Expand the reasoning system, not the default context
SKILL.md compact and add seven specialist modules plus targeted expansions to ten existing files. Make the router reason about communication process, common-ground risk, repair, emotion, power, record needs, and evidence collection. Load the new material only when those conditions apply.
This preserves the skill’s strongest architectural choice: a small execution spine with progressive disclosure. The implementation should move from 21 to 28 reference modules and grow the source registry toward roughly 120–150 curated sources. That range is a capacity estimate, not a quota. A source belongs only when it changes a decision, method, validation, or boundary.
Complete the foundation
Add channel/grounding reasoning, audience research, writing mechanics, and risk/correction. These affect almost every consequential artifact.
Add human systems
Cover feedback and conflict, organizational alignment, product/service communication, and AI-mediated trust.
Deepen specialist evidence
Strengthen graphical perception, alternate formats, spoken delivery, and research-method detail without expanding the default load.
Scope: this is an audit and implementation plan. No installed skill files were changed.
From inventory to adversarial review
Map modules, routing logic, source coverage, and validation.
Separate concepts, mechanisms, genres, methods, and safeguards.
Compare against primary scholarship and authoritative practice.
Score consequence, recurrence, evidence, overlap, and load cost.
Try to remove, merge, or constrain every proposed addition.
A candidate counted as a material gap only if it met four tests: it recurs across communication tasks; its absence can change an outcome; it has defensible evidence or authoritative operational guidance; and it can be routed without burdening unrelated tasks. Keyword absence alone did not qualify. Topics already represented under another name were marked partial rather than missing.
The existing core is sound
Mature
ContractIntent routingCognitive loadProgressive disclosureEvidence labelsProvenanceArtifact validation
The communication brief, canonical knowledge objects, layered artifacts, genre playbooks, evidence discipline, and two-pass red-team process are appropriate foundations.
Strong format coverage
PapersReportsRFCsADRsTechnical docsChartsDiagramsSlidesWebWorkshopsShort-form
The skill knows how to shape many artifacts. The missing layer is often the process around them: how understanding is established, tested, repaired, negotiated, and maintained.
| Capability family | Current status | Audit judgment |
|---|---|---|
| Artifact and information architecture | Strong | Retain; use as the stable spine. |
| Cognition and attention | Strong | Retain; connect more explicitly to grounding, correction, emotion, and research methods. |
| Medium selection | Partial | Add communication-process and feedback-loop reasoning; avoid fixed rich/lean rankings. |
| Audience understanding | Partial | Convert “requires user evidence” into a concrete research-method selector and result record. |
| Writing mechanics | Partial | Add sentence-to-section information flow, cohesion, tone, and editing passes. |
| Risk, crisis, and correction | Thin | Add a high-stakes overlay for uncertainty, numeracy, trust, rumors, and updates. |
| Interpersonal and organizational dynamics | Thin | Add listening, feedback, conflict, power, psychological safety, change, and alignment. |
| Product/service and AI-mediated communication | Thin | Add lifecycle microcopy and human–AI transparency, contestability, and handoff. |
What is genuinely missing
| Priority | Gap | Why it changes outcomes | Canonical home |
|---|---|---|---|
| P0 | Channel, process, and grounding | Medium choice depends on whether people must absorb information, converge on meaning, repair misunderstanding, or create a durable record. | Expand communication-router.md |
| P0 | Audience research and evaluation | Without method selection, audience assumptions and “user evidence” remain slogans rather than testable inputs. | New audience-research-and-evaluation.md |
| P0 | Writing and editing mechanics | Genre templates do not solve sentence order, paragraph coherence, topic continuity, tone, or revision strategy. | New writing-and-editing.md |
| P0 | Risk, crisis, uncertainty, and correction | High-stakes communication needs phase-aware messaging, calibrated uncertainty, numeracy, rumor handling, and correction mechanics. | New risk-crisis-and-correction.md |
| P1 | Interpersonal feedback and conflict | Listening, question design, face threat, power, repair, and negotiation determine whether a correct message is usable. | New interpersonal-feedback-and-conflict.md |
| P1 | Organizational change and alignment | Change communication is an evolving sensemaking and coordination system, not a launch announcement. | New organizational-change-and-alignment.md |
| P1 | Product and service lifecycle | Labels, permissions, errors, confirmations, notifications, onboarding, support, and records form one communication journey. | New product-and-service-communication.md |
| P1 | AI-mediated communication | AI involvement changes disclosure, trust, personalization, explanation, correction, consent, provenance, and escalation needs. | New ai-mediated-communication.md |
| P2 | Specialist evidence depth | Graphical perception, alternate formats, spoken/nonverbal delivery, and empirical publication methods need better evidence and narrower playbooks. | Expand existing specialist modules |
1 · Channel, process, common ground, and repair
The router presently asks sensible questions, but it lacks a theory of what a channel must enable. Add the distinction between information conveyance and meaning convergence; common-ground gaps; feedback latency; rehearsability; reprocessability; parallelism; transmission speed; privacy; emotional sensitivity; and record requirements. Clark and Brennan’s grounding model makes understanding a collaborative achievement with evidence and repair, while Media Synchronicity Theory explains why the same medium may help one communication process and hinder another.
