steph france · agents in real companies

SKILLS

Knowledge compounding machines for your business.

why my agents keep getting better — and yours start from zero

the problem

Great work Tuesday.
Amnesia Wednesday.

Knowledge leaves in closed chats

Knowledge leaves with employees

Knowledge is rented from SaaS

every session starts from zero — the agent is smart, the company learns nothing

the mechanism

A skill is a file. Not a feature.

skills/platform-slack/ ├── SKILL.md ├── CONTEXT.md ├── LEARNINGS.md ├── LOGS.md ├── references/ ├── scripts/ └── tests/
plain Markdown, in a folder you own

the agent reads it when the task matches

open on camera — no black box

works with any model, survives every model
inside one real skill

Every part is just a file

  • SKILL.md — the procedure, read when it triggers
  • CONTEXT.md — local truth: identities, paths, boundaries
  • LEARNINGS.md — gotchas from real runs · where it compounds
  • LOGS.md — what ran, when, what happened
  • references/ scripts/ tests/ — depth, code, proof
LIVE — walk the real platform-slack folder, scroll the LEARNINGS gotcha list
why it compounds

The compounding loop

do real work
capture it as a skill
next run reads it
run writes lessons back

⟲ every loop makes the next run better

the model never got smarter — the workspace did

my library, ~100 skills

Six kinds of skills

core-*
how every agent thinks: plan, build, research, review
platform-*
one external tool each: Shopify, Slack, Klaviyo…
tool-*
production methods: diagrams, video, ads
wf-*
scheduled workflows that run themselves
cd-*
commands: /save, /resume, /process
system-*
internal capabilities, agents maintaining agents
structure

Skills point to skills

TOOLS.md → "read platform-klaviyo first"

SKILL.md → loads references/ only when needed

big skills branch into small ones

progressive disclosure:
the agent reads only what the task needs

context stays small,
knowledge stays deep
auto-improvement

Skills that improve skills

Skill Forge
creates and audits skills

Workflow Forge
turns skills into scheduled jobs

Feedback rule
every correction updates the skill, not the chat

it took 6 failed versions of one skill to learn this — and that history lives in the file

why not just…

Skills vs the usual answers

fine-tuningslow, expensive, frozen at training time
one giant promptcontext bloat, no structure, no history
RAG aloneretrieves facts, not procedures or judgment
skillsowned files: procedure + context + lessons, versioned in git
where good skills come from

The best skills don't come from Twitter.
They come from your company.

a downloaded skill knows the tool · your skill knows your business

LIVE — a skill born from one real conversation (founder process → skill file)
the warnings

External skills: last, not first

  • use external skills when your own foundation is ready
  • a skill is instructions your agent will obey — read before you install
  • never let a skill smuggle credentials or hidden calls
  • inspecting a source and obeying it are different acts
the carry-away

Your agent's intelligence lives in files you own — not in the model.

models will change again next year · your skill library survives all of them

start today: take the last useful thing your agent did — and make it a file.