Thin Harness / Fat Skills
Interfaces change. Only context layer stays.
AI interfaces change fast. Only people and culture stay. Thin Harness is the generic LLM glue — 200 lines is enough. Fat Skills is the context layer on top, carrying the org's tacit knowledge. 90% of value lives in Skills. Software value has collapsed to zero. Durable revenue comes from open-source infra (IronClaw-style security) or people infra (Palantir-style FDE). AX isn't a product — it's the people and culture that accumulate Skills.
- 01Interfaces change fast. People and culture don't
- 02Harness = 200-line glue: JSON in, text out
- 0390% of value is in Skills
- 04Software value has collapsed to zero
- 05Moving tacit org knowledge into Skills doesn't automate away
- 06Cognition's $10.2B is the FDE playbook, not Devin
- 07Harnesses replicate. Skills are org-specific
- 08Don't chase interfaces. Bet on people and culture that stack Skills
A generic LLM interface abstraction. JSON in, text out — thin enough to write in ~200 lines.
The org-specific context layer on top of the harness. Domain knowledge, procedures, and judgment in markdown — where 90% of value lives.
Palantir's model — engineers embedded in client orgs to convert tacit knowledge into Skills. Cognition applies it to AI.
Not a product. The philosophy of helping orgs grow their own Skills — through people and culture.
“Interfaces change. People and culture are what stays”
— Goobong Jeong
“Software replicates. Skills are org-specific”
— Goobong Jeong
“Turning tacit org knowledge into Skills doesn't automate”
— Goobong Jeong