Skills are the procedural memory of the company. This essay covers the SOUL.md / SKILL.md / run.sh contract that every skill follows, why bash is the right language for the catalog, and the atomic capability loop that grows the catalog autonomously week after week.
A 340-skill catalog is good. A growing catalog is the moat. The discipline that makes growth possible is the structural shape of each skill — readable in 60 seconds, smoke-tested before promotion, owner-tagged for auto-wire. Read this if you have a working agent and want to know why your catalog is not compounding.
The 4 pillars are the chassis. The 340 skills are the engine.
A skill is an executable directory containing SKILL.md (frontmatter plus a description) and run.sh (executable bash). The frontmatter declares 4 fields — the skill's name, owner_agent, description, and trigger; the bash body is whatever it is.
Skills are deliberately bash. Not Python. Not TypeScript. Not a custom DSL. The reasoning is simple: every skill should be readable in 60 seconds. Bash plus jq plus curl is the lowest-common-denominator way to express "do these 3 things in order"; an LLM can read a 30-line skill and know what it does immediately.
About 340 skills currently. The number isn't the point — the generation rate is. The atomic capability loop runs weekly:
1. Watch what tasks the system gets asked to do.
2. Identify the recurring ones (3+ instances over 30 days).
3. Propose a new skill for each recurrence pattern.
4. A/B test the proposed skill against the existing path (whatever the human or agent did before the skill existed).
5. Promote the winners (skills whose A/B variant produced higher quality output or lower cost).
6. Retire the losers (skills that didn't beat the baseline after 10 trials).
The loop is what makes the catalog compound. Year 1's 340 skills are good. Year 2's catalog will be measurably better, because the system creates capabilities autonomously — and the skills get used, audited, and refined faster than any human team could iterate.
The 340 skills break into 3 categories:
Operational skills. Things the system needs to do to keep itself running: rotate credentials, recover stalled launchd plists, audit browser sessions, generate Friday ship logs, maintain the vault index. The DAI OS public bundle ships with 10 of these because they generalize across every customer.
Domain skills. Things specific to the company's market or workload: draft a Gumroad product page, score a sales lead, calibrate a recommendation system, run 1 competitor-analysis pass. These are private; they encode specific business knowledge.
Meta-skills. The 4 skills that operate on other skills: the skill creator, the skill A/B tester, the skill registry sync, the skill catalog index. Without meta-skills the catalog would be a flat list; with them it's a self-organizing library.
The skill is procedural memory. Where the vault remembers facts (Pillar 3), the skill catalog remembers how to do things. Both compound over time; both are durable when the underlying LLM models change.
**Chapter 3 summary:** Skills are bash directories with SKILL.md plus run.sh. About 340 skills today, growing weekly via the atomic capability loop. Three categories: operational, domain, meta. Skills compound; they are the procedural memory of the company.
← The 4-pillar architecture · full paper · Voice-gating and public-content scrubbing →