Audit anthropomorphic cues

Source and version

Repository path

.agents/skills/audit-anthropomorphic-cues/SKILL.md

Source revision

cfaa53f5b7500e3901d8470e1e471cd0230ff2a5

Source SHA-256

26bc66d0cc0ccee013e34f9dfcbcc1883a595c96523cd29c99d7c0fd71d6f0f0

Publication standing

Public ready

Included source files

The linked SKILL.md is the complete public source package for this skill.

Complete skill source

The Markdown below matches the exact source text. The format link above opens the file on its own.

---
name: audit-anthropomorphic-cues
description: Audits bundled person-like cues in generative-system interfaces and work language, separating human-interpretation evidence, model-output hypotheses, local system functions, and accountable authority. Applies when names, pronouns, avatars, voice, emotional language, agency claims, role metaphors, or tool surfaces may change reliance, delegation, capability expectations, responsibility, or recovery.
---

# Audit anthropomorphic cues

## Contract

- **What:** A cue-bundle audit that makes relational value, technical referents, evidence standing, authority implications, and the smallest useful change visible together.
- **When:** Person-like or social cues may shape how people interpret a generative system, or when a terminology pass risks treating one phrase as the whole intervention.
- **Boundaries:** The audit diagnoses and recommends within the current task. Source owners retain quoted, cited, participant-supplied, historical, legal, public-name, and compatibility-contract language. Named people and institutions retain interface, publication, access, consequential interpretation, and accountability decisions.
- **Result:** A compact cue ledger plus retain, qualify, separate, disclose, test, or revise decisions that preserve useful warmth while making the acting system and accountable human recoverable.

## Fit and attention

- **Useful when:** A name, persona, first-person voice, pronoun, avatar, emotional phrase, role metaphor, turn-taking pattern, customization feature, agency claim, or tool surface may affect reliance or responsibility.
- **Poor fit when:** Ordinary factual editing resolves the wording, exact source language must remain verbatim, or the requested result is a causal comparison rather than an interface and language audit.
- **Smallest useful dose:** Inspect the consequential cue bundle and its nearby referents; revise only the ambiguity that changes a reader's interpretation or action.
- **Interactions:** `context-canvas-surface-reading` owns the visitor and runtime map; `source-authority-mapping` owns source and projection lineage; the receiving writing practice owns final prose; a separately authorized evaluation owns causal comparison.
- **Rest or release when:** The cue bundle, intended value, evidence standing, acting system, accountable owner, and smallest useful move are recoverable. A new cue, population, deployment context, or measured effect can reopen the audit.

## Rules

### Rule: Inspect the bundle before changing a word

Read the phrase with its name, imagery, voice, role label, interaction pattern, tool access, and surrounding system-state signals. A personal noun may be harmless alone and consequential in a stronger bundle. Contextual reading preserves that distinction; a global replacement erases it.

### Rule: Map the actor and the arrangement separately

Use the narrowest accurate referent for the acting component: provider, product, model, configured instance, generative AI system, retrieval aid, drafting system, proposal channel, validator, state projection, or another evidenced local function. `Collaboration` can continue to name an arrangement of work among people, systems, and artifacts. A nearby actor noun identifies the person, authority holder, or technical component responsible for each move.

### Rule: Keep identity, capability, and authority distinct

An interaction alias or role metaphor is an interface cue. State product or model identity, local work role, permissions, evidence horizon, review requirement, and accountable person or institution when those distinctions change reliance or consequence. Actual system behavior and permissions establish capability; named people and institutions retain authority.

### Rule: Separate effect classes

Distinguish observed human interpretation, measured human behavior, interface affordance, observed system output, plausible model-conditioning hypothesis, and unsupported inference. Keep name-only observations separate from studies or interfaces that bundle several social cues.

### Rule: Preserve source-bearing and compatibility language

Retain proper names, quoted or cited language, participant wording, stable titles, filenames, URLs, route IDs, schema terms, serialized keys, and historical versions unless their owner explicitly changes them. Local explanation can clarify their role without rewriting their identity.

