Post 8 Part 1, build a context file for the task type you have been developing. Purpose, conventions, constraints, current state, do not repeat. Spend two to three hours on this. Load it at the start of every AI session on this task type for the rest of the month. Update it at the end of each session with what changed.

At the end of 30 days: one systematically mastered task type with a maintained context file, direct experience with the five-phase methodology, and a baseline leverage ratio measurement for that task type. This is the foundation.

Days 31–60: Expansion

Apply the same systematic approach to two to three additional recurring task types. For each: work through the five-phase methodology until fluent, build a context file, use it consistently. By day 60 you should be running three to four task types systematically. Track time before and after on at least two of them — the measurement is motivating and informs where to invest next.

Build your first template library. Identify two to three output templates that appear in your most common task types. One page of structure per template is sufficient. Build them, load them as context when the relevant task type comes up, and refine them as you discover what is missing or what could be improved. A template loaded as context produces more consistent output than a template described in the session request.

Start the specialist profile model. Based on the task types you are now handling systematically, identify the one or two that would benefit most from a dedicated specialist context profile — deeper, more thoroughly calibrated than the general context files from month one. Build those profiles and use them instead of general context files for those task types. The investment is two to four hours per profile. The return is immediately visible in first-pass output quality on complex specialist tasks.

At the end of 60 days: three to four mastered task types, measurable time savings on at least two, a template library in early form, and specialist profiles for your highest-value tasks. You are operating at a systematically higher leverage ratio than you were 60 days ago.

Days 61–90: Compounding

Establish the maintenance rhythms that keep your infrastructure current and valuable. Update context files after every project that changes the relevant context. Update templates when better approaches emerge. Review specialist profiles quarterly and update when the underlying context has changed significantly. These commitments require approximately two to three hours per month total. Without them, the infrastructure degrades gradually from accurate and useful to stale and misleading. With them, it remains current and compounds its value with each session that uses it.

Measure your leverage ratio at day 90. Take two or three representative tasks you were doing manually 90 days ago and are now doing with AI. Compare how long each takes now versus before. A well-executed 90-day adoption typically produces leverage ratios of 2:1 to 5:1 on systematically developed task types. Across a full work week with 30–40 percent of time on AI-assisted tasks, this translates to 15–30 percent more total output from the same hours — or the same total output in significantly fewer hours.

Identify the next infrastructure investment. By day 90 the highest remaining leverage opportunities should be visible. What specialist profile would most improve the next most important task type? What template would save the most time if it existed? What context file is most outdated and degrading session quality? These questions drive the next infrastructure investment cycle — which does not end. It is a continuous discipline, not a project with a finish line.

The Practice Beyond 90 Days

The 90 days establishes a foundation and demonstrates what systematic practice produces. The following months deepen and extend it. The leverage ratios that were 2:1 to 5:1 at day 90 will continue to improve as the methodology matures, the context infrastructure becomes more comprehensive, and the direction skill develops through continued practice. The professionals in this series who started with three-week build cycles for new products were running five-day build cycles within twelve months — the same five-to-six-fold improvement available to any professional who builds this practice consistently.

The compounding is the point. Start small, on one task type, this week. Build the foundation in the first 30 days. Expand in the second 30. Establish the maintenance rhythms in the third 30. Then continue. The leverage follows, consistently and measurably, for as long as the practice is maintained and developed.

What to Do When Progress Stalls

The most common stall points in a 90-day adoption, and the specific actions that address each one:

Stall at Week 2: AI output quality is not improving despite following the methodology. Root cause: almost always Phase 1 (Define). The definition is not specific enough for the AI to produce what is actually needed on the first pass. The fix: take one of the outputs that did not meet your standard and analyze specifically what it would have needed to include to be what you needed. Then articulate the definition that would have produced that output. That articulation is the definition you should have used. Apply it to future sessions of this task type.

Stall at Day 45: Three to four task types are producing good results but improvement has plateaued. Root cause: usually Phase 2 (Contextualize). The context files are not capturing the nuance that distinguishes good output from great output in your domain. The fix: spend thirty minutes reviewing your current context files and asking what is missing that a specialist in your domain would know. Add what is missing. The output quality improvement from a well-developed context file is non-linear — the last ten percent of context completeness often produces a disproportionate share of the output quality improvement.

Stall at Day 75: Progress has slowed because maintenance is taking too much time. Root cause: context files and templates have grown unwieldy, requiring significant time to read and update. The fix: simplify. Remove content that is no longer relevant. Reorganize for faster lookup. The context file is most valuable when it is accurate and concise, not when it is comprehensive. A five-hundred-word context file that is kept current is more valuable than a five-thousand-word context file that is maintained at the level of the least interesting detail.

The Mindset Shift That Accelerates Everything

Beyond the specific practices in the roadmap, one mindset shift consistently accelerates the adoption trajectory: treating every AI session as an investment in future sessions, not just as a means to complete the current task.

A session that produces a good output and then stops is a linear return. A session that produces a good output, then uses five minutes to update the context file with what made this session work well, then adds a useful phrase or structure to the template library, then notes in the specialist profile that a particular approach worked especially well — that session produces a compounding return. The future sessions that reference those updates will produce better outputs faster because of the five minutes invested in the update.

This mindset — each session as an investment in all future sessions — is what separates professionals who are still producing at roughly linear improvement rates eighteen months into AI adoption from those who are operating at five-to-ten-times leverage. The linear improvers treat each session as an isolated transaction. The compounding practitioners treat each session as a contribution to an infrastructure that makes all future sessions better. The 90-day roadmap builds toward the compounding mindset. Applying it explicitly, from day one, accelerates everything.

How Recent AI Innovations Change This Picture

The 90-day roadmap described in this post assumed a specific tooling landscape. With recent AI innovations, some steps on the roadmap accelerate significantly, and the outcomes achievable by day 90 are more ambitious than they were when this roadmap was originally developed.

The context infrastructure phase — weeks 3-6 in the original roadmap — was the most time-intensive phase because it required building custom context loading systems, CLAUDE.md files, and template libraries from scratch. Agent Skills change this phase substantially. Building shared skills from the start, using the platform’s native infrastructure, reduces the setup time and increases the portability of what is built. What took three weeks of custom infrastructure work can now be done in a few days using the Skills framework.

MCP server setup — integrating Claude with the professional’s existing tools — is now a more documented, more standardized process than it was in 2024. For common professional tools (Notion, GitHub, Jira, Slack, databases), MCP servers are available as open-source or commercial packages with standard configuration. The integration phase of the roadmap, which previously required evaluating and sometimes building custom integrations, is now primarily a configuration task for common tool stacks.

By day 90 under the original roadmap, a professional could expect to be running a mature vibe coding or AI collaboration practice with meaningful leverage ratios on their core work. With current tooling, the same 90 days can produce a multi-agent capable, MCP-integrated, Skills-based practice — a more sophisticated and more deeply integrated AI collaboration system than was achievable in 90 days with 2024 tooling. The starting point for what “mature practice” means has moved up significantly.