STOW

Focused writing governance for LLM output: anti-synthetic prose discipline, action-oriented responses, protected literals, and controlled technical writing when the task requires it.

STOW (Standardising Technical Output Writing) is a public writing specification and packaged Agent Skill. It helps a model produce clear user-facing prose, procedures, coordination artefacts, and structured output without turning every task into controlled English or an expensive validation workflow.

STOW is not an AI-authorship detector, a generic “sound human” style randomiser, or a universal final-response enforcement layer. Most of its writing rules are contextual guidance. Its callable tools decide only closed properties at their declared input boundaries.

Current release: v0.4.1.

Why STOW?

A repository instruction that only names a controlled-language standard can be a useful and inexpensive model-memory cue. It still leaves the model to reconstruct the standard, decide what applies, remember the dictionary and exceptions, preserve project terminology, and distinguish contextual judgement from mechanical checks.

STOW externalises that work and adds a broader writing policy:

Need	A name-only instruction	STOW
Anti-synthetic prose	Outside the named controlled-language standard	Context-bounded guidance for filler, hollow evaluation, fabricated specificity, formulaic transitions, needless process narration, and related pathologies
Focused output	Depends on the model's general habits	Action shaping, result-first reporting, bounded steps, tangent control, and explicit state
Controlled technical writing	Reconstructed from latent knowledge	Issue 9-derived rules, sparse dictionary access, project terminology authority, profiles, and explicit limitations
Protected content	Ad hoc	Precedence guidance for code, commands, paths, identifiers, quotations, and data values
Mechanical checking	Usually unspecified	Closed validators for the few properties that are genuinely mechanical
Inspectability	The operative policy remains mostly latent	Rules, applicability, exceptions, profiles, conflicts, evidence boundaries, and generated catalogues are public

The measured single-task advantage over name-only conditioning was small, not dramatic. STOW's larger distinction is that the writing policy is versioned, inspectable, testable, and reusable instead of existing only as a sentence in context.

Contextual prose review distinguishes functionless closers, unsupported evaluation, manufactured contrast, mechanical symmetry, blanket hedging, and inflated metaphor from legitimate summaries, voice, correction, templates, uncertainty, and literal technical usage. Action shaping now distinguishes action-first work from explanation- or safety-first contexts, preserves exhaustive contracts, surfaces only changed progress, and defers secondary issues without dropping them. These remain contextual G1 behaviours.

Scope

STOW has related operating surfaces that load only as far as the task requires:

Surface	Primary job	What normally loads
Ordinary prose	Remove recurrent model-writing pathologies, preserve requested voice, and keep the answer focused	Compact kernel only
Technical and coordination prose	Add stable terminology, explicit conditions, evidence boundaries, bounded steps, and clear status	One matching cold reference when needed
Public-facing documentation	Preserve a correct reader model, evidence strength, optionality, current state, limitations, and functional structure	Existing technical-clarity profile plus one direct cold public-documentation reference
Controlled technical writing	Apply the supported controlled-language rules, dictionary records, project terminology, procedure, description, safety, punctuation, and counting guidance	Controlled profile plus bounded cold references; sparse lookup only when a lexical question arises
Structured or protected content	Preserve raw data and validate closed formats or schemas	Raw/protected mode or an explicitly invoked validator

STOW is suited to README and documentation work, technical explanations, maintenance procedures, safety text, repository status updates, agent handoffs, and structured artefacts. It can also govern ordinary conversational prose when the user wants a more direct and less synthetic response.

STOW is not intended to replace a creative style guide, infer project terminology without authority, certify strict controlled-language conformance, or guarantee that a host will select the skill automatically.

Design genealogy

STOW is an independent synthesis, not a dependency bundle or a one-to-one copy of any source.