### Rule: Choose the smallest useful move

- **Retain** when the cue carries clear relational or technical value and the surrounding contract is legible.
- **Qualify** when one nearby noun, role clause, provenance field, or evidence statement resolves ambiguity.
- **Separate** when product identity, interaction alias, local role, and authority have collapsed into one label.
- **Disclose** when permissions, evidence horizon, review, or accountable ownership materially affects use.
- **Test** when a consequential choice depends on a causal effect the available evidence cannot establish.
- **Revise** when the wording assigns human-like agency, judgment, memory, emotion, or authority to a technical component without a source-owned reason.

### Rule: Generated carriers follow their source owner

Trace visitor copy through templates, generators, manifests, mirrors, and delivery artifacts. Edit the semantic source, update governed versions and manifests when their contract requires it, and rebuild generated carriers through their existing command. Generated output is evidence of the source route, not an alternate editing surface.

## Operating sequence

1. **Bound the surface:** Name the visitor, workflow, source family, consequence, and current authority.
2. **Inventory cues:** Record the social and person-like cues that arrive together.
3. **Map referents:** Identify the product or system, local technical function, interaction alias, permission boundary, evidence horizon, and accountable human or institution.
4. **Classify evidence:** Separate human-facing observations, measured behavior, system-output observations, hypotheses, and open questions.
5. **Locate consequence:** Note effects on reliance, delegation, capability expectations, responsibility, authority, and recovery.
6. **Select the move:** Retain, qualify, separate, disclose, test, or revise each decision-changing cue.
7. **Apply at authority:** Update the authored source and matching descriptive or generated projections through their owners.
8. **Verify semantically:** Confirm the desired role distinction across visible prose, metadata, accessibility text, serialized presentation strings, and current generated carriers while preserving allowlisted source-bearing language.

## Working ledger

| Field | Record |
| --- | --- |
| Cue and bundle | The phrase or interface cue and the signals that travel with it |
| Intended relational value | The welcome, continuity, coordination, or other useful job to preserve |
| Displayed claim and verified boundary | What the interface says the system can do; the observed behavior, permission, and enforcement evidence that supports or limits that claim |
| Acting system and local function | The narrow technical referent supported by the source |
| Human effect | Observed, measured, supported, hypothesized, or open interpretation effect |
| System-output effect | Observed output or bounded hypothesis, kept distinct from human interpretation |
| Consequence and affected population | Reliance, delegation, responsibility, authority, or recovery implication |
| Accountable owner | The person or institution holding the consequential decision |
| Move and recovery | Retain, qualify, separate, disclose, test, or revise; plus the return route |

## Representative scenarios

### Scenario: A generic role label hides the acting component

- Given visitor copy says an “AI collaborator” maintains orientation and decides what to surface
- And the implementation is a generative system that can propose from a bounded source packet
- When the cue bundle is audited
- Then the copy names the proposal or retrieval function
- And the named human owner retains selection and consequential judgment
- And `collaboration` may continue to describe the larger work arrangement

### Scenario: Exact language is part of the source contract

- Given a cited title, quotation, schema enum, serialized key, route identifier, or historical artifact contains person-like language
- When changing that value would alter identity, compatibility, or source fidelity
- Then the value remains intact
- And nearby explanation clarifies the current technical or authority boundary when readers need it

### Scenario: Recommendation, selection, approval, and execution are different moves

- Given an interface says the system “decides” which consequential application to submit
- And the observed flow only establishes that it compares options, drafts a packet, and can execute submission after a named person's approval
- When capability and authority are mapped
- Then the system's copy uses the evidenced proposal, comparison, drafting, or execution verb
- And the named person or institution is shown as the selection, approval, and accountability holder where the source establishes those roles
- And any unresolved selection or permission boundary remains an explicit open question rather than an invented assignment

## Completion

The audit is complete when a reader can distinguish the cue bundle, intended relational value, acting technical component, evidence standing, affected people, and accountable authority; each change is source-owned and proportionate; and unresolved causal effects have a bounded evaluation route rather than an implied conclusion.