algo-first-skill-sec

"The LLM reflects — the user evaluates." — roebi

"I ask for algorithms, I get algorithms." — a friend

---

The Architecture Pattern

/**
 * Architecture Pattern: Algo First, Skill Second
 * License: MIT
 */

Function Solution findPayableSolution(String problem) {

  List<Subproblem> subproblems = divideAndConquer(problem)

  List<Solution> solutions = []

  for each subproblem in subproblems {

    List<Algorithm> algos = findAlgorithms(subproblem)
    Subproblem remainder  = subproblem - algos

    List<Skill> skills = isEmpty(remainder)
                           ? []
                           : findSkills(remainder)

    solutions.add(
      Solution(
        controller : algos,   // runs first, always
        llmCalls   : skills   // only for what algos cannot do
      )
    )
  }

  return merge(solutions)
}

// Entry point
Solution s = findPayableSolution("my problem")

Key properties:

• remainder = subproblem - algos — the Skill's input is always what the algo left over, never the raw problem
• isEmpty(remainder) ? [] : findSkills(remainder) — zero Skills is a valid and preferred outcome
• controller: algos — the algo is always in charge; the Skill is a parameter it passes work to

A Skill is not losing to an algorithm. A Skill is a more complex algorithm — reserved for what only an LLM can do.

---

What is in this repo

Directory	Content	License
skills/	Agent Skills (agentskills.io spec)	CC BY-NC-SA 4.0
scripts/	Extracted algorithms (bash)	MIT

This repo proves its own methodology by applying it to itself. Every script here was extracted from a Skill using extract-algorithm-from-skill.

---

Skills

find-algorithm-for-problem

Design-time gate. Before creating a Skill, run this process to determine whether a known algorithm or script can solve the problem (or most of it). The algorithm becomes the controller; any LLM-only remainder becomes a minimal, targeted Skill.

extract-algorithm-from-skill

Refactor-time gate. Applied to existing Skills. Classifies every instruction step as ALGORITHMIC or LLM-ONLY. Extracts deterministic steps into scripts. Stamps a Quality Mark (proved) on the reviewed Skill and logs the action.

---

The Quality Mark

Every Skill reviewed by extract-algorithm-from-skill receives a proved stamp in its metadata:

metadata:
  proved: 'YYYYMMDD | <skill-name>:<skill-version> | outcome:<outcome> | who:<reviewer>'

Absence of proved = unreviewed.

---

License structure

Every top-level subdirectory contains its own LICENSE file. The root LICENSE file does not govern content — it only explains the structure and links to each subfolder license.

algo-first-skill-sec/
├── LICENSE          ← explainer + links (governs nothing)
├── skills/LICENSE   ← CC BY-NC-SA 4.0
└── scripts/LICENSE  ← MIT