Source	Contribution to STOW	Boundary
ASD-STE100 Simplified Technical English, Issue 9	The controlled-technical lineage: bounded vocabulary, technical nouns and verbs, sentence and procedure constraints, safety writing, descriptions, punctuation, and the principle of specifying observable linguistic behaviour instead of saying only “write clearly”	ASD-STE100 is owned and maintained by ASD. STOW is independent, is not endorsed by ASD, and keeps strict conformance locked
realrossmanngroup/no_ai_slop_writing_rules	Observable anti-synthetic patterns, worked contrasts, structural repetition, filler, hollow claims, dramatic headings, corrective contrasts, and false-positive safeguards	STOW does not import the excluded voice profile or treat statistical AI tells as universal bans; a pattern survives only when it identifies an independent writing problem and a legitimate countercontext
ayghri/i-have-adhd	General focus and runtime mechanisms: action-first output, visible progress, bounded steps, tangent deferral, explicit exceptions, progressive disclosure, and pre-send discipline	STOW adopts only those transferable output-design mechanisms; it does not add a source-themed mode or a universal step cap

The resulting product is broader than controlled technical language and narrower than a universal prose normaliser:

anti-synthetic prose discipline
+ focused action and status shaping
+ evidence and semantic preservation
+ controlled technical writing when applicable
+ closed validation only where the property is closed

What STOW changes

These examples are illustrative. They show the intended behaviour, not a guarantee that every model will make the same edit.

Ordinary prose: remove functionless framing

Before:

Great question! There are several angles worth unpacking here before we get to the answer itself.

After:

The build fails because the lockfile pins a version the registry no longer serves. Repin it, then rerun the build.

Focused work: expose the next bounded actions

Before:

First open the file, find the function, swap it out, then run the tests and check whether anything else broke.

After:

1. Open `src/auth.ts`.
2. Replace `verifyToken`.
3. Run `npm test -- auth.spec.ts`.

Controlled technical writing: apply the narrow route

Source:

Prior to commencement, utilize the approved bonding agent on SEAL_A; if ERR-17 occurs, terminate the operation.

Controlled-guided rewrite:

Before the operation, apply the approved bonding agent to SEAL_A.
If ERR-17 occurs, stop the operation.

STOW must preserve source force. Advice, permission, uncertainty, and an authorised command are not interchangeable merely because an imperative is shorter.

Protected output: leave the artefact alone

When the requested result is raw JSON, code, a command, a path, an identifier, or quoted text, lower-precedence prose preferences yield:

{"service":"api","status":"ok"}

Quick start

Install

With skills.sh

npx skills add theislampill/stow

This installs STOW through the Agent Skills ecosystem. Installation makes the skill available to a compatible host; host selection and activation remain separate.

From the release artifact

Download STOW.skill from the latest GitHub release.

Linux or macOS:

mkdir -p ~/.claude/skills
unzip -o STOW.skill -d ~/.claude/skills

Windows PowerShell (Expand-Archive requires a .zip suffix):

Copy-Item .\STOW.skill .\STOW.zip -Force
Expand-Archive -Path .\STOW.zip -DestinationPath "$HOME\.claude\skills" -Force

Build from source

Build the same package layout from this repository:

python tools/build_skill.py

Then use dist/STOW.skill in place of the downloaded STOW.skill in the extraction commands above.

The expected layout is:

<skills-dir>/stow/SKILL.md

These commands write or replace the stow directory under the selected skills directory. Back up local modifications before reinstalling; remove that directory to uninstall the skill.

The repository also includes Claude plugin manifests. Hosts that consume compatible skill or plugin packages can install from the repository instead of extracting the archive manually.

Invoke

Invoke STOW explicitly when host selection matters. For example:

Use STOW to rewrite this README section. Preserve every technical claim, command, path, identifier, and quotation.

For controlled technical work:

Use STOW under the controlled-technical-guided profile. Rewrite this procedure, preserve source force and protected literals, and do not claim strict conformance.

Automatic skill selection is host-dependent. Installation alone does not prove that a particular host or task loaded STOW.

Smallest working flow

1. Invoke STOW explicitly and name the output type or profile when it matters.
2. Supply the draft or task plus the facts, literals, terminology authority, and voice that must survive.
3. Review the returned candidate. Invoke a runtime helper only for a closed check that the task needs; advisory findings do not block delivery by themselves.

STOW reads no repository files and runs no helper merely because it is installed. A host or agent performs only the reference reads and explicit tool calls required by the selected route. Unknown project terms remain unresolved unless the caller supplies an approved terminology surface.

