ScopeSmart - AI-assisted project scoping for agencies

ScopeSmart was an AI-assisted scoping tool for digital agencies and consultancies. From a one-paragraph project brief, it produced a scope document, a phased plan with estimates and team assignments, and a critical-path timeline.

The idea came from my years on the agency side, writing scopes and proposals for pitches that often didn't win. That work is unpaid and repetitive, and by 2024 LLMs were good enough to take a real share of it.

Challenges

In 2024, a single prompt couldn't reliably turn a vague brief into a usable plan. Quality dropped as the task grew, and the model would invent requirements or produce estimates no one could defend.

Agencies also need to own the scope they sign off on, so the tool had to show its assumptions and let people correct them before building on top.

Solution

I built ScopeSmart as a chain of LLM steps, where each step's output becomes the next step's input and the user confirms along the way:

1. Clarifications: the brief is checked for gaps (stakeholders, objectives, dependencies, risks, success criteria), each returned with a suggested default the user can accept or edit.
2. Scope document: a structured scope with inclusions, exclusions, deliverables and constraints, editable in a markdown editor.
3. Requirements: a categorized requirements list, written as input for the planning steps.
4. Phases, milestones and tasks: a structured plan, with each task assigned to a role on the project team.
5. Estimates and dependencies: a three-point (PERT) estimate per task, plus an inferred dependency graph between tasks.

The timeline uses no LLM. It combines PERT expected durations, a critical-path pass over the dependency graph and per-person resource leveling, shown as a Gantt view and exportable to Excel for the client.

Prompt quality was tested separately with promptfoo. For a period, ScopeSmart also ran as an embedded partner app inside the LiftOS project-management platform.

Outcome

Breaking generation into small, reviewable steps gave far more reliable output than one large prompt, and letting users correct assumptions early kept errors from carrying into the plan. Keeping the numbers out of the model, with estimates feeding a deterministic schedule, made the timeline something an agency could actually defend.

ScopeSmart ran from 2024 to 2025, with over 200 sign-ups, and has since been retired.

Building an AI product? Let's connect.