cmd/ghsearchdump: improve prompt

Author: wenyihu6

pkg/cmd/kv/ghsearchdump/prompt.md

@@ -2,28 +2,59 @@
 
 Use this prompt with an LLM to generate a monthly progress summary from the JSON output of `ghsearchdump`.
 
+**Recommended Model:** Claude Opus 4.5 (`claude-opus-4-5-20251101`) - exceptional at narrative writing and strategic synthesis.
+
 ---
 
 You are helping write a monthly progress summary for a software engineering team.
 
 I have JSON containing PRs worked on this month (attached). Please help me create a ~400 word narrative summary suitable for inclusion in a larger milestone document.
 
-Requirements:
-- Write in narrative prose, not bullet lists (e.g., "Wenyi went deep on X and fixed Y")
-- Attribute work to authors by first name (map: wenyihu6=Wenyi, tbg=Tobias, sumeerbhola=Sumeer, angeladietz=Angela)
+**Focus on project advancement, not individual git logs:**
+- Lead with team-level outcomes and what projects advanced (e.g., "Our most notable output this month was...")
+- Explain strategic context: why work matters, what problems it solves, how it fits into larger goals
+- Discuss priority adjustments, trade-offs, and "side quests" that emerged
+- Include honest assessment of challenges, learning curves, and constraints faced
+- Cover meta-work: planning sessions, backlog reviews, epic closures with realistic assessments
+- Mention individuals by first name only when their specific expertise or leadership is relevant (map: wenyihu6=Wenyi, tbg=Tobias, sumeerbhola=Sumeer, angeladietz=Angela)
+
+**Technical requirements:**
+- Write in narrative prose, not bullet lists or per-person summaries
 - Use inline markdown links for Jira compatibility: `[#12345](https://github.com/...)`
-- Group PRs thematically behind descriptions, not one link per PR
+- Group PRs thematically behind descriptions as evidence, not one link per PR
 - Extract Epic IDs using `/Epic[:\s]+CRDB-(\d+)/i` to identify major workstreams
-- Focus on outcomes and themes, not individual PR details
 - Skip bot comments, CI noise, and mechanical changes
 
-Tone: Technical but readable for engineering leadership. Emphasize what was accomplished and why it matters.
+Tone: Technical but readable for engineering leadership. Write like you're explaining project status to stakeholders who care about outcomes, strategy, and realistic progress—not just a list of completed tasks.
 
 ---
 
+## Example Output
+
+Here's an example that demonstrates the desired style—focused on project advancement, strategic context, and honest assessment rather than individual contributions:
+
+> Our small team's most notable output this month was work to align the existing allocator's lease-count balancing for replicas and leases with the multi-metric allocator, in order to prevent thrashing. We are pleased to report that this works.
+>
+> We also spent a lot of time poring over the behavior of existing simulation tests together, to understand details of both the old allocator and mma. It's a steep learning curve, we are accelerating and making many useful testing, interpretability, and observability improvements along the way. Still, we expect to continue spending significant amounts of time in this area, paying down the cognitive debt that has built up in this area of the product and making it more accessible to newcomers.
+>
+> We also began a small but important side quest to disable "follow the workload" lease rebalancing by default (in 26.1+), more background here. By removing this somewhat obscure feature, we would facilitate a (future) full transition off the old allocator. We still need to work through the deprecation plan, so please reach out to us if you have thoughts on this.
+>
+> October also included virtual and in-person planning, so significant amounts of energy went into a backlog review, as well as formulating a plan for the multi-metric allocator in 26.1 and beyond. (TLDR: cloud preview of the multi-metric allocator).
+>
+> We will be closing out our 25.4 epic as PARTIAL. We have met the main goal of a master non-production version of the multi-metric allocator that performs the basic advertised functions, but had hoped to have addressed a number of additional work items related to production readiness at this point. Given the substantial non-project (mainly reactive support and admission control onboarding) burden that the team had to shoulder and the thin staffing, we are content with what we have achieved.
+
+Note how this example:
+- Leads with team outcomes ("Our small team's most notable output")
+- Explains why work matters ("in order to prevent thrashing", "facilitate a future full transition")
+- Honestly discusses challenges ("steep learning curve", "cognitive debt")
+- Mentions priority shifts ("side quest to disable...")
+- Covers planning and meta-work ("backlog review", "formulating a plan")
+- Provides realistic epic assessment ("PARTIAL" with clear context on constraints)
+
 ## Tips
 
 - Attach the JSON file generated by `ghsearchdump` to your prompt
 - You may need to iterate: ask for condensing, different formatting, or additional links
 - Consider a two-phase approach: first get a thematic analysis, then write your own narrative and ask the LLM to "add relevant PR links"
+- Reference the example above when prompting: "Write in a style similar to the example provided"