We Built a Pipeline That Turns Individual AI Know-How Into Org-Wide Assets

A few months ago, a GitHub project called "colleague.skill" went viral: feed a departing employee's data into AI, get a digital clone. The debate it triggered made us realize something simpler — every company with 50+ AI users is hemorrhaging invisible assets every time someone leaves. Here's the pipeline we built to stop it.

The problem nobody was tracking

Traditional knowledge management captures files. Specs go in the drive, code in the repo. When someone leaves, access gets revoked and the artifacts stay.

AI output is different. What makes a senior engineer fast with AI isn't what they produce — it's how they ask. The same task, the same model: one person ships production code, another gets unusable output. That gap is invisible and it walks out the door.

We found three categories of "blind-spot assets" hiding in personal AI accounts:

| Asset Type | What It Is | Why It Disappears |
|---|---|---|
| Session Patterns | Proven context combos + guidance strategies | Buried in personal chat history |
| Skill Chains | Branching workflows refined over 6+ months | Locked in platform config UIs |
| Memory Snippets | Org-specific convention knowledge fed into prompts | Lives only in seniors' brains |

❌ None of these live in Confluence, Notion, or the company wiki.
❌ Newcomers don't know they exist, so they can't ask for them.
❌ After a departure, they're gone. Forever.

The five-stage pipeline

1. Converge the ingress

If API calls scatter across personal accounts, there's nothing to distill. We deploy a sidecar proxy at the edge — zero business code changes. Every request gets identity tagged, department attached. Call path shifts from individual → model platform to individual → org proxy → model platform.

2. Capture at runtime

The proxy forwards requests AND captures context on a side channel — session structures, skill paths, memory traces. Collected as structured data in-flight, not retroactive logs. Suddenly you can answer: whose prompts are highest quality, where is the budget actually going, which call chains are worth reusing?

3. Refine, don't hoard

Raw data hits a refinement module. The principle: don't save everything, sift for gold. Auto-strip PII and sensitive fields, discard noise, keep high-value patterns. Retention policy is admin-configurable — selective by default, deeper for key roles.

4. Catalog with access controls

Refined assets enter a multi-axis catalog — searchable by capability type, source department, scenario. But searchable ≠ open access. Cross-department reuse requires admin approval. What gets preserved is "how it's done," not "who said what."

5. Distill into reusable policies

High-frequency sessions and skill chains get solidified as preset policies. Newcomers don't start from zero — there's a verified path in the call workflow. Memory snippets auto-match in similar scenarios. From "ask Bob for his prompts" to "the policy library has it."

What we learned

Each stage depends on the previous one. Skip ingress convergence, distillation is impossible. Skip refinement, you're distilling noise. Skip access controls, sharing becomes chaos.

The hard truth: every month you delay, another batch of assets becomes permanently unrecoverable after the next resignation. The pipeline isn't about making AI smarter — it's about making sure the money you're already spending on AI calls doesn't just produce this month's output, but leaves behind capability that compounds.

If you're building something in this space or thinking about it, would love to hear your take.
Learn more at https://aikeylabs.com/zh/i/ih10
For enterprise inquiries: [email protected]