Maya Patel — craft posture

The prompt → generate → verify → correct loop, graded — teachable checks at each loop stage, each tied to value and a fix. Private, growth-oriented, trajectory-framed.

20of 27 checks passing

+7pts vs last 30 days· BLENDED — 4 measured + 3 proxy · never a single opaque 0–100 grade

The graded loop

The same loop HackerRank grades at hiring — measured continuously, on real work.

1Plan· 6/8 checks passing

Plan-mode disciplineFair

3/4

Skips plan mode on multi-file / risky changes

Faking it: Plan-mode events are logged commands — passing means actually planning; costs real work to fake.

Fewer wrong-direction reworks on risky codeMeasured

Prompt qualityStrong

3/4

Specifies target file, goal, and done-condition

Faking it: Occurrence-based today — content grading coming.

Fewer iterations → fewer tokens per outcomeProxy

2Build · Supervise· 5/7 checks passing

Waste patternsStrong

3/4

Right-sized models and strong cache reuse

Faking it: Occurrence-based today — content grading coming; waste-rule hits are measured, attribution is heuristic.

≈ 4.7M tokens/mo recoverable (est.)Proxy

Iteration efficiencyFair

2/3

Thrash loops — many tries, low acceptance

Faking it: Non-monotonic and measured: churn loops AND zero-verification single-shots both flag — shrinking it means actually converging.

Fewer tokens burned on loops with no outcomeMeasured

3Verify· 6/8 checks passing

Edge-case handlingFair

3/4

Accepts diffs on risky paths without touching tests

Faking it: Occurrence-based today (test-file touches) — content grading coming.

Risk avoided on high-blast-radius codeProxy

Verification disciplineFair

3/4

Accepts diffs with no verification between diff-proposed and accept

Faking it: Passing requires a real verify event between diff and accept — faking it costs an actual test run.

Catches defects before accept — the cheapest point to catch themMeasured

4Correct · Review· 3/4 checks passing

Execution qualityFair

3/4

Accepted diffs get reworked shortly after

Faking it: Post-accept churn is visible in later diffs on the same files — hiding it costs more than fixing it.

Cleaner diffs, less reworkMeasured

Findings

5 issues that actually waste tokens or add risk — ranked by value at stake

medium
Skips plan mode on multi-file / risky changesPlan-mode disciplineMeasured
Plan-mode discipline: 3 of 4 checks passing on eligible (complexity-weighted) sessions.
Fewer wrong-direction reworks on risky code11 sessions →
medium
Thrash loops — many tries, low acceptanceIteration efficiencyMeasured
Iteration efficiency: 2 of 3 checks passing on eligible (complexity-weighted) sessions.
Fewer tokens burned on loops with no outcome9 sessions →
medium
Accepts diffs on risky paths without touching testsEdge-case handlingProxy
Edge-case handling: 3 of 4 checks passing on eligible (complexity-weighted) sessions.
Risk avoided on high-blast-radius code7 sessions →
medium
Accepts diffs with no verification between diff-proposed and acceptVerification disciplineMeasured
Verification discipline: 3 of 4 checks passing on eligible (complexity-weighted) sessions.
Catches defects before accept — the cheapest point to catch them5 sessions →
medium
Accepted diffs get reworked shortly afterExecution qualityMeasured
Execution quality: 3 of 4 checks passing on eligible (complexity-weighted) sessions.
Cleaner diffs, less rework3 sessions →

Every check is complexity- and blast-radius-weighted, so trivial tasks never count. Proxy signals (prompt quality, waste patterns, edge-case) are coaching hypotheses, not hard failures.