Customer Support
Umbrella skill for customer-facing support workflows: triage incoming tickets, draft calibrated responses, convert resolutions into KB articles, package escalations, and research customer context. Each mode loads its own reference files on demand.
Scope: Customer-facing support work. Use professional-communication for internal business formats, csuite for strategic decisions, pr-workflow for code PRs.
Mode Detection
Classify the request into exactly one mode. If the request spans modes, choose the primary and note the secondary.
| Mode |
Signal Phrases |
Core Output |
| TRIAGE |
New ticket, categorize, prioritize, route, P1-P4, severity, SLA |
Structured triage assessment with priority, routing, initial response |
| RESPOND |
Draft response, reply to customer, follow up, de-escalate, bad news, decline |
Customer-facing message with tone calibration and internal notes |
| KB |
Knowledge base, document this, FAQ, write article, how-to guide, troubleshooting doc |
Publish-ready KB article with metadata and search optimization |
| ESCALATE |
Escalate, engineering attention, SLA breach, churn risk, multiple customers, leadership |
Structured escalation brief with impact assessment and repro steps |
| RESEARCH |
Look up, investigate, what did we tell them, has this been reported, check history |
Research brief with source attribution and confidence scoring |
Reference Loading Table
Load only the references required by the detected mode.
| Mode |
Reference |
| TRIAGE |
references/triage-methodology.md |
| RESPOND |
references/response-drafting.md |
| KB |
references/knowledge-base.md |
| ESCALATE |
references/triage-methodology.md, references/response-drafting.md |
| RESEARCH |
references/knowledge-base.md |
| Any mode |
references/llm-support-failure-modes.md (always load -- LLM failure awareness is non-negotiable in support) |
Mode: TRIAGE
Framework: PARSE -> CLASSIFY -> ROUTE -> RESPOND
Phase 1: PARSE -- Extract the actual problem from the ticket.
- Core problem vs. stated symptom (customers describe symptoms, not root causes)
- Urgency signals: production down, data loss, blocked, multiple users, time-sensitive
- Emotional state: frustrated, confused, matter-of-fact, escalating
- Customer context: account tier, history, previous tickets if available
Phase 2: CLASSIFY -- Assign category and priority.
Apply the category taxonomy from references/triage-methodology.md:
| Category |
When |
| Bug |
"It used to work and now it doesn't" |
| How-to |
"How do I make it work?" |
| Feature request |
"I want it to work differently" |
| Billing |
Payment, subscription, invoice, refund |
| Account |
Login, permissions, SSO, access |
| Integration |
API, webhook, third-party, sync |
| Security |
Data exposure, unauthorized access, compliance |
| Performance |
Slow, timeout, degraded, unavailable |
Assign priority P1-P4. When in doubt, err higher -- easier to de-escalate than recover from a missed SLA.
| Priority |
Criteria |
SLA Response |
| P1 Critical |
Production down, data loss, security breach, all users |
1 hour |
| P2 High |
Major feature broken, no workaround, many users |
4 hours |
| P3 Medium |
Partial break, workaround exists, small impact |
1 business day |
| P4 Low |
Cosmetic, feature request, general question |
2 business days |
Phase 3: ROUTE -- Determine the right team.
| Route to |
When |
| Tier 1 |
How-to, known issues with docs, billing inquiries, password resets |
| Tier 2 |
Bugs needing investigation, complex config, integration troubleshooting |
| Engineering |
Confirmed bugs needing code fixes, infrastructure, performance degradation |
| Product |
Feature requests with demand, design decisions, workflow gaps |
| Security |
Data access concerns, vulnerability reports, compliance (bypasses tier progression) |
Phase 4: RESPOND -- Draft initial response using the category templates from references/triage-methodology.md.
Gate: Triage output includes: category, priority with justification, routing recommendation, suggested initial response, internal notes.
Mode: RESPOND
Framework: CONTEXT -> CALIBRATE -> DRAFT -> VERIFY
Phase 1: CONTEXT -- Understand the full situation before writing a word.
