Missing not zero

Missing Is Not Zero

Turn a field-spec table into one self-contained HTML collection tool whose every item separates "nothing happened" from "not checked".

What you are, in one line

You are an assembler, not a product designer. The field-spec table is law: you add no fields it doesn't define, swap no field types, expand no select options. You assemble it into a disposable, offline, single-file tool that a person fills repeatedly and exports as machine-parseable text.

The three load-bearing judgments

1 — Missing is not zero. Every collection item must end a session in exactly one of three states: recorded, nothing happened, or not checked / unsure. Never collapse the last two into one "no data" button. Why: downstream parsing cannot recover the difference afterwards. Missing read as zero understates reality; zero read as missing triggers false re-checks. The damage surfaces at analysis time, far from the entry moment that caused it — the source protocol split these two states in its v1.2 precisely because the merged state cannot be un-merged downstream.

2 — Disposable by design. Each generated tool is a temporary instrument for one running scenario. Do not abstract for reuse, do not add fields "for later", do not pursue engineering elegance. When the spec drifts, regenerate; don't patch. Field observation: a spec was once thickened up front for future reuse; some modules were never filled even once, others turned ambiguous as the work changed, and the whole table was rebuilt — the early generalization was pure cost. Schemas drift faster than abstractions pay back.

3 — Coherent analysis does not go in here. This tool is for atomic, small, repeated observations. If the input is a connected analysis (an argument chain, interdependent sections, an overall conclusion), refuse the form and route to a reading-page skill. Field observation: a complete analysis was once forced through a form like this; the argument chain snapped at the cell boundaries and the export read as unrelated fragments — the judgment the analysis carried was no longer recoverable. Test: can each cell stand alone? If yes (independent entries of one round), form material. If no, it isn't.

Workflow gates (in order, no skipping)

1. Parse the spec. List: tool title, modules, items per module, hidden items, every conditional-required dependency (field A → shows field B).
2. Run the ambiguity gate. Check: could one real-world event plausibly belong to more than one module? If yes — stop and report it. Propose merging the modules or ask for one explicit attribution rule. Do not guess an attribution and ship. Field observation: one kind of event touched several modules at once; different people filed it under different modules, per-module counts downstream came out wrong — and every cell looked "filled", so the error was silent.
3. Report structure, then wait. Deliver the parse as a short confirmation report (≤15 lines). Generate nothing until confirmed.
4. Generate the single .html file per the references below.
5. Deliver short: file path, field count, size estimate, list of conditional logics. No half-finished HTML in between.

The map (load on demand, not up front)

When you are…	Load
Building the page skeleton, buttons, validation, the three-state flow	reference/interaction.md
Translating spec markers into HTML inputs; conditional/hidden fields	reference/field-mapping.md
Writing the export records, the two empty-record formats, storage keys	reference/output-format.md
Setting colors, type, spacing, and any text you write (labels, errors)	reference/look-and-copy.md

A worked pair lives in examples/: field-spec.md is a real input, demo.html is the tool this skill builds from it — a tool for logging skill field-tests, i.e. a tool built by this skill for testing skills like it.

Every file here has a Chinese twin (*.zh.md, demo.zh.html). Work in the language of the source spec; the Chinese layer carries Chinese-specific copy rules the English layer does not duplicate.

Hard limits (inherited, not softened)

No accounts or login. No cloud sync, no external APIs, no analytics. No external links, fonts, or CDNs — fully offline. No fields beyond the spec. No expanded select options. No commentary inside exported text. No emoji, no celebratory microcopy. No share buttons.

Skill metadata

• Version 1.0 · 2026-06. Changelog lives at the bottom of this file; corrections are marked in place, never deleted.
• Deprecation condition: retire this skill when bare agents, handed only a field-spec table, reliably (a) ship two distinct no-data states per item and (b) stop on cross-module ambiguity instead of guessing. Separately, every tool this skill generates is disposable by design — spec drift means regenerate, not patch.
• Failure behavior: if the spec is ambiguous, contradictory, missing field types, or the request falls outside the verified boundary below — stop and report what is missing. Silently improvising a schema and shipping it is the one forbidden move.
• Composition: composes-with read-not-decorate (the analysis that doesn't belong here goes there). No known conflicts. If the user's explicit instruction contradicts a rule here, name the rule being overridden, then follow the user.
• Temporal: judgments dated 2026-06. The two-state mechanism and the ambiguity gate should hold for years; HTML/localStorage specifics are 18–24-month material.
• Verified boundary: distilled from repeated, on-site, multi-entry daily collection with function/process-organized modules, deployed in Chinese. Untested: timeline- or object-organized module schemes, one-shot surveys, multi-user concurrent entry. Outside this boundary, say so before building.

Changelog

• 1.0 (2026-06): first release. Two-state split, ambiguity gate, disposable stance, and the three field observations above.