Ubiquitous Language

Create or update a project vocabulary so users, domain experts, and agents mean the same thing without forcing the user to know specialist terms.

Outputs

Produce canonical terms, aliases, relationships, ambiguities, sources, and prompt translations. When corrective feedback is the source, also produce its intent radius, generalized rule, pattern sweep, dispositions, and enforcement handoff. Use schema_version: 1 when automation consumes the output.

Workflow

1. Determine scope and output path.
2. Resolve the active glossary with the output-format routing; ask only when multiple contexts remain ambiguous.
3. Read any existing glossary first and preserve intentional choices.
4. Extract domain nouns, workflow verbs, actor names, lifecycle states, aliases, and overloaded phrases.
5. When the input is corrective feedback, run Corrective Feedback Mode before choosing scope.
6. Choose canonical terms that improve execution; keep natural-language aliases when useful.
7. Write or update the active ubiquitous-language file.
8. Add a concise pointer in the nearest active agent instruction surface.
9. Report the highest-value terms, prompt translations, sources, pattern-sweep dispositions, enforcement handoffs, and skipped evidence.

Corrective Feedback Mode

Use this mode when the user corrects a specific line, function, file, command, API, workflow, or implementation detail and the correction may express wider engineering intent.

1. Treat the visible example as evidence, not as the presumed scope boundary.
2. Classify the Feedback Intent Radius as line, function, file, package, repository, architecture_rule, or durable_memory.
3. Run a bounded Pattern Sweep across structurally similar implementations, glossary entries, prompt translations, validators, schemas, tests, and policy surfaces. Do not equate textual similarity with equivalent semantics.
4. State the Generalized Feedback Rule without the incidental identifier, path, function, error, or example that exposed it.
5. Give every relevant sibling a Similar-Case Disposition: align now, different semantics, defer with reason, or not applicable.
6. Decide whether language alone is sufficient. When recurrence needs mechanical prevention, route the rule to the owning validator, lint rule, schema, reusable abstraction, shared utility, repository convention, style rule, CI check, or architecture policy.
7. Preserve authority boundaries. Systemic intent justifies broader inspection, not unrelated or cross-repository mutation; record or hand off out-of-scope enforcement work explicitly.

Worked Corrective-Feedback Output

Input: "This parser fix is not doing what I want; do not patch only this function."

Feedback Intent Radius: repository
Generalized Feedback Rule: Parse failures use typed results; callers never infer failure from missing values.
Pattern Sweep: parsers/** and their direct callers
Similar-Case Dispositions: 3 aligned now; 1 different semantics; 1 deferred to another owner
Enforcement: schema constraint plus parser-contract regression test

Worked Glossary Output

Input: "What does 'make it available' mean in this repository?"

**Installed Runtime Copy**:
The selected checked package version installed through the authorised runtime route.
_Avoid_: copied source, approval inferred from a holding directory

## Prompt Translations

| User phrase | Canonical action |
| --- | --- |
| "make it available" | Resolve the active repository's installation policy and authority. For this lifecycle, select the SDK-checked private Tessl version or verified OpenAI-provider exemption, then prove installed identity and discovery. Report a missing route rather than defaulting to workspace or user sync. |

This is a translation, not authority to install, relink or change runtime files. Recheck the active repository's glossary before reusing an example elsewhere.

Domain Grill Mode

Use this mode when the user asks to be grilled, challenged, interviewed deeply, or stress-tested against the project domain model. Follow the focused domain-grill procedure.

Output Format

Use the output template for new single-context or multi-context glossaries. Preserve UBIQUITOUS_LANGUAGE.md when an existing repository contract consumes that filename. Translate external CONTEXT.md wording to the active local glossary surface before writing.

Term Rules

• Be opinionated. When multiple words exist for the same concept, pick one canonical term and list the others as aliases to avoid.
• Flag conflicts explicitly. If a term is ambiguous, add it to Flagged Ambiguities with a clear resolution.
• Keep definitions tight: one or two sentences, defining what the term is rather than what it does.
• Show relationships. Use bold term names and express cardinality where obvious.
• Include only concepts specific to this project's context. General programming concepts do not belong just because the project uses them.
• Group terms under subheadings when natural clusters emerge. Use a flat list when all terms belong to one cohesive area.
• Write an example dialogue between a developer and a domain expert that demonstrates how terms interact and clarifies boundaries between related concepts.
• Preserve the principle behind corrective feedback independently of the local example that revealed it.

Safety

• Do not copy raw private transcripts, secrets, tokens, or unnecessary personal data.
• Do not overwrite an existing glossary wholesale.
• Ask before finalizing a material policy or scope decision that cannot be inferred.
• Mark low-confidence terminology choices.

Execution Boundaries

• Read only the current conversation, active instruction files, existing ubiquitous-language files, and directly relevant project docs unless the user asks for history-backed vocabulary.
• Write only the active ubiquitous-language file, context map, and the smallest necessary agent-instruction pointer.
• Do not run network commands, publish artifacts, change runtime projections, or edit unrelated docs for a terminology-only update.
• Do not infer domain authority from generic code symbols; require project-specific source evidence or user wording before adding a term.
• Do not use this skill for ordinary symbol renaming, generic copyediting, or broad documentation rewrites without reusable terminology.
• Do not create ADRs, Linear notes, .harness/decisions/**, or architecture artifacts from this skill. Hand off to the repo's decision-record or architecture workflow after the terminology question is resolved.

Failure Mode

If the scope, source glossary, or authority for a terminology change is unclear, stop with one missing input rather than rewriting vocabulary by guesswork.

Validation

Confirm the active glossary exists, every canonical term has a one- or two-sentence definition, aliases to avoid are listed for terms with competing names, flagged ambiguities include a resolution, prompt translations include at least one user phrase when informal wording exists, and the agent instruction pointer references the active ubiquitous-language file. For corrective feedback, also confirm the intent radius, generalized rule, searched scope, similar-case dispositions, and enforcement decision are explicit; a local-only result must be supported by the pattern sweep rather than assumed from the named example. Fail fast: stop at the first failed gate and do not proceed until the blocker is fixed.

References

Never drop required context for brevity; move it into references or deferred context and link it here.

• Local contract, evals, and task profile: references/
• Output template: references/output-format.md
• Domain-grill procedure: references/domain-grill.md
• Archived long-form workflow: Infrastructure/references/deferred-skill-context/agent-ops-ubiquitous-language/

When to use

Use this skill when a named concept, boundary, or domain term is ambiguous across source, tests, interfaces, or operator language and a governed vocabulary decision is required.

Required Inputs

Provide the target repository or product surface, the terminology under review, authoritative sources where available, and the intended consumer of the resulting vocabulary decision.

Gotchas

Do not turn one naming preference into a global rule without an independent pattern sweep. Keep generated text, retrospective evidence, and user terminology distinct from canonical domain authority, and stop when ownership or approval is missing.