Design implication: choose a primary channel plus companion channels. A pre-read may support conveyance, a meeting convergence, and a decision log the durable record. Never encode a universal “richer is better” hierarchy.
2 · Audience research and empirical evaluation
Add a method selector covering interviews, contextual observation, moderated and unmoderated usability testing, cognitive interviewing, surveys, diary studies, experiments, analytics, and mixed methods. Select by question, consequence, sample access, ecological validity, and needed confidence. Record participant boundaries, protocol, findings, disconfirming evidence, and what changed.
Design implication: every “requires user evidence” status must name the next method and its minimum evidence record. The skill should distinguish formative research, comprehension testing, behavioral validation, and outcome evaluation.
3 · Writing, discourse, and editing
Add information flow from sentence to section: given-before-new ordering, stable topics, explicit referents, coherent paragraph roles, rhetorical relations, transitions, controlled terminology, tone, and voice. Provide separate editing passes for argument, structure, evidence, clarity, compression, accessibility, and surface correctness.
Design implication: this module improves papers, reports, briefs, emails, documentation, scripts, and presentations without duplicating their genre structures.
4 · Risk, crisis, uncertainty, numeracy, and correction
Add a high-stakes overlay for preparedness, initial event, maintenance, resolution, and evaluation; audience emotion and trust; what is known, unknown, changing, and next; absolute versus relative risk; natural frequencies; calibrated visual uncertainty; rumor monitoring; prebunking; and correction. Corrections should provide a coherent alternative explanation, avoid unnecessary repetition of false claims, and account for continued influence rather than assuming a retraction erases belief.
Design implication: gate this module on safety, health, security, incident, material uncertainty, or active misinformation. Require named update ownership and timestamps.
5 · Interpersonal, conflict, power, and psychological safety
Add intent clarification, active listening, open and diagnostic questions, observable behavior versus inference, impact, needs and boundaries, feedback consent, repair, de-escalation, negotiation, facilitation, power asymmetry, dissent channels, and escalation. Psychological safety should be treated as a team condition that permits interpersonal risk-taking—not a promise that conversations will feel comfortable or have no accountability.
Design implication: distinguish a document problem from a relationship or coordination problem. Sometimes the correct artifact is a structured conversation plus a written record.
6 · Organizational change, alignment, and shared mental models
Add stakeholder and impact mapping, local meaning-making, leader and manager cascades, feedback loops, rumor sensing, decision rights, transition states, readiness evidence, shared vocabulary, and alignment checks. Shared mental models matter most where people must coordinate interdependent work; they should be tested through predictions and scenario responses, not assumed from message exposure.
Design implication: model change communication as a lifecycle with listening and adaptation, not a one-time broadcast campaign.
7 · Product/service microcopy and lifecycle communication
Unify labels, instructions, forms, consent, permissions, validation, error recovery, empty states, confirmations, transactional records, notifications, onboarding, help, support, and handoff. Each message should say what happened, why it matters when appropriate, what the person can do, and how to recover. Interruptive channels should be reserved for timely value, not engagement inflation.
Design implication: route by lifecycle state and user task, and test the complete journey rather than isolated strings.
8 · AI-mediated and adaptive communication
Add disclosure of material AI involvement; source and uncertainty boundaries; personalization consent; explanations appropriate to impact; contestability; human review and handoff; correction of AI errors; and synthetic-media considerations. Avoid generic “AI-generated” badges that do not help the recipient assess trust, action, or recourse.
Design implication: keep provenance as the evidence record, while this module owns the human communication consequences of mediation and automation.
9 · Specialist expansions
Strengthen visual-data-communication.md with graphical-perception evidence and uncertainty tasks; audience-inclusion-and-localization.md with alternate formats and an ACR/VPAT/EN 301 549 boundary; speaking/media modules with prosody, pacing, gesture, turn-taking, interpreters, captions, and audio description; and research-and-publications.md with qualitative, quantitative, mixed-method, sampling, instrument, causal, synthesis, and academic-genre decisions.