- Who: customer name, account tier, relationship stage (new/established/frustrated)
- What: situation type (question, issue, escalation, bad news, good news, decline)
- Channel: email, ticket, chat (adjusts length and formality)
- Stakeholder level: end user, manager, executive, technical
- History: previous communications, commitments made, tone of thread
Phase 2: CALIBRATE -- Select tone from references/response-drafting.md.
| Situation |
Tone |
Key Characteristic |
| Good news |
Celebratory |
Forward-looking, enthusiastic |
| Routine update |
Professional |
Clear, concise, friendly |
| Technical response |
Precise |
Accurate, patient, structured |
| Delayed delivery |
Accountable |
Honest, action-oriented |
| Bad news / won't-fix |
Candid |
Direct, empathetic, alternative-offering |
| Issue / outage |
Urgent |
Transparent, actionable, reassuring |
| Escalation |
Executive |
Composed, ownership-taking, plan-presenting |
| Billing |
Precise |
Factual, resolution-focused |
Adjust by relationship stage:
- New customer: More formal, extra context, proactive help
- Established: Warm, direct, reference shared history
- Frustrated: Extra empathy first, concrete plan, shorter feedback loops
Phase 3: DRAFT -- Write the response following the structure:
- Acknowledgment (1-2 sentences) -- show you understand their situation
- Core message (1-3 paragraphs) -- the actual answer, update, or information
- Next steps (1-3 bullets) -- what you will do, what they need to do, when they hear from you
- Close (1 sentence) -- warm, professional, available
Phase 4: VERIFY -- Run the quality checks before presenting.
Gate: Draft includes internal notes covering: rationale for tone, facts to verify before sending, risk factors, follow-up actions needed.
Mode: KB
Framework: SOURCE -> STRUCTURE -> DRAFT -> OPTIMIZE
Phase 1: SOURCE -- Understand what you're documenting.
- Original problem, question, or error
- Resolution, workaround, or answer
- Who this affects (user type, plan level, configuration)
- Frequency: one-off or recurring
- Article type: how-to, troubleshooting, FAQ, known issue, reference
Phase 2: STRUCTURE -- Choose the right template from references/knowledge-base.md.
| Type |
Purpose |
Structure |
| How-to |
Step-by-step task completion |
Prerequisites -> Steps -> Verify -> Common Issues |
| Troubleshooting |
Diagnose and fix a problem |
Symptoms -> Cause -> Solution(s) -> Prevention |
| FAQ |
Quick answer to common question |
Direct Answer -> Details -> Related Questions |
| Known issue |
Document a bug with workaround |
Status -> Symptoms -> Workaround -> Fix Timeline |
Phase 3: DRAFT -- Write the article following formatting standards.
- Title: specific, searchable, uses customer language ("How to configure SSO with Okta" not "SSO Setup")
- Opening sentence: restate the problem in plain language
- Headers for scannability. Numbered lists for sequences. Bullet lists for non-sequential items.
- Code blocks for commands, error messages, configuration values
- Short paragraphs (2-4 sentences max). One idea per section.
Phase 4: OPTIMIZE -- Search optimization and metadata.
- Include exact error messages (customers copy-paste into search)
- Use customer language, not internal jargon
- Add common synonyms (delete/remove, dashboard/home page, export/download)
- Tag with product areas matching customer mental models
- Set review date and identify SME for technical verification
Gate: Article includes metadata (title, type, category, tags, audience), full content, publishing notes (source, related articles, review needed, suggested review date).
Mode: ESCALATE
Framework: CONFIRM -> GATHER -> ASSESS -> PACKAGE
Phase 1: CONFIRM -- Verify this warrants escalation.
Escalate when:
- Bug confirmed, needs code fix
- Multiple customers affected (3+ = pattern)
- Production down or data at risk
- SLA breach imminent or occurred
- Customer threatening churn
- Requires access or authority beyond support tier
Handle in support when:
- Documented solution or workaround exists
- Configuration or setup issue you can resolve
- Known limitation with documented alternative
Phase 2: GATHER -- Pull together all context.
- Timeline: when it started, how long the customer has waited
- What's been tried: troubleshooting steps and results
- Reproduction steps (for bugs): from clean state, specific values, environment details, frequency, evidence
- Related tickets and pattern detection
Phase 3: ASSESS -- Quantify business impact.
| Dimension |
Assess |
| Breadth |
How many customers/users? Growing? |
| Depth |
Blocked vs. inconvenienced? |
| Duration |
How long? Getting worse? |
| Revenue |
ARR at risk? Deals affected? |
| Contractual |
SLA breach? Contractual obligations? |
Severity shorthand:
- Critical: Production down, data at risk, security breach. Immediate attention.
- High: Major function broken, key customer blocked, SLA at risk. Same day.
