tkersey/ms

ms

Overview

Create and update Codex skills with minimal diffs. Use this as the direct skill-surgery primitive: work directly in the skill folder (no external registry) and keep the skill lean.

Hard constraints

• Minimal diff: change only what is needed to satisfy the new requirement.
• No extra docs: do not add README/INSTALL/CHANGELOG-style files.
• Frontmatter:
  • Default: only name and description.
  • Preserve existing allowed keys (e.g., metadata, license, allowed-tools) when updating system skills.
  • name must be hyphen-case (<=64 chars) and match the folder name.
  • description is the trigger surface: include "when to use" cues; no angle brackets; <=1024 chars.
• SKILL.md body:
  • Imperative voice.
  • Keep under 500 lines; spill deep details into references/.
  • Do not put trigger guidance in the body; put it in frontmatter description.
• agents/openai.yaml (if present/needed):
  • Prefer generating via generate_openai_yaml.py.
  • short_description must be 25-64 chars.
  • default_prompt must be one short sentence and mention $skill-name.
• Proof of completion (required):
  • Record agents/openai.yaml disposition as one of: regenerated, verified unchanged, or not present.
  • Include the exact quick_validate.py command used and its pass/fail result in the final summary.
• Always run quick_validate.py before calling it done.
  • Recommended (no global installs):
    • uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/quick_validate.py <path/to/skill>
  • Note: skill-creator scripts require PyYAML (import yaml).

Workflow Decision Tree

• If a matching skill already exists: run Update Workflow.
• If no matching skill exists: run Create Workflow.
• If the turn is read-only, asks whether a skill should change, or only asks for evidence/proof checklists: do not use ms yet.
• If the user already invoked $refine for an existing skill: let $refine own the refinement turn and treat $ms as implicit/redundant.
• If the name, location, or triggers are unclear: ask 1-3 targeted questions, then proceed.
• If refining ms itself: run the Seq Feedback Loop first and use that evidence to choose the smallest update set.

Create Workflow

1. De-duplicate: search for an existing skill that already covers the intent; prefer updating over creating a near-duplicate.
2. Discover and define: collect 2-3 concrete user prompts, then write:
   • Problem statement (1 line)
   • Success criteria (how we'll know it works)
3. Plan reusable assets: decide whether scripts/, references/, or assets/ are required; create only what is necessary.
4. Initialize (scaffold):
   • uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/init_skill.py <skill-name> --path codex/skills
   • Add --resources scripts,references,assets only when needed.
   • If you need UI metadata now, pass --interface key=value (repeatable).
5. Author SKILL.md:
   • Update frontmatter description to include concrete triggers (file types, tools, tasks, key phrases).
   • Keep the body procedural and composable (decision tree + steps + pointers into references/).
6. Sync UI metadata (optional but recommended for new skills):
   • Generate or regenerate: uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/generate_openai_yaml.py codex/skills/<skill-name> --interface key=value
   • If the skill already had default_prompt, include it again in overrides (generator does not infer it).
   • For field constraints, read codex/skills/.system/skill-creator/references/openai_yaml.md.
7. Validate:
   • uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/quick_validate.py codex/skills/<skill-name>
8. Iterate with the user: tighten triggers, remove redundancy, and promote repeatable code into scripts/.

Update Workflow (In Place)

1. Locate the target skill folder in codex/skills (or codex/skills/.system for system skills).
2. Read the current SKILL.md and resources to identify the smallest set of required changes.
3. Edit in place:
   • Update frontmatter description if triggers changed.
   • Adjust workflows, tasks, or references with minimal diffs (no formatting churn).
   • Add or remove resource folders only when they create real reuse.
4. Sync UI metadata:
   • If agents/openai.yaml exists, keep it consistent with the skill and regenerate if stale.
   • If missing and UI metadata is desired, create it with uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/generate_openai_yaml.py <path/to/skill> --interface key=value.
5. Validate with uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/quick_validate.py <path/to/skill>.
6. Summarize changes and next steps.
   • Include openai_yaml: regenerated|verified unchanged|not present.
   • Include quick_validate: <exact command> -> <result>.

Seq Feedback Loop (for improving ms)

When the target skill is ms, mine recent sessions before editing:

1. Collect assistant outputs that mention $ms:
   • seq query --root ~/.codex/sessions --spec '{"dataset":"messages","where":[{"field":"role","op":"eq","value":"assistant"},{"field":"text","op":"contains","value":"$ms"}],"select":["path","timestamp","text"],"sort":["timestamp"],"format":"jsonl"}'
2. Compute adherence rates for required proof markers:
   • quick_validate.py
   • uv run --with pyyaml -- python3 codex/skills/.system/skill-creator/scripts/quick_validate.py
   • agents/openai.yaml
   • generate_openai_yaml.py
3. Promote repeated misses into minimal-incision SKILL.md constraints/workflow edits.

Trigger Examples

• "Create a skill to manage OpenAPI specs and generate SDKs."
• "Update skill X to include a new workflow and regenerate agents/openai.yaml."
• "Refactor this skill's SKILL.md and add references/ for schemas."
• "Refactor this skill's SKILL.md, move deep details into references/, and keep it under 500 lines."