Validate a structured artefact

python ~/.claude/skills/stow/runtime/validate.py --format json some-file.json
python ~/.claude/skills/stow/runtime/validate.py --schema handoff my-handoff.md

A valid instance exits 0. A validator result applies only to the supplied candidate; STOW does not automatically intercept or approve a host's final response.

How STOW works

STOW uses progressive disclosure rather than one always-hot rule dump.

Layer	Purpose	Normal runtime role
Kernel (skills/stow/SKILL.md)	Request mode, precedence, protected regions, ordinary prose discipline, and bounded routing	Loaded when the skill is selected
References (skills/stow/references/)	Technical, public-documentation, controlled, safety, procedure, format, and other predicate-specific guidance	Read only when one named predicate applies; normal generation should not walk neighbouring files
Corpus (skills/stow/corpus/)	Full rule statements, qualifications, examples, and audit anchors	Deep application, review, and maintenance; not a normal-turn payload
Runtime (skills/stow/runtime/)	Parsers, schema checks, term-map checks, dictionary lookup, advisory lint, profile resolution, and rule queries	Invoked explicitly when its closed contract is useful; normal prose generation must not probe --help, create temporary candidates, or run advisory lint by default

Architecture at a glance

The v0.4.1 registry contains 65 active canonical rules. That number is an implementation shape, not a coverage target.

• Sixty-one G1 semantic owners provide model-mediated guidance.
• Ten G1 owners are compacted into the ordinary kernel.
• Fifty-one G1 owners are cold or predicate-loaded.
• Four G2 predicates decide closed properties at declared input boundaries.
• Ten advisory signals report surface patterns; they do not decide contextual compliance.
• 57 G1 owners have qualifying behavioural evidence.
• STOW-PRO-005 remains an open contextual limitation.
• STOW-WRD-001 terminates at an external project-authority boundary.
• STOW-WRD-003 and STOW-VRB-005 are explicit contextual deferrals.

The original audit population was 96 rule IDs. The governed reconciliation record preserves how those IDs became the current 65-rule architecture; retired IDs do not remain active merely for traceability.

Profiles

Profile	Use	Mechanical behaviour
stow-default	Ordinary user-facing prose, focus, evidence, and protected-content guidance	No controlled punctuation, contraction, vocabulary, or sentence-length rules
technical-clarity	Technical explanations, coordination prose, and public-facing technical documentation that need stable names, explicit conditions, bounded steps, and evidence-aware claims	Same mechanical prose checks as stow-default; adds contextual technical guidance
controlled-technical-guided	Procedures, safety instructions, and controlled technical writing	Activates the supported controlled rule families, sparse dictionary access, and the semicolon, contraction, Latin-abbreviation, and sentence-length checks
controlled-technical-strict	Strict conformance	Locked. STOW does not ship or claim this capability

Raw and protected artefacts are a mode, not another profile. A host must identify the task and retain custody of the actual candidate.

Public-documentation route

README files, landing documentation, install guides, release notes, and public product or architecture documentation can select one direct cold read of references/public-documentation.md. That reference composes existing canonical owners for reader-task order, claim strength, semantic repetition, structure, terminology, protected content, and limitations; it is not a second ruleset. The route adds no profile, validator, renderer, repository census, or automatic claim audit. Repository inspection remains conditional on the claims the task must ground.

The unchanged seven-family paired suite qualified the exact candidate after an earlier failed candidate was preserved as evidence. All seven pathology cases and all seven already-good controls passed fresh blinded Luna Max review, with the declared facts and protected literals preserved. This is bounded evidence for the shipped composition, not a guarantee of universal documentation quality or rendered-consumer equivalence.

Deep-absorption evidence

Two source-neutral clustered suites exercise the product's deep prose and action-shaping semantics. The prose suite covers six pathology families and six legitimate countercontexts; the action suite covers eight pathology families and eight controls. All declared literals passed mechanical custody checks, and two blinded Luna Max reviewers per suite accepted every case. The frozen case and owner map is tests/evals/deep-absorption-v1.yaml. This bounded same-model result does not establish universal behaviour, cross-model equivalence, or final-output custody.

Callable tools

The tools are optional accelerators with explicit evidence ceilings.

