A Plan You Can Grade
PREVC's Planning phase has the agent decompose the work the way project managers always have, then do something no methodology asked of it - grade its own plan. The work breakdown structure, the fifteen-minute waterfall, and the three scores I added.
Part 1 made the case that agent work needs gates, and named the first phase Planning without saying much about what happens inside it. Here is the short version: the agent turns a goal into a structured, hierarchical plan in minutes — and then does something I have never seen a methodology ask of a human planner. It grades its own plan, and hands you the grades. That second half is the part I added to the borrowed skeleton, and it is the part that changes how the next phase spends its time. But the first half deserves its due, because it is a very old discipline running at a very new speed.
Everything in, before anything out
Planning starts with a gathering, not a generating. Every requirement, every constraint, every relevant convention and prior decision is assembled and fed into the agent's context before it produces a word of plan. This ordering is not pedantry — it is the entire difference between a plan and a guess. An agent plans from what it can see; whatever you leave out of its context, it will fill in from the general shape of similar projects, and the general shape of similar projects is precisely what your project is not.
With the material in place, the agent enters plan mode and produces a work breakdown structure: a hierarchical decomposition of the deliverable into work packages small enough to execute, verify and hand around independently. If that term sounds like it predates AI by half a century, it does — and the lineage is worth a moment, because the discipline transfers almost untouched.
A discipline older than the agents
The work breakdown structure is project management's oldest reliable tool. The Project Management Institute's PMBOK Guide defines it as “a deliverable-oriented hierarchical decomposition of the work to be executed by the project team” — the PMI's practice standard is the canonical treatment. The operative word is deliverable-oriented: a WBS decomposes what will exist when you are done, not the activities you will perform along the way, and each element splits into progressively smaller components until every leaf is estimable and assignable. On depth, the defence-programme handbooks that first standardised WBS practice — MIL-STD-881 and its descendants — treat three levels as the sensible default, going deeper only where the cost or risk of an element justifies the extra resolution. That heuristic transfers directly to agent plans: a breakdown that is uniformly six levels deep is not thorough, it is procrastinating.
What changed is not the discipline but the price. A WBS for a real system used to be weeks of workshops; an agent produces a credible first draft in minutes. Addy Osmani, describing his own planning workflow, quotes Les Orchard's phrase for what this feels like: a “waterfall in 15 minutes” — the rigour of upfront planning without the months that made waterfall a cautionary tale. Osmani's version compiles the results of an iterative question-and-answer session into a structured spec file holding “requirements, architecture decisions, data models, and even a testing strategy”. The point both of them are circling is the one this whole phase rests on: upfront structure was never the thing that failed about waterfall — the feedback delay was — and the delay is gone.
The three scores
So far, everything in this part is inherited or credited. Now the extension — and to keep the ledger from part 1 honest: the grading system below is mine, nothing upstream grades its own plan, and its faults are mine.
A fast plan is not the same as a ready plan, and the gap between the two is exactly what the Review phase exists to close. The problem is that the gap is invisible from outside. The agent knows which of its work packages it genuinely understands and which it papered over — but that knowledge is trapped in the model unless you demand it. So alongside the breakdown, I have the agent produce three self-assessment scores. They are not evaluations of the project. They are evaluations of the plan's readiness for execution.
| Score | What it measures | What it tells the reviewer |
|---|---|---|
| Clarity | How unambiguous and actionable each work package is - could the agent execute it without asking further questions? | Where to add detail, examples or constraints |
| Completion | How comprehensive the coverage is - are there missing requirements, unhandled failure paths, unaddressed constraints? | What to add or surface before execution |
| Complexity | How entangled the packages are - can they decompose further, and are there hidden dependencies? | Where to split, simplify or resequence |
The scores earn their keep in combination. High clarity, high completion, high complexity says: the requirements are understood and nothing is missing, but the breakdown is too entangled for an agent to execute reliably — decompose further before anything runs. Low clarity with low complexity says the opposite: the structure is fine, the descriptions need precision. Low completion is the one that should stop you cold, because it is the agent telling you it knows there are requirements it never saw.
What the scores actually correct is an information asymmetry. Without them, the human reviewer must discover the plan's weak points by careful reading — expensive, slow, and biased towards the sections that happen to look wrong. With them, the agent's private knowledge of its own uncertainty becomes the reviewer's map. The plan arrives pre-annotated with “look here first”.
What a graded plan looks like
Concretely, for an invented project: Rosterly, a small service that builds staff rotas from availability and demand forecasts. The task is adding shift-swap requests. A trimmed slice of the graded breakdown:
# Plan: shift-swap requests for Rosterly
1. Swap-request model and lifecycle [clarity 9, completion 8, complexity 3]
1.1 Request entity: requester, target shift, offered shift, state
1.2 State transitions: proposed -> accepted -> applied / declined / expired
2. Eligibility rules [clarity 5, completion 4, complexity 7]
2.1 "Qualified for the shift" check - source of qualification data unclear
2.2 Rest-period and max-hours rules - which regulations apply is unstated
3. Notification of affected staff [clarity 8, completion 7, complexity 2]
4. Apply the swap to the published rota [clarity 7, completion 6, complexity 8]
4.1 Entangled with rota versioning - depends on how amendments are stored
Read the scores before the packages and the review writes its own agenda. Package 1 is nearly ready — a light pass. Package 2 is a flare: the agent is saying, in numbers, that it does not know where qualification data lives or which working-time rules bind, and that guessing would be irresponsible. Package 4's complexity score is the quiet warning that a dependency on rota versioning is about to entangle two work streams. Fifteen minutes of reading, aimed by the agent's own confession, replaces an hour of undirected scrutiny.
Notice also what the scores are quietly doing to the workflow's economics. The gate between Planning and Review no longer costs a full plan-read; it costs a glance at three numbers and a decision about where to dig. That is what makes the routing from part 1 affordable at MEDIUM scale and above — a reviewed plan per feature is only sustainable when the review starts pre-aimed.
One caution from practice, so the scores do not get worshipped: they are the agent's self-assessment, and self-assessment can be miscalibrated — a plan can score itself clear because it is confidently wrong about what you meant. I treat the scores as a routing signal, never as approval. A high-scoring plan still passes through human judgement; it just passes through faster. The grades direct attention; they do not replace it. Over successive cycles you also learn your agent's grading habits — where it flatters itself, where it is oddly humble — and that calibration knowledge is itself worth writing down.
Planning ends with a graded, structured breakdown sitting at a gate, waiting for the one thing no score can supply: judgement about what the organisation actually wants. That arrives in the next phase — and it is where the whole workflow is won or lost. Next, Review is the most important handoff.