Stackwich — Development skill for Claude Code
Installs an operating architecture into Claude Code as five layers — prompt, context, harness, loop, graph: meta-prompted delegation, forked research, a single-writer harness, capped failure loops, an.
How to install Stackwich
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open ovhirup/stackwich and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Stackwich does
Installs an operating architecture into Claude Code as five layers — prompt, context, harness, loop, graph: meta-prompted delegation, forked research, a single-writer harness, capped failure loops, and a plan→execute→verify→review sandwich with three tiered subagents.
Alternatives in Development
- LifeOS — ⛰️ The Life Operating System — an intent engineering platform that moves you from your current state to your i 18.7k ★
- Portaljs Add Chart — Add a chart (line, bar, area, pie, or scatter) to a dataset's showcase in a PortalJS portal 2.3k ★
- Fable OS — An agentic operating system where the kernel is controlled directly by Claude 324 ★
README
🥪 stackwich
[](https://github.com/ovhirup/stackwich/actions/workflows/validate.yml) [](LICENSE) [](https://docs.claude.com/en/docs/claude-code/skills)
**Delivery discipline for AI agents. Five layers, one command.**
Most people run AI coding agents like slot machines: prompt → pull the lever → hope. The agent writes code 10× faster than you review it, declares itself done, and you find out later. We would never accept that from a junior engineer — we accept it from our tools every day.
Stackwich is 15 years of delivery coaching — plan → build → test → review, the oldest idea in software — compressed into standing policy for [Claude Code](https://docs.claude.com/en/docs/claude-code). Install it once and every future session works this way.

The five layers
| Layer | Habit |
|---|---|
| Prompt | Meta-prompt every delegation. Context-full, with a stated done condition. Understanding never gets delegated. |
| Context | Fork research and read-heavy scans off the main thread. Memory, plans, and tasks each do their own job. |
| Harness | One writer. Read-only enforced in practice, not just in the tool list. Nothing "done" until observed. |
| Loop | Recurring work runs on /loop. Failure loops are capped at 2 cycles and never silent. |
| Graph | advisor plans → executor implements → verifier gates → advisor reviews, with explicit hand-off contracts. |
Cost discipline isn't a sixth layer — it's how the five are practiced: cheap models grind, expensive models only plan and review, and nothing already established gets re-derived.
The layer names are structure for the
Related Skills
Subagent Orchestration
Use subagents for bounded delegation, not for vague “go think more” loops.
Analyzing Systemic Governance Loops 20260822
Diagnosing functional failure by interrogating underlying structural mechanics rather than observed behavior o
Observe
Render the 7-section ATLAS observability dashboard (tool health, safety hooks, skill usage, scheduled tasks, a
Draft Agent
Meta-agent agent-builder. From a one-line description, generates a new subagent definition OR a mega-goal sub-
SEO Blog Writer Claude
SEO blog writer Claude skill — turn any URL, notes, or topic into a human-sounding blog post that ranks on Goo
Skills Marketing
Marketing skills for Claude Code, Codex, Cursor — funnel analysis, CRO, growth loops, paid ads (Google/Meta),
Related Agents
Copyeditor
Sole owner of deterministic quantitative checks across all Layers. Runs the 10 tools/lint.py groups (graph·hub
Agent Tuner
Reflexion-style agent tuner. Given evidence of a recurring failure by a specific agent, proposes the smallest
Repo Assessor
Assesses a repository against the five harness dimensions and produces one page a person can act on. Use when