Tool	What it can establish	What it cannot establish
runtime/validate.py	Whether the supplied JSON, JSONL, YAML, or schema instance satisfies its closed parser/schema contract	That the model's final response was the validated candidate, or that a host will block delivery
runtime/lint_prose.py	Advisory surface findings under a caller-supplied profile	Contextual writing quality or compliance; findings do not make the command fail
runtime/validate_terms.py	Compliance with an explicit closed term map over caller-labelled editable/protected segments	That the segment labels are semantically correct or that an unknown term is authorised
runtime/dictionary_lookup.py	Fixed dictionary membership, listed alternatives, and listed forms from the bundled projection	Intended sense, part of speech in context, or project-specific terminology authority
runtime/query_rules.py	The public registry record, applicable profiles, conflicts, and corpus anchor for a rule ID	Automatic semantic routing or live host activation

validate.py requires ruamel.yaml and jsonschema; the other listed runtime helpers are standard-library only.

Evidence

Controlled-language benchmark

A frozen, crosswalk-derived comparison used the same Luna Max model across no conditioning, a name-only controlled-language instruction, raw Issue 9 source conditioning, and STOW. The benchmark accounted for sixty-one top-level requirements plus 25 explicit child requirements.

Arm	PASS	FAIL	NOT_SCORED
Name-only	78	6	2
STOW	80	4	2

STOW uniquely passed three rows involving paragraph grouping, an ambiguous with relation, and paragraph sentence count. Name-only uniquely passed one procedure row because STOW distorted advisory force. Later targeted regressions now cover all four former STOW failure distinctions: same-item noun-cluster repair, source modality, paragraph topic grouping, and safety operation/force preservation. The complete four-arm trial was not rerun, so its frozen score remains unchanged.

This supports a bounded claim: STOW operationalised more of the tested requirement surface on that model and corpus. It does not establish universal output superiority, strict conformance, cross-model proof, or long-thread durability. Complete accounting is not complete behavioural compliance.

The frozen Trial 2 result and current targeted follow-up are summarised here and in CHANGELOG.md. docs/FUNCTIONAL-EVIDENCE.md preserves the earlier enabled-versus-disabled programme, and docs/evaluation-results.md documents the fixture and detector baseline; neither is presented as the owner of Trial 2.

Normal runtime instrumentality

A separate one-host probe measured cumulative logical input for normal installed-skill use after removing an overengineered runtime path that had caused reference walks, helper discovery, --help probes, temporary candidates, advisory lint calls, and repeated model turns.

Task	Post-fix STOW / name-only logical input
Ordinary README edit	3.2044×
Technical documentation edit	3.1848×
Controlled procedure edit	5.7776×
Two-turn repository workflow	1.5651×
Second turn only	1.0651×

These are logical-input ratios, not billing multipliers. The STOW arms had higher cache-hit fractions, but the Pro-plan receipts did not expose an authoritative cached-input price, so monetary cost was not derived. Controlled use remains materially more expensive at cold start; repeated use in the measured established thread was close to name-only input.

The result is bounded to one Codex host. Cross-host economics, cross-model behaviour, and the hypothesis that externalised governance degrades less over long or compacted conversations remain unproved.

This README is the public owner of the normal-runtime summary carried into v0.4.1. docs/design.md supplies deeper architecture detail; the README keeps the measured ratios and their evidence limits together so a reader does not mistake logical input for billing cost.

Rule classes at a glance

The current registry indexes 65 primary rules under STOW's own functional taxonomy. The registry defines each active rule's operational metadata; the governed reconciliation record preserves the original population and retired IDs.

Rule class	What it governs
Action and task shaping	How a reply opens, closes, tracks progress, and stays actionable.
Prose integrity	No filler, no fabricated specificity, no synthetic voice.
Words and terminology	Word choice and consistent naming under the controlled profile.
Multi-word nouns	Length and clarity limits for noun clusters.
Verbs and voice	Verb forms, tense, and the active voice.
Sentences and paragraphs	Sentence completeness, length discipline, and paragraph focus.
Procedures	Instruction shape and sentence limits for executable steps.
Descriptive writing	Structure and length discipline for explanatory text.
Safety instructions	Complete, correctly formed warnings, cautions, and notes.
Punctuation and word counting	Punctuation limits and how words are counted against caps.
Writing style	Rewriting guidance when a word-for-word fix is not enough.
General writing practice	Cross-cutting recommendations for controlled technical text.

