Skill Synthesis

Skill Synthesis

The third time you do the same thing, Ptah notices — and for a single repeated workflow, it doesn’t wait for you to say yes.

Skill Synthesis watches your sessions for trajectories — sequences of turns, tool calls, and outcomes that look like a reusable workflow — and turns the ones that hold up into permanent skills at:

~/.ptah/skills/<slug>/SKILL.md

From that point on, any agent can invoke it like a hand-authored skill — same trigger semantics, same on-the-fly context injection.

Two tracks, and only one of them asks first

A single workflow that succeeds enough times promotes itself. No review screen, no Accept button — it clears a quality judge and a couple of safety gates, then Ptah writes the SKILL.md and it’s live in your library. This is the intended design: the gates are the safety mechanism, not a person.

A cluster of similar-but-not-identical sessions is different. Ptah distills the cluster into one candidate, judges it, and stops — it only reaches your library after you review it in Recommended and click Accept.

If you’ve only ever noticed the Recommended queue, it’s worth knowing the first path exists. How It Works covers the full split; Background Learning covers what runs unattended and how to control it.

What’s in this section

The Skills tab

Recommended, Sessions & Library — and what to delete. Learn more →

How it works

Capture, the two promotion tracks, and the gates between them. Learn more →

Background learning

The queue, the cron tiers, and why it doesn’t burn your chat quota. Learn more →

Reviewing candidates

Promote, reject, or just watch what Ptah found. Learn more →

SKILL.md anatomy

What the generated file looks like — and what it deliberately leaves out. Learn more →

Settings

Thresholds and caps. Learn more →

Why it exists

Hand-authored skills are great when you know up-front what’s worth abstracting. Most workflows aren’t like that — they emerge from repetition. Skill Synthesis catches those without you having to notice them yourself.

Limits

Active skills are capped at 200 by default (skillSynthesis.maxActiveSkills). When the cap is exceeded, the weakest resident is demoted to dormant — kept on disk and in the database but skipped when skills load. Dormant skills are never deleted, and authored skills are exempt. Cosine dedup against the active set prevents near-duplicates from polluting the directory.