## 1
1. A Manager's Embarrassment: The Issue Tracker Has Become a "Junk Room"
Anyone who has led a small team of five or fewer has probably experienced this: fast product iterations cause the issue tracker to balloon. User feedback, bug reports, feature requests, duplicate questions, "+1" follow-ups, and even off-topic chatter all get mixed into a single list.
The result is three classic pain points:
Pain point one: Nobody wants to open the issue list. When a developer sees 200+ uncategorized issues sitting in the backlog, their first reaction isn't to work through them—it's to procrastinate. A genuinely urgent bug may be buried on page 7, while a pinned emotional rant sits at the top.
Pain point two: The cost of manual cleanup is absurd. We once had an intern spend a week cleaning up issues, producing an Excel spreadsheet in the end—plenty of labels applied, but with inconsistent classification standards, and a week later the newly added issues had rendered the spreadsheet obsolete.
Pain point three: Decision-making lacks grounding. Trying to determine "should this module be refactored?" or "what actually hurts users the most?" relied solely on people's impressions from browsing issues. Impressions are unreliable, and meetings devolve into everyone talking past each other.
Essentially, the problem with the issue tracker isn't a technical problem—it's a problem of information processing workflow. And information triage, extraction, and categorization happen to be exactly what today's AI is best at.
2. What AI Can Do for You Here
Let's be clear about the division of labor: AI doesn't fix bugs, and it doesn't make product decisions. What AI does is turn unstructured information into structured decision inputs. Specifically, it can do four things:
- Batch triage: Given a classification system, AI can categorize hundreds of issues one by one—genuine bugs, feature requests, documentation issues, duplicates, invalid entries—and provide reasoning for each classification.
- Information extraction: Pull out the core elements from lengthy discussion threads: symptom, reproduction conditions, scope of impact, modules involved, and degree of community consensus.
- Merging and deduplication: Identify semantically duplicate issues, even when they're phrased completely differently. This is the most time-consuming part of manual cleanup.
- Trend summarization: Hand off work like "aggregate feature requests from the past month by module" to AI, producing summaries that management can read directly.
3. My Implementation Process: Four Steps, with Human Checkpoints at Every Step
As a manager, I care about two things: whether the process runs smoothly, and how risk is controlled. After all, issues may contain user information and internal details, and dumping them all into AI poses risks.
Step one: Sanitization and export. Use GitHub CLI to export issues as JSON, then run them through a local sanitization script that replaces emails, customer names, and internal domains with placeholders. No AI in this step—pure rule-based processing, which offers maximum controllability.
Step two: Define the classification system. This is a step the manager must personally participate in. Classification standards aren't invented by AI—they're defined by the team. Ours is a six-level labeling system: bug-confirmed / bug-unverified / feature-request / documentation / duplicate / invalid. The simpler the system, the more stable the AI's categorization.
Step three: Batch AI processing. Feed issues to the model in batches of 20, requiring structured JSON output that includes classification, confidence, a one-sentence summary, and duplicate associations. Items with confidence below the threshold go into a manual review queue—this is the key risk-control design: AI handles the 80% of clear-cut items, humans handle the 20% of ambiguous ones, rather than humans processing 100% or blindly trusting AI's 100%.
Step four: Manual spot-checking and write-back. Weekly, spot-check 10% of the AI results; if systematic bias is found, revise the classification definitions in the prompt. Final results are written back to GitHub labels so the whole team benefits, rather than living in a private Excel spreadsheet.
On cost: the token expenditure for processing a few hundred issues is quite modest—check your model's official pricing page for specifics. I recommend running a small trial batch before scaling up.
4. Prompt Template: Ready to Copy and Use
你是一位开源项目的issue分诊助手。请对以下issue逐条分析,严格输出JSON。
分类体系(只能从中选择):
- bug-确认:有明确复现路径的缺陷
- bug-待验证:疑似缺陷但信息不足
- 功能请求:希望增加新能力
- 文档:文档缺失、错误或不清晰
- 重复:与已存在issue语义重复
- 无效:跑题、情绪化表达、信息不足且无法跟进
输出格式:
[{
"issue编号": "",
"分类": "",
"置信度": "高/中/低",
"一句话摘要": "",
"关键信息": {
"现象": "",
"复现条件": "",
"涉及模块": ""
},
"归类理由": "",
"疑似重复于": "issue编号或null"
}]
规则:
1. 只依据issue内容判断,不臆测
2. 信息不足时选"bug-待验证"或"无效",不要猜
3. 置信度为"低"的条目,归类理由必须写清缺失了什么信息
4. 摘要不超过50字,给没读过原文的人看
待分析issue:
{issues内容}One detail worth noting: requiring the AI to state what's missing when confidence is low tells the reviewer exactly what to follow up on in the issue—which actually makes review more efficient.
5. Before and After: What Changed Isn't Efficiency—It's the Process
Before AI: Weekly manual cleanup producing a spreadsheet that quickly became outdated; developers distrusted the issue tracker, and urgent bugs were passed along verbally; retrospectives featured rounds of "I think users care most about X" versus "No, it's Y."
After AI: The triage runs automatically every Monday; low-confidence items go to the review queue and are handled in half an hour; the issue tracker's labels are trustworthy, and developers are willing to pick up issues by label; monthly retrospectives cite the aggregated summaries directly, moving discussions from "what's happening" to "what do we do about it."
Let's do a rough calculation: what used to be roughly 20 hours of manual cleanup per week has been compressed to about 2 hours of review plus spot-checking. But more important than labor hours—cleanup went from "one person's burden" to "a routine part of the process", and that's the real reason things stopped spiraling out of control.
6. A Few Reminders for Managers
- Define standards before deploying AI; the classification system is set by humans, and AI is merely the executor;
- Low-confidence results must have a human outlet, or errors will accumulate silently;
- Sanitize data first, even if it's just out of habit;
- AI conclusions should support decision-making, not replace direct communication with users.
If you're also looking for a stable model API for batch text processing tasks like this, you can try this gateway service—registration here: https://api.thistoken.ai/register
Turning your issue tracker from a "junk room" back into an "information source" might be just one batch triage away.
---
Every example in this post runs with a single API key — get yours at https://api.thistoken.ai/register and start in minutes.
Хотите попробовать Token.AI?
Создайте API Key уровня проекта, включите каналы в консоли и настройте маршрутизацию, бюджеты и журналы аудита.
注册 ThisToken.AI 并获取 API Key