The table is the landing-page summary. Exact rule wording, applicability, exceptions, and status remain generated from the registry; the complete generated catalogue is retained as an appendix below and the rule index remains the stronger navigation owner.

Important boundaries

• Contextual guidance is not deterministic. G1 rules can fail even when read. STOW-PRO-005 remains open because a measured model twice preserved unsupported evaluative labels.
• Only closed properties get G2 claims. Four rules have callable compliance predicates. Advisory surface matches do not prove contextual harm.
• No final-output custody ships. A host must hold the actual final candidate, run any required checks, block invalid or unknown results, authorise repairs, and revalidate before delivery.
• Automatic activation is not guaranteed. Installed skill availability, host selection, and actual reads are separate facts.
• Strict controlled-language conformance is locked. Dictionary lookup, contextual guidance, project terminology, and bounded validators do not establish full conformance.
• Project terminology needs authority. Unknown words are not self-authorising technical terms. A caller or repository must supply an approved terminology surface.
• Dictionary facts have limits. The bundled 2,198-record projection can expose membership, forms, and alternatives. It cannot decide the intended sense or grammatical role by itself.
• Protected-region handling is bounded. Guidance and advisory masking cover declared or recognisable regions; STOW has no universal byte comparator for final output.
• Public-documentation evidence is bounded. The paired qualification supports the exact cold composition and cases; it does not establish universal documentation quality, automatic repository verification, or rendered-consumer equivalence.
• Runtime evidence is local. The measured token ratios come from one Codex/Luna Max environment. Billing cost, cross-host parity, cross-model parity, and long-run durability were not established.
• STOW is not an AI detector. It suppresses specified writing pathologies without claiming to identify the author of a text.

Documentation map

Need	Public owner
Full active rule catalogue	Generated appendix below; authoritative navigation in skills/stow/references/rule-index.md
Rule conflicts and precedence	docs/rule-conflicts.md
Current v0.4.1 product/evidence summary	This README and CHANGELOG.md
Architecture and profile model	docs/design.md
Earlier enabled-versus-disabled evidence	docs/FUNCTIONAL-EVIDENCE.md
Fixture and detector baseline	docs/evaluation-results.md
Rule usability and implementation status	docs/RULE-USABILITY.md
Package and install evidence	docs/INITIAL-PACKAGE-HEALTH.md
Self-dogfood and public-claim checks	docs/SELF-DOGFOOD.md
Historical rewrite-readiness checkpoint	docs/REWRITE-READINESS.md
Source skill contract	skills/stow/SKILL.md

The complete primary-rule catalog

Every primary rule, exactly once, grouped by rule class. The one-sentence summaries are navigational, not authoritative: the registry (skills/stow/rules/registry.yaml) defines operational metadata, and each rule's corpus module carries the full statement, qualifications, and examples. Expand a class to see its rules.

Status meanings: Callable means a shipped validator checks it mechanically. Planned means the mechanism is specified but not implemented. Review-fallback means a model applies it by reading it; no program checks it.

These statuses describe the mechanical implementation route, not whether the semantic guidance exists or has behavioural evidence. Planned does not mean that the rule is absent from the writing specification; it means that no callable checker implements the stated mechanism.

Action and task shaping (ACT-001 through ACT-011)

Rule	Summary	Applies when	Status
STOW-ACT-001	Action-first response opening	the request is an actionable task; exception: an informational request leads with the answer; safety or decision context can precede the action	Planned
STOW-ACT-002	Number ordered multi-step work as bounded, task-complete actions	the work contains more than one ordered action; exception: preserve exhaustive required material and use a nonprocedural structure when order is not part of the reader's task	Planned
STOW-ACT-004	Defer secondary issues without dropping them	a secondary issue surfaces during the main task; exception: answer a blocking question in place; otherwise preserve the deferred issue in one bounded later note	Review-fallback
STOW-ACT-005	Surface changed progress and remaining state	a multi-turn task has material changed state or the reader needs a resume boundary; exception: a single-turn answer or a turn with no material state change needs no repeated progress ledger	Review-fallback
STOW-ACT-006	Concrete effort estimates	a defensible range exists for the estimate; exception: with no defensible range, omit the figure; accuracy outranks the preference	Review-fallback
STOW-ACT-007	Surface completed outcomes	work ran and produced a result this turn	Planned
STOW-ACT-008	Neutral error reporting	the turn reports an error	Planned
STOW-ACT-011	Lists, not tables, for action sequences	action sequences, not comparison data	Planned