Progressive disclosure by decision, not topic popularity
The core should remain a routing and validation system. Specialist knowledge loads only when a brief activates it. Consequential constraints—safety, action, uncertainty, deadline, owner, accessibility, and source status—must never be hidden inside collapsed detail.
Always load
Communication contract, brief, router, core cognitive safeguards, canonical knowledge objects, final validation.
Conditionally load
Intent/genre module, medium module, and overlays for risk, research, interpersonal dynamics, product lifecycle, AI mediation, inclusion, or governance.
Load for proof
Source registry, method details, reporting standards, compliance references, and deep mechanism notes only when claims or validation require them.
Router additions
| New input | Question it answers | Routing effect |
|---|---|---|
| Process stage | Convey information, converge on meaning, decide, coordinate, repair, or record? | Selects channel sequence and feedback pattern. |
| Common-ground gap | How much context, vocabulary, history, and assumption is shared? | Controls orientation, examples, checks, and interaction. |
| Equivocality | Are facts missing, or can the same facts support conflicting interpretations? | Favors research for uncertainty and dialogue/facilitation for equivocality. |
| Grounding need | What evidence of understanding is required? | Adds read-backs, teach-back, confirmation, scenario, or next-action evidence. |
| Emotion, trust, and power | Is the recipient threatened, skeptical, dependent, or unable to dissent safely? | Changes messenger, pacing, privacy, participation, and escalation. |
| Record requirement | Must the communication be reviewable, auditable, signed, or recoverable? | Adds durable companion artifact and version ownership. |
| AI involvement | Did AI materially generate, personalize, recommend, decide, or mediate? | Adds disclosure, evidence, recourse, and human-handoff requirements. |
Standard shape for each new module
Use when → decision inputs → mechanisms → design patterns → failure modes → validation methods → source boundaries. This keeps the references executable rather than encyclopedic.
Six gated phases
Contract
Capability matrix and acceptance tests
Create a coverage matrix before writing new prose. Define one canonical home for every concept, activation conditions, expected outputs, validation methods, and exclusions. Add scenario fixtures for an ambiguous paper, high-stakes presentation, incident update, risk notice, feedback conversation, organizational change, product error, AI disclosure, and mixed-method study.
- No content is added without a recurring decision or validation it changes.
- Every gap has a route and a test before it has a module.
Foundation
Grounding, research, writing, and risk
Expand communication-router.md; add audience-research-and-evaluation.md, writing-and-editing.md, and risk-crisis-and-correction.md. Update the brief with common-ground gap, emotional state, trust, power, correction state, AI involvement, and record needs.
- Router returns a primary medium, companion medium, grounding evidence, and unresolved assumptions.
- Audience-evidence requirements map to a concrete method and result record.
- Risk outputs distinguish known, unknown, changing, action, owner, and next update.
Human systems
Interpersonal, organizational, product, and AI
Add the remaining four new modules. Keep boundaries explicit: interpersonal covers person-to-person exchange and repair; organizational covers collective sensemaking and change; product covers service lifecycle messages; AI-mediated covers trust, disclosure, contestability, and handoff.
- Each module names when a document is insufficient and a conversation or feedback loop is required.
- Power, consent, privacy, escalation, and recovery are first-class inputs.
Specialists
Evidence-depth expansions
Update persuasion/public, visual data, inclusion/localization, meetings/media, presentations/speaking, research/publications, quality measurement, and provenance modules. Add primary research methods and reporting standards without turning the skill into a statistics textbook.
- Each prescription states evidence class, scope, and transfer limits.
- Current standards are dated and linked rather than paraphrased as timeless rules.
Integration
Skill spine, router code, and source registry
Update SKILL.md load rules, mandatory checks, companions, and direct links; extend route_communication.py with the new inputs and forms; add source metadata for source class, scope, transfer limit, freshness, and last review. Extend validation for tables of contents, direct-link reachability, duplicate URLs, and required source metadata.
SKILL.mdremains comfortably below 500 lines.- All references are one link away from
SKILL.md; files over 100 lines have a table of contents. - The router explains uncertainty; it never presents a score as objective truth.
Release
Validation, optimization, and second red team
Run static validation, source/link checks, router fixtures, and at least ten fresh-context scenario tests. Red-team for omissions, contradictions, unsafe persuasion, cultural transfer, critical information hidden by disclosure, and default-context bloat. Optimize duplication and navigation, then repeat the red team before installation and release.
- All fixtures and validation scripts pass.
- A removal pass proves every new section earns its token and maintenance cost.
- The final report records residual risks and deferred topics rather than claiming completeness.
