Weighted Planning and the 90-Day Loop
How the framework becomes a form and the form becomes a plan - per-sub-discipline ratings with evidence, a 500-point ceiling, a gap-times-weight priority formula and a 90-day goal loop, walked through with a worked example.
Part 10 closed the tour of the disciplines: eight for the AI era, seven for the traditional craft, every weight argued for. Which leaves the question a framework must answer or admit to being decoration. What do you do with it on a Tuesday? My answer is a self-assessment template — one per framework, deliberately mirrored — that turns the weighted model into a development plan in five moves: rate, evidence, score, prioritise, commit. This part walks the template end to end, then fills it in for an invented engineer so you can watch the arithmetic produce a plan.
A form with opinions
The template opens with an identity block — assessor, role, career stage from Junior through Mid, Senior and Staff to Principal, assessment date, review period, and a field that matters more than it looks: the name of a peer reviewer. We will come back to that field, because the whole design assumes it is filled in.
Then the rating key. Every sub-discipline is rated 1 to 5, and the numbers are not vibes — they are the five levels from part 6, compressed to a sentence each:
| Rating | Level | Meaning |
|---|---|---|
| 1 | Novice | Understand conceptually - cannot apply independently |
| 2 | Apprentice | Apply basic techniques with guidance - handle routine situations |
| 3 | Practitioner | Apply independently - create reusable patterns, support others |
| 4 | Strategist | Lead and design at team level - resolve novel problems |
| 5 | Architect | Set direction at organisational scale - advance the practice |
The body of the form is one table per discipline. Each row is a sub-discipline carrying its sub-weight and its absolute weight — the two-level arithmetic from part 1 pre-computed, so Knowledge Architecture & Curation shows its 4.50 and Example Curation & Few-Shot Design its 1.60 right next to the empty rating cell. That adjacency is a design decision. You should know what a row is worth before you rate yourself against it, because the weight is the model telling you where an honest 2 actually hurts. Each discipline table ends with a reflection prompt — not decoration, but a trap for inflated ratings: if you cannot answer “when did you last create a reusable prompt template that others adopted”, your Prompt Craft rating has just been argued down a notch by your own silence.
Evidence or it does not count
Every rating row has an evidence field, and the template's instruction is the strictest sentence in it: concrete artefacts, not assertions. “Designed the context architecture used across three projects” is evidence. “Good at context engineering” is a mood. The discipline here is borrowed from how the progression levels were built in part 6 — observable evidence or it did not happen — and it exists because the form's biggest enemy is the person filling it in. An evidence field you cannot populate is the template quietly telling you the rating above it is wrong.
The arithmetic of a scorecard
Scoring is one multiplication per row: rating times absolute weight. Sum a discipline's rows and you get its weighted score; sum the disciplines and you get the overall score. Because the absolute weights across a framework total 100 and the maximum rating is 5, the ceiling is a fixed 500 for every person, every team and every year — which is what makes scores comparable across all three. A score means nothing in isolation and something quite precise in context: the software template's scoring guide makes the point that a Senior engineer at 300 is on track while a Junior at 300 is exceptional. The number does not judge you; the number plus your career stage does.
The summary section renders the same data twice more — a radar of discipline averages against the expected profile for your career stage, and, in the AI template only, the stack profile from part 9 with your rating slotted into each layer, so a weak foundation under a strong apex is visible as a picture rather than an inference.
