The research skill lives at .agents/skills/iwac-mcp/ — it moved out of .claude/skills/, which is where this issue originally pointed. It is written for Claude: a depth-choice prompt, prose density tuned for Opus/Sonnet, and ~895 lines across SKILL.md plus four reference files. Driving the same workflow from Qwen-Agent or another local-LLM frontend needs an adapted variant.
Why
A Qwen 3.5 + Qwen-Agent chatbot would benefit from the same five-phase research methodology, but smaller models stumble on dense instructions and long context.
Status (July 2026)
Re-checked against the current skill — two of the original assumptions have drifted:
Strip <thinking> tags / XML-ish constructs — already gone. SKILL.md contains none.- The depth-choice prompt (
SKILL.md § Research Depth) is deliberate Claude-side behaviour and stays in the original; the fork is what drops it.
- The size problem is the real remaining work. 895 lines total:
SKILL.md 223, references/tools-by-phase.md 315, research-domains.md 178, biases-and-limitations.md 117, capabilities-overview.md 62. That is far past an ~8k-context budget.
Acceptance
Out of scope
- The chatbot itself (separate repo / project)
Related: #5 (zero-config semantic search via local embeddings). Together these are the same "run IWAC without cloud APIs" story — a local embedding model and a local LLM frontend.
The research skill lives at
.agents/skills/iwac-mcp/— it moved out of.claude/skills/, which is where this issue originally pointed. It is written for Claude: a depth-choice prompt, prose density tuned for Opus/Sonnet, and ~895 lines acrossSKILL.mdplus four reference files. Driving the same workflow from Qwen-Agent or another local-LLM frontend needs an adapted variant.Why
A Qwen 3.5 + Qwen-Agent chatbot would benefit from the same five-phase research methodology, but smaller models stumble on dense instructions and long context.
Status (July 2026)
Re-checked against the current skill — two of the original assumptions have drifted:
Strip— already gone.<thinking>tags / XML-ish constructsSKILL.mdcontains none.SKILL.md§ Research Depth) is deliberate Claude-side behaviour and stays in the original; the fork is what drops it.SKILL.md223,references/tools-by-phase.md315,research-domains.md178,biases-and-limitations.md117,capabilities-overview.md62. That is far past an ~8k-context budget.Acceptance
Out of scope
Related: #5 (zero-config semantic search via local embeddings). Together these are the same "run IWAC without cloud APIs" story — a local embedding model and a local LLM frontend.