notes-v3.6 — Revision Notes Generator: Verified Coverage & Source Transparency
Complete rewrite of the revision-notes-generator skill, driven by a systematic diff of generated outputs against their sources (PMT Mechanics & Materials, SME Newton's Laws, SME Classification of Particles). The previous version produced accurate notes but with three hidden failure modes: coverage varied between runs, knowledge was added from outside the source without any marker, and the accuracy-verification instruction produced no observable behaviour. This release makes all three visible and testable.
What's New
Coverage map and mandatory verification
The skill now builds an explicit checklist of every subheading, method, equation, and condition in the source before writing, and runs a mandatory verification pass against it before output. This closes the run-to-run drift where method steps (e.g. resolving projectile velocity components, the maximum-height condition) appeared in one run and vanished in the next.
Flagged beyond-source additions
Adding correct knowledge beyond the source is permitted when needed for complete understanding, but every addition must now be marked inline:
(Beyond source: the added fact.)
Previously, silent additions could contradict nearby source content. In testing, an unmarked insertion produced notes stating both "strong force → gluon" and "pion is the exchange particle of the strong force" with no reconciliation.
Error flagging instead of silent correction
Source statements that conflict with standard A-Level treatment are kept and flagged:
⚠ Check: source states X; standard treatment is Y.
The old "correct any errors" instruction never visibly fired and risked invisible rewrites. In validation, the new behaviour caught a genuine error in a PMT source (a duplicated specification section number) on its first run.
Worked example policy
Worked examples that teach reusable methods (conservation-law checks, SUVAT strategy, quark deduction, equilibrium proofs) are compressed and retained inline. Pure number-substitution examples are dropped. Previously this was unspecified, so examples survived or vanished at random.
Signal density replaces length-based conciseness
Conciseness is now defined as every sentence defining, explaining, or connecting a concept, with no length cap. Long, dense sources produce long, dense notes.
Self-contained output
Diagram content must be translated into prose or tables. "See source" references are banned. In validation, Feynman-diagram-only content (kaon production via gluon, decay via W boson) was correctly extracted into text.
Conditional tier labelling
(HT) marking now applies only to GCSE sources. A-Level sources have no tiers, so the previous always-on instruction was dead weight.
Output Changes
Notes generated by the updated skill differ visibly from previous releases:
*(Beyond source: ...)*and⚠ Check:markers may appear in the body- Source section numbering is retained in headings for spec cross-referencing
- Equations are rendered in LaTeX rather than code blocks
- Compact inline worked examples appear throughout
- Outputs are typically ~40% longer due to retained methods and examples
If you post-process these notes (e.g. into flashcards), account for the new marker syntax.
Validation
Tested on two source families with output-vs-source diffs:
- PMT Mechanics & Materials: previous coverage gaps closed, one genuine source error caught and flagged, two legitimate beyond-source additions correctly marked
- SME Classification of Particles: all five method-teaching worked examples retained, Feynman diagram content translated, one false-positive flag traced to a source-specific example embedded in the skill (fixed by replacing it with a generic placeholder)
Known Limitations
- Source-internal contradictions (the source disagreeing with itself across sections) can occasionally pass the verification pass
- The worked-example filter occasionally retains a trivial number-substitution example
- Flagging thresholds are conservative: muddled-but-defensible source phrasing may go unflagged
Skill-Writing Lesson
Never embed real content from the target source corpus as an illustrative example inside a skill. During validation, the model pattern-matched an embedded gluon/pion example and mis-flagged that exact content as beyond-source when it actually appeared in the source's tip box. Examples in skills should always be generic placeholders.