Prose integrity (PRO-001 through PRO-023)

Rule	Summary	Applies when	Status
STOW-PRO-001	Use em dashes only under an explicit style contract	an explicit style contract governs punctuation; exception: a deliberate house style that permits em dashes remains permitted	Review-fallback
STOW-PRO-002	Require attributable numbers	any numeric claim; exception: no attributable source: omit the number rather than invent one	Review-fallback
STOW-PRO-005	End claims on a concrete detail	factual claims in editable prose; exception: a conceptual definition satisfies this with a precise, checkable statement	Review-fallback
STOW-PRO-006	Functionless semantic repetition	a later statement repeats an earlier meaning without adding function; exception: functional repetition for correction, safety, navigation, emphasis, or stable terminology remains permitted	Review-fallback
STOW-PRO-007	Avoid mechanical repetition that obscures function	consecutive repeated structures obscure the function of the content; exception: preserve deliberate parallelism, recurring terminology, house style, and required layouts	Planned
STOW-PRO-009	Use urgency or intensified emphasis only when a decision-relevant reason is stated	urgency or intensified emphasis lacks a decision-relevant reason; exception: preserve a supported deadline-led command or requested functional emphasis	Review-fallback
STOW-PRO-011	Remove framing or process language only when it adds no information or decision value	framing or process language adds no information or decision value; exception: preserve a material limitation, method, audience, progress state, or requested voice	Review-fallback
STOW-PRO-013	Evidence-grounded requested voice	the response uses the default or explicitly requested voice; exception: the requested voice governs while factual claims remain evidence-grounded	Review-fallback
STOW-PRO-015	Grounded uncertainty	uncertainty lacks an evidence boundary or stacked qualifiers obscure the uncertain proposition or reason; exception: justified uncertainty and bounded capability statements remain permitted	Review-fallback
STOW-PRO-016	Concrete, descriptive headings	a heading teases, dramatizes, or abstracts instead of naming the section content; exception: preserve a functional expressive heading in a requested voice when it still identifies the subject	Review-fallback
STOW-PRO-017	No fabricated scenarios	Conditional	Review-fallback
STOW-PRO-018	No fabricated history	Conditional	Review-fallback
STOW-PRO-019	No fabricated attributions	Conditional	Review-fallback
STOW-PRO-020	Review formulaic lexical patterns	a listed pattern or corrective contrast may obscure its function or reject a characterization absent from the discourse; exception: ordinary connectors, technical uses, and a discourse-present correction naming the real differentiator remain permitted	Review-fallback
STOW-PRO-023	Quote sources accurately	Quoted sources	Review-fallback

Words and terminology (WRD-001 through WRD-014)

Rule	Summary	Applies when	Status
STOW-WRD-001	Use dictionary-approved words; admit technical nouns and technical verbs only under a defined category supplied by project authority, and prefer an approved dictionary verb when one exists	controlled vocabulary is requested and dictionary or project terminology authority is available; exception: unknown technical terms require external authority and are not rejected by lexical lookup alone	Review-fallback
STOW-WRD-002	Use each approved word only in its dictionary-specified part of speech and listed forms; a listed past participle can act as an adjective	Controlled profile	Planned
STOW-WRD-003	Use approved words only with their dictionary-approved, often restricted, meanings	an approved meaning is supplied for contextual review; exception: lexical membership and listed alternatives do not establish the intended sense or an equivalent action	Review-fallback
STOW-WRD-007	Do not use a technical noun as a verb; keep it a noun or adjectival modifier	Controlled profile	Planned
STOW-WRD-008	Prefer the technical noun already approved by your company, industry, or subject field	Controlled profile	Planned
STOW-WRD-010	Do not use regional, slang, or jargon words as technical nouns	Controlled profile	Review-fallback
STOW-WRD-011	Use one technical noun consistently for one item, preserve key words and key phrases that organize the logic, and reuse recurring wording for the same context	guidance-level under the technical-clarity profile; binding under the controlled profile	Planned
STOW-WRD-014	Use American English spelling unless another official directive overrides; do not change quoted-text spelling	Controlled profile	Planned