- Medium: Significant issue with workaround, important but not urgent. This week.
Phase 4: PACKAGE -- Structure the escalation brief.
Determine target:
- Engineering: confirmed bugs, infrastructure, code changes needed
- Product: feature gaps, design decisions, competing priorities
- Security: data exposure, vulnerability, compliance (bypasses normal tier progression)
- Leadership: high-revenue churn risk, SLA breach on critical account, policy exception needed
Include: severity, target team, impact summary, issue description, what's been tried, reproduction steps, customer communication status, specific ask with deadline, supporting context.
Gate: Escalation brief complete. Follow-up cadence set (Critical: every 2h internal / 2-4h customer. High: every 4h / 4-8h. Medium: daily / 1-2 business days).
Mode: RESEARCH
Framework: SCOPE -> SEARCH -> SYNTHESIZE -> CAPTURE
Phase 1: SCOPE -- Define what you're looking for.
Research types:
- Customer question: needs a factual answer
- Issue investigation: has this been reported, what's the workaround
- Account context: what was previously communicated
- Topic research: best practices, general domain knowledge
Clarify: factual vs. contextual, audience (internal vs. customer), scope boundaries.
Phase 2: SEARCH -- Work through source tiers systematically.
| Tier |
Source Type |
Confidence |
| 1 |
Official docs, KB, policies |
High |
| 2 |
CRM, support tickets, meeting notes |
Medium-High |
| 3 |
Chat, email, calendar |
Medium |
| 4 |
Web, forums, third-party docs |
Low-Medium |
| 5 |
Inference, analogies, best practices |
Low -- flag explicitly |
Cross-reference across multiple sources.
Phase 3: SYNTHESIZE -- Compile findings with attribution.
- Lead with the bottom-line answer
- Assign confidence level (High / Medium / Low / Unable to Determine)
- Attribute every finding to its source
- Note contradictions explicitly -- present both sides, recommend the conservative answer for customer-facing use
- Identify gaps and unknowns
Phase 4: CAPTURE -- Suggest knowledge preservation.
If the research took significant effort, was a common question, or corrected a misunderstanding, offer to create a KB article or FAQ entry. Knowledge capture prevents duplicate research.
Gate: Research brief includes: direct answer, confidence level, key findings with source attribution, gaps and unknowns, recommended next steps.
Cross-Mode Patterns
These apply regardless of mode. Internalize them.
Empathy is not performance. Acknowledge the customer's situation genuinely. "I understand how frustrating this must be" is fine when they're frustrated. The same phrase when they asked a simple how-to question is patronizing. Match the emotional register. Read references/llm-support-failure-modes.md on tone mismatch.
Own it. Active voice. "We" not "the system." "I'll investigate" not "this will be investigated." Take responsibility where appropriate. Take ownership in all customer-facing communication.
Specificity over reassurance. "I'll update you by Friday at 3pm" beats "I'll get back to you soon." Concrete details build trust. Vague reassurance erodes it.
Close the loop. Every interaction ends with clear next steps: what you will do, what they need to do, when they hear from you next.
Confirm feature existence before stating it. Say 'let me check' when uncertain. A wrong answer that sends a customer down a dead-end path is worse than "let me check and get back to you." See references/llm-support-failure-modes.md.
Limit commitments to actions within your authority. No timeline commitments on behalf of engineering. No policy exceptions without approval. No "we'll definitely build that." The trust cost of a broken promise exceeds the short-term relief of making one.
1---2name: customer-support3description: Customer support workflows — ticket triage, response drafting, knowledge base articles, escalation handling, customer research. Use when triaging support tickets, drafting customer responses, creating KB articles, managing escalations, or researching customer context.4---5
6# Customer Support
7
8Umbrella skill for customer-facing support workflows: triage incoming tickets, draft calibrated responses, convert resolutions into KB articles, package escalations, and research customer context. Each mode loads its own reference files on demand.
9
10**Scope**: Customer-facing support work. Use professional-communication for internal business formats, csuite for strategic decisions, pr-workflow for code PRs.
11
12---
13
14## Mode Detection
15
16Classify the request into exactly one mode. If the request spans modes, choose the primary and note the secondary.
