Skills

Skills are capability definitions that enable the FDE to execute specific business tasks. Unlike conventional AI tools where users design and implement skills, Deprex’s FDE generates skills autonomously.

Skills Are Born from Observation

The FDE’s skill generation pipeline works as follows:

Workflow observation -> Pattern detection -> Task flow recording -> Skill auto-generation -> Shadow Comparison -> PM approval

1. Workflow Observation

The FDE continuously observes Slack conversations, meeting transcripts, and recurring task patterns.

Observation example:
- The PM manually compiles the support backlog every Monday
- The same format is used for the Slack post every time
- The KPI definitions used for compilation are fixed

2. Pattern Detection and Skill Generation

When recurring patterns are detected, the FDE records the task flow and auto-generates a skill (SKILL.md + scripts).

app/orgs/acme/skills/
  weekly-support-summary/      # Auto-generated by the FDE
    SKILL.md                   # Business procedure definition
    scripts/
      collect_metrics.py       # Data collection script

3. Proposal to PM

When a skill is generated, the PM is notified via Slack.

[Skill Generated] weekly-support-summary
Purpose: Automate the weekly Monday support backlog compilation
Inputs: Support tool data, KPI definitions
Output: Summary report posted to Slack
Shadow Comparison: Compared against 3 weeks of manual output, quality equivalent
Approve?

SKILL.md Structure

Auto-generated skills follow this structure:

---
name: weekly-support-summary
description: Auto-execute Acme's weekly support backlog compilation
---

# Weekly Support Summary

## Purpose
Compile the weekly support backlog, classify major themes, and post to Slack.

## Inputs
- Support backlog data
- KPI definitions memory
- Reporting rules memory

## Workflow
1. Retrieve data from the support backlog source
2. Classify inquiries into the top 3 categories
3. Extract metrics with significant week-over-week changes
4. Generate a Slack-ready summary

## Output
- Summary in natural language posted to Slack
- List of items requiring review

## Rules
- Do not state speculation as fact
- Mark items with unclear data sources as "needs verification"
- Escalate to PM if backlog exceeds 50 items

PMs review this content for business accuracy. Code-level review is not needed.

Shared and Tenant Skills

Type	Location	How Created	Examples
Shared	app/_shared/skills/	Developed by 0ai engineers	Browser automation, image generation, data analysis
Tenant	app/orgs/<company>/skills/	Auto-generated by the FDE	Tenant-specific reports, notifications, analysis

The FDE combines shared skill capabilities with tenant-specific context to execute tasks.

Autonomous Skill Improvement

Skills are not static after generation. The FDE continuously improves them through:

Learning from Feedback

When PMs provide feedback on output, the FDE automatically modifies the skill’s Rules or Workflow.

PM: "Include category-level counts in the support summary"
FDE: Modified Workflow step 4 to include category-level counts.
     This will be reflected starting from the next execution.

Self-Improvement from Execution Results

The FDE automatically evaluates its own output quality, detects improvement opportunities, and modifies skills. Quality gradually improves even without human feedback.

Ground Truth Comparison

Shadow Comparison periodically compares human output (ground truth) with FDE output. When quality gaps exist, the FDE modifies skills to converge quality.

The PM’s Role

PMs are not in the position of creating skills, but of reviewing skills proposed by the FDE:

• Approve: If the skill’s purpose and output match business needs
• Request changes: Point out needed modifications to output format or decision criteria via Slack
• Reject: If the task should not be automated
• Provide feedback: Share improvement points based on execution results

Important Notes

• Credentials are never written directly into skills (managed in config)
• PMs do not need to write skill code
• Skill improvement is performed autonomously by the FDE, but changes to business decision criteria require PM feedback