Meera fills in the form
Time to make it concrete. Meera is invented — a senior engineer, strong on the traditional framework, assessing herself against the AI framework for the first time. Here is her Specification Engineering table as she filled it in:
| Sub-discipline | Abs. weight | Rating | Weighted | Evidence |
|---|---|---|---|---|
| Problem Decomposition & Task Design | 4.40 | 3 | 13.20 | Breaks features into agent-sized tasks routinely |
| Acceptance Criteria & Definition of Done | 4.40 | 2 | 8.80 | Criteria written, but reviewers still ask questions |
| Constraint Architecture | 4.40 | 2 | 8.80 | Musts and must-nots, ad hoc, not systematic |
| Self-Contained Problem Statements | 4.40 | 2 | 8.80 | Tasks still assume shared context |
| Evaluation Criteria Embedding | 4.40 | 1 | 4.40 | Has never embedded evaluation in a spec |
| Discipline total | 22.00 | avg 2.0 | 44.00 |
Work breakdown is the muscle she brought over from years of delivery practice; embedding evaluation criteria is a skill she has literally never exercised. Her full scorecard, discipline by discipline:
| Discipline | Weight | Avg rating | Weighted score |
|---|---|---|---|
| Specification Engineering | 22 | 2 | 44 |
| Context Engineering | 18 | 3 | 54 |
| Intent Engineering | 15 | 2 | 30 |
| Agent Architecture & Orchestration | 12 | 2 | 24 |
| Evaluation & Quality Engineering | 10 | 3 | 30 |
| Prompt Craft | 8 | 4 | 32 |
| AI Safety, Security & Governance | 8 | 2 | 16 |
| Domain Translation & Systems Thinking | 7 | 3 | 21 |
| Total | 100 | 251 |
251 out of 500, and the shape is more informative than the total. Her single highest average is Prompt Craft — the most visible skill, the one the framework prices at 8 — while her lowest scores sit on the heaviest disciplines. Meera has optimised for the discipline she can see herself performing. The template's scoring guide calls this pattern out by name, because it is the most common profile the framework meets: if your Prompt Craft is your peak, your investment does not match the model's opinion of where leverage lives.
Gap times weight
Now the move that makes the whole apparatus worth the form-filling. For each sub-discipline, take the gap between current rating and target rating — Meera, as a Senior aiming at Strategist-level practice in her highest-leverage rows, sets targets a level or two up — and multiply the gap by the row's absolute weight. Highest product first. That single formula is the development plan.
| Sub-discipline | Current | Target | Gap | Abs. weight | Weighted gap |
|---|---|---|---|---|---|
| Acceptance Criteria & Definition of Done | 2 | 4 | 2 | 4.40 | 8.80 |
| Evaluation Criteria Embedding | 1 | 3 | 2 | 4.40 | 8.80 |
| Knowledge Architecture & Curation | 3 | 4 | 1 | 4.50 | 4.50 |
Notice what the formula quietly refuses to do. Meera's Example Curation & Few-Shot Design sits at 4 — even if it sat at 2, a three-level climb there would be worth 4.80, still less than a two-level climb on either specification row. The formula will not let her polish the light rows she already enjoys, and it will not send her after every gap equally, which is the failure mode of the unweighted grid this series opened on. Gaps are everywhere; weighted gaps have an ordering.
Three goals, ninety days
The template caps the plan at one to three goals for the next ninety days, each with success criteria and a target date. The cap is the point — a development plan with nine goals is a wish list wearing a spreadsheet. Meera's three, straight from her priority table:
- Specifications with verifiable acceptance criteria — for the next three delegated features, write the specification before any agent runs; success is a reviewer confirming each one is executable without a single clarifying question.
- Evaluation embedded, not appended — every one of those specifications defines its quality gates and measurement method inside the document; success is one run where the output was accepted or rejected against the embedded criteria alone.
- A personal eval folder — twenty prompts with expected, judged results from her real work, run against the next model release; success is a written keep-or-switch decision backed by the results.
Ninety days later she reassesses, the scores move or embarrassingly do not, the priorities re-rank, and the loop turns again. Two or three turns of that loop generate something the first pass cannot: a trend line.
The second signature
The last section of the template belongs to someone else. Calibration notes: a peer or manager reviews the ratings against the evidence and records their own calibrated rating next to each self-rating, with room for the disagreement. The template carries an explicit warning here, and part 7 spent its whole length on why — in METR's randomised trial, experienced developers forecast a 24 per cent speed-up from AI tools, actually took 19 per cent longer, and afterwards still believed they had been sped up by about 20 per cent. Perception of one's own performance with these tools is not merely noisy; it can point the wrong way entirely. A self-assessment that no second person has challenged is a draft, and the design treats it as one — the form is not finished until both signature lines are.
That is the full loop: rate against levels, defend with evidence, score against 500, rank by gap times weight, commit to three goals, calibrate, repeat in ninety days. What the loop cannot do — what no scorecard can — is the subject of the final part: what a scorecard cannot see.