17
18| Mode | Signal Phrases | Core Output |
19|------|---------------|-------------|
20| **TRIAGE** | New ticket, categorize, prioritize, route, P1-P4, severity, SLA | Structured triage assessment with priority, routing, initial response |
21| **RESPOND** | Draft response, reply to customer, follow up, de-escalate, bad news, decline | Customer-facing message with tone calibration and internal notes |
22| **KB** | Knowledge base, document this, FAQ, write article, how-to guide, troubleshooting doc | Publish-ready KB article with metadata and search optimization |
23| **ESCALATE** | Escalate, engineering attention, SLA breach, churn risk, multiple customers, leadership | Structured escalation brief with impact assessment and repro steps |
24| **RESEARCH** | Look up, investigate, what did we tell them, has this been reported, check history | Research brief with source attribution and confidence scoring |
25
26---
27
28## Reference Loading Table
29
30Load only the references required by the detected mode.
31
32| Mode | Reference |
33|------|-----------|
34| TRIAGE | `references/triage-methodology.md` |
35| RESPOND | `references/response-drafting.md` |
36| KB | `references/knowledge-base.md` |
37| ESCALATE | `references/triage-methodology.md`, `references/response-drafting.md` |
38| RESEARCH | `references/knowledge-base.md` |
39| Any mode | `references/llm-support-failure-modes.md` (always load -- LLM failure awareness is non-negotiable in support) |
40
41---
42
43## Mode: TRIAGE
44
45**Framework**: PARSE -> CLASSIFY -> ROUTE -> RESPOND
46
47**Phase 1: PARSE** -- Extract the actual problem from the ticket.
48
49- Core problem vs. stated symptom (customers describe symptoms, not root causes)
50- Urgency signals: production down, data loss, blocked, multiple users, time-sensitive
51- Emotional state: frustrated, confused, matter-of-fact, escalating
52- Customer context: account tier, history, previous tickets if available
53
54**Phase 2: CLASSIFY** -- Assign category and priority.
55
56Apply the category taxonomy from `references/triage-methodology.md`:
57
58| Category | When |
59|----------|------|
60| Bug | "It used to work and now it doesn't" |
61| How-to | "How do I make it work?" |
62| Feature request | "I want it to work differently" |
63| Billing | Payment, subscription, invoice, refund |
64| Account | Login, permissions, SSO, access |
65| Integration | API, webhook, third-party, sync |
66| Security | Data exposure, unauthorized access, compliance |
67| Performance | Slow, timeout, degraded, unavailable |
68
69Assign priority P1-P4. When in doubt, err higher -- easier to de-escalate than recover from a missed SLA.
70
71| Priority | Criteria | SLA Response |
72|----------|----------|-------------|
73| P1 Critical | Production down, data loss, security breach, all users | 1 hour |
74| P2 High | Major feature broken, no workaround, many users | 4 hours |
75| P3 Medium | Partial break, workaround exists, small impact | 1 business day |
76| P4 Low | Cosmetic, feature request, general question | 2 business days |
77
78**Phase 3: ROUTE** -- Determine the right team.
79
80| Route to | When |
81|----------|------|
82| Tier 1 | How-to, known issues with docs, billing inquiries, password resets |
83| Tier 2 | Bugs needing investigation, complex config, integration troubleshooting |
84| Engineering | Confirmed bugs needing code fixes, infrastructure, performance degradation |
85| Product | Feature requests with demand, design decisions, workflow gaps |
86| Security | Data access concerns, vulnerability reports, compliance (bypasses tier progression) |
87
88**Phase 4: RESPOND** -- Draft initial response using the category templates from `references/triage-methodology.md`.
89
90**Gate**: Triage output includes: category, priority with justification, routing recommendation, suggested initial response, internal notes.
91
92---
93
94## Mode: RESPOND
95
96**Framework**: CONTEXT -> CALIBRATE -> DRAFT -> VERIFY
97
98**Phase 1: CONTEXT** -- Understand the full situation before writing a word.
99
100- Who: customer name, account tier, relationship stage (new/established/frustrated)
101- What: situation type (question, issue, escalation, bad news, good news, decline)
102- Channel: email, ticket, chat (adjusts length and formality)
103- Stakeholder level: end user, manager, executive, technical
104- History: previous communications, commitments made, tone of thread
105
106**Phase 2: CALIBRATE** -- Select tone from `references/response-drafting.md`.
