Turn support questions into proposed knowledge improvements Leading Deeper with AI — Jasmine James https://jasminejames.dev/leading-deeper/support-improvement Skill source (GitHub): https://github.com/jasminelsjames/leading-deeper-ai-skills/blob/e1d2e65d3922c5e376fc0c6e8cfbb2f85722b1a7/skills/review-feedback/SKILL.md Adaptable template based on the talk; not the original internal prompt. Version 1 / 2026-11-04 START HERE: ONE FIRST RUN Turn two questions into a proposed improvement. Use the AI tool approved for your work with supplied material. Wording and layout will vary by model; look for the evidence, uncertainty and review decisions shown here. Connections and recurring runs depend on your setup. All examples use invented sources and results. They are illustrations, not reports from Jasmine’s workplace or measured outcomes. FICTIONAL INPUTS **[S1] Two selected fictional threads · 26–30 October 2026:** Thread A: a Harbor v2 owner asks, 'Which part of Settings contains Export?' Thread B: a Harbor v2 member asks, 'Why do I not see Export?' These are the only threads in the sample; no resolution evidence was supplied. **[S2] Current quick-start guide:** To download a CSV, export your project data from Settings. **[S3] Approved Harbor v2 product note and review scope:** Only project owners can create CSV exports: Project settings → Data → Export. Members do not see Export. A workspace administrator can identify an unknown project owner. Reviewer and publication owner: Lee. Compare an owner asking where to click, a member missing the action, and an unknown product version. For an unknown version, ask for the version rather than assuming these steps apply. No replay runner is available. WORKFLOW 1. Human — Bound the evidence and decision: Supply the selected threads, existing guide and approved product note. Name Lee as reviewer and ask for one proposed change, not publication. 2. AI — Connect the gap to a minimal patch: Identify what the current guide omits, propose replacement wording supported by the product note, and prepare the representative comparison. Label it unrun because no replay runner is available. 3. Human — Test before deciding: Lee checks the evidence, runs the comparison with the actual answering setup, and approves, revises or rejects the wording before any shared guide changes. ILLUSTRATIVE OUTPUT Export guidance review item · unpublished draft **Finding and evidence:** Both sampled threads concern locating or accessing Export [S1]. The current guide omits the owner-only restriction and exact navigation [S2], which are present in approved v2 guidance [S3]. This supports a targeted clarification in this sample; it does not establish the frequency of the gap across all support. **Proposed replacement:** For Harbor v2, only a project owner can create a CSV export. Owners open Project settings → Data → Export. Members do not see Export. If you do not know the project owner, ask your workspace administrator to help identify them. [S3] Update the existing quick-start entry rather than adding a second guide. **Comparison plan · not run:** Ask the same three questions with the current and proposed material: owner asking where to click → expect the exact path; member missing Export → expect the owner-only rule and administrator route; version unknown → expect a version question before v2 instructions. Record each actual answer, supporting reference and unsupported claim. No before/after result or pass is claimed. [S3] **Recommendation and uncertainty:** Recommend review and comparison before a publication decision. The change is source-supported, but its effect on answers and later support questions is unmeasured. Lee owns approval, revision or rejection. No guidance has been published. WHAT TO LOOK FOR ACROSS MODELS - Finding and evidence - Proposed replacement - Comparison plan · not run - Recommendation and uncertainty HUMAN REVIEW - Is the current quick-start entry the right source to change, and does every proposed sentence follow the approved note? - Does the actual comparison preserve the unknown-version question and avoid unsupported permission advice? MCP AND TOOL SETUP Use read access to the scoped support threads and current knowledge. A separate runner is needed to execute replay comparisons; otherwise return a test plan labeled not run. ### Read team conversations - Needed when: When analyzing live support history instead of a supplied export. - MCP or tool examples: Slack’s official MCP server; An organization-approved Teams-capable MCP, such as Work IQ where available - Tools must support: Search approved channels within the reporting period; Read relevant messages and complete threads; Retain message links, dates and thread context - Access needed: Read access to the channels and threads included in the task. Verify the installed server exposes the needed search and read tools; Teams availability varies by tenant, licensing and administrator policy. Sending messages and administration are not needed. - Without the connection: Provide approved thread exports or excerpts with dates, source links and enough surrounding context to interpret them. - Provider documentation: https://docs.slack.dev/ai/slack-mcp-server/ ; https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/add-work-iq ; https://learn.microsoft.com/en-us/microsoft-agent-365/tooling-servers-overview ### Read current knowledge and source artifacts - Needed when: When comparing a proposed patch with current approved knowledge that is not supplied directly. - MCP or tool examples: Local file or repository tools; An approved documentation/search connector or MCP; GitHub or GitLab MCP for remote repository material - Tools must support: Read the authorized documentation, code and referenced artifacts; Search relevant content; Capture stable source identifiers, dates and versions - Access needed: Read access to the selected sources only. Drafting these outputs does not require publishing or editing the source of truth. - Without the connection: Provide an approved local snapshot or excerpts with source identifiers and relevant versions. - Provider documentation: https://github.com/github/github-mcp-server ; https://docs.gitlab.com/user/model_context_protocol/mcp_server/ ### Replay representative questions - Needed when: Only when claiming an executed comparison between current and proposed material. - MCP or tool examples: A local replay