What could make this plan worse than the current skill?
| Attack | Failure | Plan correction |
|---|---|---|
| “Comprehensive” becomes accumulation | The skill turns into a library that cannot decide what to load. | Require activation conditions, decision impact, canonical ownership, and a removal test for every addition. |
| Theories become deterministic rules | Media richness, cognitive effects, or persuasion findings are applied outside their scope. | Store evidence class and transfer limits; phrase theories as diagnostic lenses and validate with audience evidence. |
| The router creates false precision | A numeric score disguises uncertainty and context dependence. | Return ranked options, tradeoffs, assumptions, and a confidence label—not an objective winner. |
| Critical information is hidden | Progressive disclosure conceals safety, action, deadline, uncertainty, or ownership. | Declare these as always-visible invariants across all formats and overlays. |
| Source volume becomes quality | References multiply while decisions stay unchanged and currentness decays. | No quota; deduplicate, prefer primary/authoritative sources, add freshness metadata and scheduled review. |
| Human-systems modules duplicate each other | Feedback, change, meetings, and narrative repeat the same advice. | Enforce boundaries and link to shared mechanisms rather than copying them. |
| Regulation is frozen into prose | Jurisdiction-specific rules become stale or overgeneralized. | Encode a research-and-review protocol; link current standards with dates and jurisdiction, and require expert review when consequential. |
| Audience research becomes ritual | Teams run interviews or tests without matching the method to the uncertainty. | Route from question to method, define stopping evidence, record negative cases, and distinguish comprehension from behavior and outcomes. |
| AI gets a fashionable catch-all module | Existing provenance, privacy, and product rules are duplicated. | Limit the module to communication consequences of AI mediation; keep evidence custody in provenance and interface states in product communication. |
Result of the second red team
The original expansion concept was narrowed in four ways: channel theory stays inside the existing router instead of becoming an eighth new module; legal/regulatory communication becomes a boundary protocol rather than a static knowledge module; visual, spoken, accessibility, and publication gaps expand existing specialist files; and the target source count is explicitly non-binding. These changes reduce duplication and maintenance risk while preserving the missing capabilities.
Sources that materially changed the plan
This is a condensed research registry for the audit, not the proposed final skill registry. The implementation phase should add full metadata, evidence class, scope, transfer limits, and review dates.
- Dennis, Fuller & Valacich — Media Synchronicity Theory · channel capabilities and conveyance/convergence.
- Clark & Brennan — Grounding in Communication · common ground, evidence of understanding, and repair.
- Daft & Lengel — Media richness and organizational information requirements · historical diagnostic lens.
- GOV.UK Service Manual — User research · operational method families.
- GOV.UK — Moderated usability testing · behavioral observation and task testing.
- U.S. Census Bureau — Cognitive interviewing · testing question interpretation and response process.
- CDC Clear Communication Index · research-based assessment criteria for public communication.
- APA Journal Article Reporting Standards · quantitative, qualitative, and mixed-method reporting.
- COREQ · qualitative interview and focus-group reporting.
- CDC Crisis & Emergency Risk Communication Manual · phase-aware crisis practice.
- WHO — Risk communication and community engagement · emergency trust and participation.
- Evidence-based risk communication review · absolute risk, uncertainty, and visual aids.
- Gigerenzer & colleagues — Natural frequencies · communicating conditional risk.
- Ecker et al. — Misinformation belief and resistance to correction · continued influence and correction design.
- Mann & Thompson — Rhetorical Structure Theory · relations that support textual coherence.
- Text-structure instruction meta-analysis · evidence for structure-aware comprehension support.
- Edmondson — Psychological safety and learning behavior · interpersonal risk in teams.
- Weick, Sutcliffe & Obstfeld — Organizing and sensemaking · change as ongoing meaning construction.
- DeChurch & Mesmer-Magnus — Shared mental models meta-analysis · team cognition and performance.
- GOV.UK Design System — Error messages · clear problem identification and recovery.
- GOV.UK Design System — Confirmation pages · transaction completion and next steps.
- Apple Human Interface Guidelines — Writing · product-language practice.
- NIST AI Risk Management Framework · trustworthiness and risk governance.
- OECD AI principle — Transparency and explainability · awareness, explanation, and challenge.
- Jakesch et al. — AI-mediated communication and trust · interpersonal effects of algorithmic mediation.
- Heer & Bostock — Crowdsourcing graphical perception · empirical chart-decoding performance.
- Section508.gov — Accessibility Conformance Reports · product accessibility documentation.
- ETSI EN 301 549 · European ICT accessibility requirements.
- W3C — Audio description · alternate access to video information.
- Rains — Reactance meta-analysis · autonomy threat and resistance.
- Inoculation theory meta-analysis · prebunking effects and limits.
- Pashler et al. — Learning styles review · evidence against matching instruction to fixed learning styles.