107
108| Situation | Tone | Key Characteristic |
109|-----------|------|--------------------|
110| Good news | Celebratory | Forward-looking, enthusiastic |
111| Routine update | Professional | Clear, concise, friendly |
112| Technical response | Precise | Accurate, patient, structured |
113| Delayed delivery | Accountable | Honest, action-oriented |
114| Bad news / won't-fix | Candid | Direct, empathetic, alternative-offering |
115| Issue / outage | Urgent | Transparent, actionable, reassuring |
116| Escalation | Executive | Composed, ownership-taking, plan-presenting |
117| Billing | Precise | Factual, resolution-focused |
118
119Adjust by relationship stage:
120- **New customer**: More formal, extra context, proactive help
121- **Established**: Warm, direct, reference shared history
122- **Frustrated**: Extra empathy first, concrete plan, shorter feedback loops
123
124**Phase 3: DRAFT** -- Write the response following the structure:
125
1261. **Acknowledgment** (1-2 sentences) -- show you understand their situation
1272. **Core message** (1-3 paragraphs) -- the actual answer, update, or information
1283. **Next steps** (1-3 bullets) -- what you will do, what they need to do, when they hear from you
1294. **Close** (1 sentence) -- warm, professional, available
130
131**Phase 4: VERIFY** -- Run the quality checks before presenting.
132
133- [ ] Tone matches situation and relationship stage
134- [ ] No unauthorized commitments (timelines, features, exceptions)
135- [ ] No product roadmap details that shouldn't be external
136- [ ] Clear ownership and next steps
137- [ ] Appropriate length for channel
138- [ ] No corporate jargon, no passive voice to dodge accountability
139
140**Gate**: Draft includes internal notes covering: rationale for tone, facts to verify before sending, risk factors, follow-up actions needed.
141
142---
143
144## Mode: KB
145
146**Framework**: SOURCE -> STRUCTURE -> DRAFT -> OPTIMIZE
147
148**Phase 1: SOURCE** -- Understand what you're documenting.
149
150- Original problem, question, or error
151- Resolution, workaround, or answer
152- Who this affects (user type, plan level, configuration)
153- Frequency: one-off or recurring
154- Article type: how-to, troubleshooting, FAQ, known issue, reference
155
156**Phase 2: STRUCTURE** -- Choose the right template from `references/knowledge-base.md`.
157
158| Type | Purpose | Structure |
159|------|---------|-----------|
160| How-to | Step-by-step task completion | Prerequisites -> Steps -> Verify -> Common Issues |
161| Troubleshooting | Diagnose and fix a problem | Symptoms -> Cause -> Solution(s) -> Prevention |
162| FAQ | Quick answer to common question | Direct Answer -> Details -> Related Questions |
163| Known issue | Document a bug with workaround | Status -> Symptoms -> Workaround -> Fix Timeline |
164
165**Phase 3: DRAFT** -- Write the article following formatting standards.
166
167- Title: specific, searchable, uses customer language ("How to configure SSO with Okta" not "SSO Setup")
168- Opening sentence: restate the problem in plain language
169- Headers for scannability. Numbered lists for sequences. Bullet lists for non-sequential items.
170- Code blocks for commands, error messages, configuration values
171- Short paragraphs (2-4 sentences max). One idea per section.
172
173**Phase 4: OPTIMIZE** -- Search optimization and metadata.
174
175- Include exact error messages (customers copy-paste into search)
176- Use customer language, not internal jargon
177- Add common synonyms (delete/remove, dashboard/home page, export/download)
178- Tag with product areas matching customer mental models
179- Set review date and identify SME for technical verification
180
181**Gate**: Article includes metadata (title, type, category, tags, audience), full content, publishing notes (source, related articles, review needed, suggested review date).
182
183---
184
185## Mode: ESCALATE
186
187**Framework**: CONFIRM -> GATHER -> ASSESS -> PACKAGE
188
189**Phase 1: CONFIRM** -- Verify this warrants escalation.
190
191Escalate when:
192- Bug confirmed, needs code fix
193- Multiple customers affected (3+ = pattern)
194- Production down or data at risk
195- SLA breach imminent or occurred
196- Customer threatening churn
197- Requires access or authority beyond support tier
198
199Handle in support when:
200- Documented solution or workaround exists
201- Configuration or setup issue you can resolve
202- Known limitation with documented alternative
203
204**Phase 2: GATHER** -- Pull together all context.
205
206- Timeline: when it started, how long the customer has waited
207- What's been tried: troubleshooting steps and results
208- Reproduction steps (for bugs): from clean state, specific values, environment details, frequency, evidence
209- Related tickets and pattern detection
210
211**Phase 3: ASSESS** -- Quantify business impact.