Multi-word nouns (MWN-001 through MWN-001)

Rule	Summary	Applies when	Status
STOW-MWN-001	Keep multi-word nouns to a maximum of three words and keep coined terms short and easy; for a longer approved noun, write it in full first, then restructure it or use a declared short form, approved abbreviation, or clear hyphenation while you preserve enough identity-bearing words to refer to the same item	a coined or approved noun phrase contains stacked nouns or modifiers in controlled prose; exception: project-authorized terms, protected identifiers, item identity, and clear natural pronoun coordination take priority	Review-fallback

Verbs and voice (VRB-002 through VRB-007)

Rule	Summary	Applies when	Status
STOW-VRB-002	Use only the infinitive, imperative, simple present, simple past, simple future, and listed past participle; do not use perfect, progressive, or other complex constructions	controlled prose requires a bounded tense or aspect choice; exception: preserve a time relation when the source or procedure requires it	Planned
STOW-VRB-005	Use an -ing word only as a technical noun or as a modifier inside a technical noun	a word ending in ing appears outside a declared technical noun; exception: project authority can classify the form as a noun term or noun modifier	Review-fallback
STOW-VRB-006	Use active voice; passive is allowed only in descriptive writing when the agent is unknown	Controlled profile	Planned
STOW-VRB-007	Describe an action with an approved verb, not a nominalization; technical verbs stay verbs, except that a listed past participle can act as an adjective	Controlled profile	Planned

Sentences and paragraphs (SEN-002 through SEN-005)

Rule	Summary	Applies when	Status
STOW-SEN-002	Do not omit words or use contractions; write every word in full	Controlled profile	Review-fallback
STOW-SEN-003	Break complex text into a vertical list with the prescribed layout	Controlled profile	Planned
STOW-SEN-004	Use approved connecting words and phrases to link related sentences	Controlled profile	Planned
STOW-SEN-005	Use articles and demonstratives before nouns where grammatically correct	Controlled profile	Planned

Procedures (PRC-001 through PRC-005)

Rule	Summary	Applies when	Status
STOW-PRC-001	Limit each procedural sentence to a maximum of twenty words	Controlled profile	Callable
STOW-PRC-002	Write only one instruction per sentence unless actions occur at the same time	Controlled profile	Planned
STOW-PRC-003	Write instructions in the imperative command form	the source already authorizes a command rather than advice or a permission statement; exception: do not silently strengthen advice, permission, or uncertainty; preserve source force or request an authority decision	Planned
STOW-PRC-004	State a required condition first and separate it from the command with a comma	Controlled profile	Planned
STOW-PRC-005	A note in a controlled procedure gives information and does not introduce an action	a note is attached to a controlled procedure; exception: a higher-precedence literal or output contract takes priority	Planned

Descriptive writing (DSC-001 through DSC-006)

Rule	Summary	Applies when	Status
STOW-DSC-001	Introduce information gradually, one subject per sentence	Controlled profile	Review-fallback
STOW-DSC-003	Limit each descriptive sentence to a maximum of twenty-five words	Controlled profile	Callable
STOW-DSC-004	Group related information into paragraphs, each led by a topic sentence	Controlled profile	Review-fallback
STOW-DSC-006	Keep every paragraph to a maximum of six sentences	Controlled profile	Planned

Safety instructions (SAF-001 through SAF-003)

Rule	Summary	Applies when	Status
STOW-SAF-001	Label each safety instruction with a word that identifies the level of risk	Safety notices	Review-fallback
STOW-SAF-002	Begin a safety instruction with a clear, accurate command or condition	Safety notices	Planned
STOW-SAF-003	State the risk or the possible result of not obeying the safety instruction	Safety notices	Review-fallback

Punctuation and word counting (PCT-001 through PCT-007)

