Build workflow
Claude Code Skills: Our Agentic Engineering System
Ad-hoc prompting does not scale past one developer. Our answer is a skills system: every repeatable engineering task (publishing, audits, image pipelines, SEO checks, deploy safety) is captured as a versioned Claude Code skill with its own rules, scripts, and references. New projects inherit the whole library on day one, and the agent behaves like a team member who has already read the runbook.
Typical timeline
1-3 weeks depending on how many processes we capture as skills
Stack
Claude Code · Manas Skills (our open library) · Manas Framework (agent swarm + tracking layer) · n8n x Claude Code template · GitHub
What we need to start
- Your repository and deployment setup (any stack; we adapt the skill pack)
- The processes your team repeats weekly (we turn the top 5 into skills first)
- A safety baseline: what the agent may and may not touch
How it works
- 01
Process inventory
We list what your team actually repeats: releases, reviews, content publishing, data pulls, QA passes. Each candidate gets scored by frequency and rule-clarity; fuzzy judgment calls stay human.
- 02
Skill authoring
Each selected process becomes a SKILL.md with explicit triggers, steps, guardrails, and bundled scripts. Skills are code-reviewed like any other artifact.
Tools: Claude Code, Manas Skills - 03
Framework wiring
Skills plug into the Manas Framework: a tracking board, task locks, handoff notes, and an assumptions ledger so agent work is auditable and resumable across sessions.
Tools: Manas Framework - 04
Guardrails and review gates
Hooks enforce the rules that must never depend on model mood: commit gates, permission prompts, secret scanning, and human-approval checkpoints for anything outward-facing.
- 05
Team rollout
Your engineers get the skill library, the conventions doc, and a working session. The measure of success: a new hire (human or agent) ships a real task on day one by following skills, not tribal knowledge.
You get
- A versioned Claude Code skill library tailored to your stack
- Agent guardrails: hooks, permission rules, review gates
- The Manas Framework tracking layer (board, locks, ledger) installed in your repo
- A recorded working session + conventions doc for your team
When to choose another approach
- ·You want a one-off script, not a repeatable process (a skill is overkill)
- ·The process changes weekly; capture it after it stabilizes
- ·Nobody on your team will own the library; unowned automation rots
Open resources
Frequently asked
What is a Claude Code skill?
A versioned instruction pack (SKILL.md plus scripts and references) that teaches the agent one repeatable process: when to trigger, exact steps, guardrails, and how to verify the result. It turns prompting into engineering.
Do skills work with our existing stack?
Yes. Skills are markdown plus ordinary scripts, so they wrap whatever you already use: your CLI tools, CI, databases, and APIs. We adapt the library to your repo rather than forcing a new stack.
How is this different from a prompt library?
Prompts suggest; skills execute. A skill carries runnable scripts, hard guardrails enforced by hooks, and verification steps, so output quality does not depend on who wrote the prompt that day.
Is this an official Anthropic product or partnership?
No. Claude Code is Anthropic's tool; the skills system, Manas Framework, and templates are our own engineering practice built on top of it, and the library is open source on GitHub.
FROM A PATTERN TO YOUR SYSTEM
Want this working in your business?
Bring your tools, data, and operating requirements. ManasFlow can help you scope the implementation around the way your business works.