script or evaluation harness; An approved test runner; A custom MCP exposing the organization’s replay/evaluation workflow - Tools must support: Run the same representative questions against current and proposed material; Capture inputs, outputs and actual comparison results; Include cases that should still escalate to a person - Access needed: Access to the approved test material and runner. Keep proposed knowledge separate from published sources. Publishing or messaging permissions are not needed. - Without the connection: Return a concrete replay plan labeled not run and leave the proposal unpublished. Do not turn a reasoned prediction into a reported test result. Provide approved thread exports and a knowledge snapshot. Propose a patch locally; run comparisons with an available local test runner or explicitly leave the replay plan unrun. MODEL STARTING POINT Reviewed October 30, 2026. Start with Claude Sonnet 5.5 (claude-sonnet-5-5). A balanced starting point for a minimal knowledge patch and analysis of replay results. Move up to Claude Opus 5.5 (claude-opus-5-5) when: Contradictions, difficult regressions or changes to the knowledge structure require deeper investigation. ALTERNATIVES - GPT-6.1 Sol (gpt-6.1-sol): An alternative for drafting and reviewing the proposal in an approved OpenAI-based agent host. - Claude Haiku 5.5 (claude-haiku-5-5): Use for initial thread classification or gap triage, then review the proposed patch with the stronger starting model. HUMAN RESPONSIBILITY Run the replay comparison with actual tools or label it not run. Model judgment is not a test result; a named person still approves, revises or rejects publication. Suggested fit based on official provider documentation, not our own speed or quality benchmarks. Test on representative examples before relying on a model. A model does not connect your tools or grant access. MCP support, available tools and model selection depend on the agent host and your organization's approved setup. API PRICE REFERENCE USD per one million input and output tokens, using standard first-party API pricing. These are not subscription prices or total workflow costs. Tool calls, reasoning, retries, caching and different tokenizers affect the final bill. CLAUDE SONNET 5.5 Provider: Anthropic API model ID: claude-sonnet-5-5 Context: 1 million tokens Standard API: $2 input / $10 output per 1M tokens. Model documentation: https://platform.claude.com/docs/en/models/sonnet-5-5/overview Pricing: https://platform.claude.com/docs/en/about-claude/pricing CLAUDE OPUS 5.5 Provider: Anthropic API model ID: claude-opus-5-5 Context: 1 million tokens Standard API: $4 input / $20 output per 1M tokens. Model documentation: https://platform.claude.com/docs/en/models/opus-5-5/overview Pricing: https://platform.claude.com/docs/en/about-claude/pricing GPT-6.1 SOL Provider: OpenAI API model ID: gpt-6.1-sol Context: 1.05 million tokens Standard API: $2 input / $10 output per 1M tokens. Model documentation: https://developers.openai.com/api/docs/models/gpt-6.1-sol Pricing: https://developers.openai.com/api/docs/pricing CLAUDE HAIKU 5.5 Provider: Anthropic API model ID: claude-haiku-5-5 Context: 1 million tokens Standard API: $0.10 input / $0.50 output per 1M tokens. For requests with prompts up to 100,000 tokens. Standard API: $0.50 input / $2.5 output per 1M tokens. For requests with prompts over 100,000 tokens; the higher rates apply to that request. Model documentation: https://platform.claude.com/docs/en/models/haiku-5-5/overview Pricing: https://platform.claude.com/docs/en/about-claude/pricing Machine-readable catalog: https://jasminejames.dev/leading-deeper/model-guidance.json PROMPT Review these support threads for improvements to our shared knowledge. PREFLIGHT Verify which tools are actually available and which source locations they can access. Do not assume that naming an MCP server means it is connected. Local file and terminal tools are valid and do not necessarily use MCP. If required access is missing, ask for an approved export or excerpt; state what cannot be checked or run. INPUTS Reporting period: [fill in] Authorized threads: [provide] Current approved knowledge sources: [provide] Reviewed criteria / expected behavior: [provide, or ask for a proposal to review] Representative questions / expected supporting evidence: [provide] Reviewer and publication owner: [fill in] WORK Use the reviewed criteria or expected behavior to assess the supplied material. If criteria are absent, propose them for review instead of silently choosing a standard. Treat AI assessments as candidate findings until a person verifies the evidence. Identify recurring questions, missing context, stale guidance and conflicting instructions. Link each finding to evidence from the supplied material. Select a small, useful improvement rather than summarizing everything. Check whether an existing source should be updated instead of creating duplicate guidance. Propose a minimal patch and explain why it addresses the gap. TEST Replay the representative questions against the current and proposed material. Where you have tools to run that comparison, record inputs, outputs and actual results. Otherwise, provide a test plan clearly labeled not run. Include cases that should still escalate to a person, and inspect for contradictory or unsafe guidance. Do not call the proposal an improvement merely because it produces a longer answer. OUTPUT Return the finding and evidence, proposed patch, replay comparison or unrun test plan, known regressions or uncertainties, and a recommendation to approve, revise or reject. STOP Leave the proposal unpublished. A named human reviews the evidence and decides whether to approve, revise, reject or change the structure of the source. Never change or publish shared guidance automatically. WHAT A PERSON REVIEWS - Is this a real recurring gap, and is the proposed source the right place to fix it? - Did the replay expose regressions, overconfident answers or cases that should escalate? - Would rejecting the patch or changing the source structure be a better decision? EVIDENCE TO CHECK - Count proposed, tested, reviewed and published changes separately; do not imply they form one cohort unless tracked that way. - Check later questions for the targeted gap using the same definition and comparable periods. - Keep assistance, confirmed resolution and human attention as distinct measures.