Rule	Summary	Applies when	Status
STOW-PCT-001	Do not use the semicolon; write two separate sentences instead	Controlled profile	Callable
STOW-PCT-003	Use parentheses only for references, item identifiers, step identifiers, abbreviations, singular/plural forms, explanations, or alternatives	parentheses appear in controlled prose; exception: protected text and the listed parenthetical purposes remain unchanged	Planned
STOW-PCT-004	In a vertical list, a colon counts as a period for word count and ends a sentence	Controlled profile	Planned
STOW-PCT-005	Parenthetical text counts as one word in the host sentence	Controlled profile	Planned
STOW-PCT-006	Count a number, number with unit, abbreviation, identifier, quoted text, title or label, or proper name as one word	Controlled profile	Planned
STOW-PCT-007	Use hyphens only between directly related words that operate as one unit; a hyphenated group counts as one word	Controlled profile	Planned

Writing style (STY-001 through STY-003)

Rule	Summary	Applies when	Status
STOW-STY-001	When a word-for-word replacement is insufficient, rewrite the sentence while preserving the meaning	Controlled profile	Review-fallback
STOW-STY-003	Do not combine approved words into unlisted phrasal verbs	Controlled profile	Planned

General writing practice (GEN-002 through GEN-007)

Rule	Summary	Applies when	Status
STOW-GEN-002	Rewrite a with phrase only when it has two plausible attachments	a with phrase has two plausible attachments; exception: leave a clear with phrase unchanged	Planned
STOW-GEN-003	Use only approved pronouns; replace an ambiguous pronoun with its noun	Controlled profile	Planned
STOW-GEN-005	Confirm a possible false friend against a supplied source-language meaning	a source-language form or intended English meaning is supplied for controlled review; exception: do not infer a false friend from spelling resemblance alone	Review-fallback
STOW-GEN-006	Avoid Latin abbreviations; use English words instead	Controlled profile	Callable
STOW-GEN-007	When gender is unknown or irrelevant, name the role or use an inclusive reference	a human role is named and gender is unknown or irrelevant; exception: preserve gender when it is a supplied fact or materially relevant	Planned

Contributing and regeneration

The registry is canonical. Change the owned source, regenerate derived public surfaces, and run the checks:

python tools/gen_rule_index.py
python tools/gen_always_on.py
python tools/gen_rule_conflicts.py
python tools/gen_readme_catalog.py
python -m pytest tests/ -q
python tools/check_provenance_leak.py --local

Do not hand-edit generated regions such as references/rule-index.md, references/always-on.md, or docs/rule-conflicts.md.

Troubleshooting

The skill does not activate. Confirm that the install resolves to <skills-dir>/stow/SKILL.md. A nested <skills-dir>/STOW/stow/SKILL.md layout will not resolve correctly. Then invoke STOW explicitly; automatic selection is host-dependent.

The strict profile fails. That is intentional. controlled-technical-strict is locked because STOW does not have the authority, contextual decisions, final-output validation, and delivery custody required for a strict conformance claim.

A validator rejects JSON that looks valid. Check for a byte-order mark, duplicate key, trailing comma, NaN, Infinity, or a pasted code fence.

YAML values changed type. Quote scalars whose string form matters. Keys that coerce to the same scalar are duplicate keys even when their spelling differs.

STOW edited code or a literal. That is a precedence violation unless editing the literal was the task. G1 guidance and the advisory linter cannot prove final byte identity; use an independent comparison when exact preservation is required.

A generated documentation file returns after editing. Change the registry or conflict owner and regenerate. Do not hand-edit generated output.

Upgrading. Remove the previous <skills-dir>/stow/ directory and extract the new release artefact in its place. Verify the installed payload against the release you intended to install.

Repository layout

skills/stow/
  SKILL.md              compact kernel and source skill contract
  references/           predicate-loaded public guidance and generated indexes
  corpus/               full rule modules and stable audit anchors
  rules/                registry, profiles, conflicts, dictionary data, and schemas
  runtime/              bounded validators, lookup helpers, lint, and profile resolution
  schemas/              structured artefact schemas
  templates/            worked structured artefact templates
docs/                   design, evidence, package health, readiness, self-dogfood, and conflicts
tests/                  test suite, evaluations, and fixtures
tools/                  builders, generators, measurements, and checks
dist/                   built skill archive, checksum, and manifest