Build Skills Loop

Outcome

Produce a production-ready skill and an iteration record, not just a draft prompt. Deliver clear artifacts for each stage so the workflow can be repeated.

Workflow

1) Define Business SOP from the user need

Capture these fields before writing any skill file:

• Goal and scope
• Trigger phrases (what user requests should activate the skill)
• Execution actions in SOP order
• Expected outputs and acceptance criteria
• One concrete input/output example

Write the result using references/dual-sop-template.md under the Business SOP section.

2) Discover existing skills with find-skills

Run targeted discovery before reinventing:

npx skills find <keywords>

If the user already has a candidate package, capture:

• Install command
• Package/repo and skill name
• Claimed popularity metrics (if provided by user)
• Attribution status (confirmed or unconfirmed)

For unknown sharer/author, mark attribution as unconfirmed and include a claim note in final output.

3) Compare candidate skill vs custom SOP

Build a gap matrix with these checks:

• Requirement coverage
• Missing technical steps
• Determinism (scripts/templates vs free text)
• Edge-case handling and failure recovery
• Testability and validation steps
• Context size risk (what should move to references/)

Use references/comparison-template.md.

4) Merge into one upgraded skill

Combine the stronger parts from both sides:

• Keep frontmatter trigger-focused (name, description)
• Keep SKILL body short and procedural
• Move long specs/checklists/templates into references/
• Add scripts only when repetitive steps need deterministic execution
• Add validation command(s)

5) Run two-round simulation and retrospective

Run at least two rounds:

1. Round 1: execute the merged skill on a realistic prompt.
2. Capture misses/friction.
3. Update skill/references/scripts.
4. Round 2: rerun the same prompt or a harder variant.
5. Record improvements and remaining gaps.

Store the loop in references/iteration-retro-template.md. If runtime memory is available, save the retrospective summary to memory; otherwise keep it in a local iteration log.

6) Iterate 2-3 cycles until stable

Stop only when outputs are stable across varied prompts:

• Triggering is reliable
• Required outputs are complete
• Failure handling is explicit
• Validation passes consistently

Output Contract

Return these artifacts in each build cycle:

1. Business SOP snapshot
2. Candidate skill discovery summary
3. Comparison matrix
4. Merged skill design decisions
5. Round-1 and Round-2 retrospective notes
6. Final changes applied to SKILL.md and resources

Example Request

Use this style to trigger the skill:

“帮我做一个用于周报自动生成的 skill。先按 SOP 写出执行流程和产出，再去找现有 skill，对比差异后合并。最后做两轮复盘并给出迭代版本。”

Attribution Note Template

Use this when source sharer is unknown:

“已确认技能包来源：vercel-labs/skills@find-skills。分享者身份暂未确认，欢迎原作者或传播者补充认领。”