212
213| Dimension | Assess |
214|-----------|--------|
215| Breadth | How many customers/users? Growing? |
216| Depth | Blocked vs. inconvenienced? |
217| Duration | How long? Getting worse? |
218| Revenue | ARR at risk? Deals affected? |
219| Contractual | SLA breach? Contractual obligations? |
220
221Severity shorthand:
222- **Critical**: Production down, data at risk, security breach. Immediate attention.
223- **High**: Major function broken, key customer blocked, SLA at risk. Same day.
224- **Medium**: Significant issue with workaround, important but not urgent. This week.
225
226**Phase 4: PACKAGE** -- Structure the escalation brief.
227
228Determine target:
229- **Engineering**: confirmed bugs, infrastructure, code changes needed
230- **Product**: feature gaps, design decisions, competing priorities
231- **Security**: data exposure, vulnerability, compliance (bypasses normal tier progression)
232- **Leadership**: high-revenue churn risk, SLA breach on critical account, policy exception needed
233
234Include: severity, target team, impact summary, issue description, what's been tried, reproduction steps, customer communication status, specific ask with deadline, supporting context.
235
236**Gate**: Escalation brief complete. Follow-up cadence set (Critical: every 2h internal / 2-4h customer. High: every 4h / 4-8h. Medium: daily / 1-2 business days).
237
238---
239
240## Mode: RESEARCH
241
242**Framework**: SCOPE -> SEARCH -> SYNTHESIZE -> CAPTURE
243
244**Phase 1: SCOPE** -- Define what you're looking for.
245
246Research types:
247- **Customer question**: needs a factual answer
248- **Issue investigation**: has this been reported, what's the workaround
249- **Account context**: what was previously communicated
250- **Topic research**: best practices, general domain knowledge
251
252Clarify: factual vs. contextual, audience (internal vs. customer), scope boundaries.
253
254**Phase 2: SEARCH** -- Work through source tiers systematically.
255
256| Tier | Source Type | Confidence |
257|------|------------|------------|
258| 1 | Official docs, KB, policies | High |
259| 2 | CRM, support tickets, meeting notes | Medium-High |
260| 3 | Chat, email, calendar | Medium |
261| 4 | Web, forums, third-party docs | Low-Medium |
262| 5 | Inference, analogies, best practices | Low -- flag explicitly |
263
264Cross-reference across multiple sources.
265
266**Phase 3: SYNTHESIZE** -- Compile findings with attribution.
267
268- Lead with the bottom-line answer
269- Assign confidence level (High / Medium / Low / Unable to Determine)
270- Attribute every finding to its source
271- Note contradictions explicitly -- present both sides, recommend the conservative answer for customer-facing use
272- Identify gaps and unknowns
273
274**Phase 4: CAPTURE** -- Suggest knowledge preservation.
275
276If the research took significant effort, was a common question, or corrected a misunderstanding, offer to create a KB article or FAQ entry. Knowledge capture prevents duplicate research.
277
278**Gate**: Research brief includes: direct answer, confidence level, key findings with source attribution, gaps and unknowns, recommended next steps.
279
280---
281
282## Cross-Mode Patterns
283
284These apply regardless of mode. Internalize them.
285
286**Empathy is not performance.** Acknowledge the customer's situation genuinely. "I understand how frustrating this must be" is fine when they're frustrated. The same phrase when they asked a simple how-to question is patronizing. Match the emotional register. Read `references/llm-support-failure-modes.md` on tone mismatch.
287
288**Own it.** Active voice. "We" not "the system." "I'll investigate" not "this will be investigated." Take responsibility where appropriate. Take ownership in all customer-facing communication.
289
290**Specificity over reassurance.** "I'll update you by Friday at 3pm" beats "I'll get back to you soon." Concrete details build trust. Vague reassurance erodes it.
291
292**Close the loop.** Every interaction ends with clear next steps: what you will do, what they need to do, when they hear from you next.
293
294**Confirm feature existence before stating it.** Say 'let me check' when uncertain. A wrong answer that sends a customer down a dead-end path is worse than "let me check and get back to you." See `references/llm-support-failure-modes.md`.
295
296**Limit commitments to actions within your authority.** No timeline commitments on behalf of engineering. No policy exceptions without approval. No "we'll definitely build that." The trust cost of a broken promise exceeds the short-term relief of making one.