hk-skill-creator

Name: hk-skill-creator
Author: deepklarity

84stars12forks

How to install hk-skill-creator?

deepklarity/harness-kit/hk-skill-creator

$npx -y skills add deepklarity/harness-kit --skill hk-skill-creator

Installs into the current project.

Summary

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's descr

Files · 1

View on GitHub

SKILL.md

1	# /hk-skill-creator — Create and Improve Skills
2
3	Create new skills and iteratively improve existing ones through a structured draft-test-evaluate-improve loop.
4
5	## Context
6
7	<skill_context> $ARGUMENTS </skill_context>
8
9	If the context above describes what skill to create or improve, proceed. If empty or unclear, ask the user what they want the skill to do.
10
11	## The Core Loop
12
13	The process of creating a skill:
14
15	1. Understand what the skill should do, with concrete examples
16	2. Draft the SKILL.md and any bundled resources
17	3. Test by running Claude-with-the-skill on realistic prompts
18	4. Evaluate outputs qualitatively (does it do what the user wants?) and quantitatively (assertions)
19	5. Improve based on feedback — generalize, don't overfit
20	6. Repeat until the user is satisfied
21	7. Optimize the description for reliable triggering
22
23	Figure out where the user is in this process and help them progress. Maybe they want a skill from scratch — help narrow intent, draft, test. Maybe they already have a draft — go straight to eval/iterate. Maybe they just want to vibe — be flexible.
24
25	## Communication Style
26
27	Skills are used by people across a wide range of technical familiarity. Pay attention to context cues:
28
29	- "evaluation" and "benchmark" are borderline but OK to use without explanation
30	- For "JSON", "assertion", "frontmatter" — look for cues the user knows these before using them without a brief definition
31	- When in doubt, briefly explain terms inline
32	- Match the user's register — if they're casual, be casual back
33
34	## Step 1: Capture Intent
35
36	Start by understanding what the skill should enable. The conversation might already contain a workflow to capture (e.g., "turn this into a skill"). If so, extract answers from context first — tools used, sequence of steps, corrections made, input/output formats. The user fills gaps, then confirms before proceeding.
37
38	Questions to resolve:
39	- What should this skill enable Claude to do?
40	- When should this skill trigger? (what user phrases/contexts)
41	- What's the expected output format or behavior?
42	- Are there concrete examples of input/output?
43
44	### Interview and Research
45
46	Proactively ask about edge cases, input/output formats, example files, success criteria, and dependencies. Check existing skills in `.claude/skills/` — spawn a haiku subagent to scan for overlapping or related skills. Come prepared with context to reduce burden on the user.
47
48	Do not proceed to drafting until the functionality is clearly understood.
49
50	## Step 2: Plan Resources
51
52	For each concrete example from Step 1, analyze:
53	1. How to execute from scratch (what steps, what tools)
54	2. What scripts, references, or assets would help when executing repeatedly
55
56	Build a list of reusable resources to include. Common signals:
57	- If all test cases would independently write similar helper scripts, bundle that script in `scripts/`
58	- If the skill needs detailed reference docs (API schemas, large examples), put them in `references/`
59	- If the skill produces output from templates, put templates in `assets/`
60
61	## Step 3: Initialize or Locate the Skill
62
63	### New skill
64
65	Scaffold with:
66
67	```bash
68	python ${CLAUDE_SKILL_DIR}/scripts/init_skill.py <skill-name>
69	```
70
71	This creates the directory structure at `.claude/skills/<skill-name>/` with a template SKILL.md. Follow `hk-` prefix convention for this project.
72
73	### Existing skill
74
75	Read the existing SKILL.md and understand its current structure before making changes.
76
77	## Step 4: Write the Skill
78
79	The skill is written for another instance of Claude to use. Focus on non-obvious procedural knowledge — things the model wouldn't know or would get wrong without guidance.
80
81	### Implementation order
82
83	1. Bundled resources first (scripts/, references/, assets/)
84	2. Delete unused example files from initialization
85	3. Complete SKILL.md last (it references the resources)
86
87	### SKILL.md Structure
88
89	Read `references/writing_guide.md` for the complete writing guide including all frontmatter fields, dynamic context injection, and subagent execution patterns. Key points:
90
91	Frontmatter — The `description` is the primary triggering mechanism. Descriptions longer than 250 characters get truncated in the skill listing, so front-load the key use case. Include both what the skill does AND specific trigger

People who liked this also viewed

6 picks

Type

hk-skill-creator

84stars12forks

View on GitHub

How to install hk-skill-creator?

deepklarity/harness-kit/hk-skill-creator

$npx -y skills add deepklarity/harness-kit --skill hk-skill-creator

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# /hk-skill-creator — Create and Improve Skills
2
3	Create new skills and iteratively improve existing ones through a structured draft-test-evaluate-improve loop.
4
5	## Context
6
7	<skill_context> $ARGUMENTS </skill_context>
8
9	If the context above describes what skill to create or improve, proceed. If empty or unclear, ask the user what they want the skill to do.
10
11	## The Core Loop
12
13	The process of creating a skill:
14
15	1. Understand what the skill should do, with concrete examples
16	2. Draft the SKILL.md and any bundled resources
17	3. Test by running Claude-with-the-skill on realistic prompts
18	4. Evaluate outputs qualitatively (does it do what the user wants?) and quantitatively (assertions)
19	5. Improve based on feedback — generalize, don't overfit
20	6. Repeat until the user is satisfied
21	7. Optimize the description for reliable triggering
22
23	Figure out where the user is in this process and help them progress. Maybe they want a skill from scratch — help narrow intent, draft, test. Maybe they already have a draft — go straight to eval/iterate. Maybe they just want to vibe — be flexible.
24
25	## Communication Style
26
27	Skills are used by people across a wide range of technical familiarity. Pay attention to context cues:
28
29	- "evaluation" and "benchmark" are borderline but OK to use without explanation
30	- For "JSON", "assertion", "frontmatter" — look for cues the user knows these before using them without a brief definition
31	- When in doubt, briefly explain terms inline
32	- Match the user's register — if they're casual, be casual back
33
34	## Step 1: Capture Intent
35
36	Start by understanding what the skill should enable. The conversation might already contain a workflow to capture (e.g., "turn this into a skill"). If so, extract answers from context first — tools used, sequence of steps, corrections made, input/output formats. The user fills gaps, then confirms before proceeding.
37
38	Questions to resolve:
39	- What should this skill enable Claude to do?
40	- When should this skill trigger? (what user phrases/contexts)
41	- What's the expected output format or behavior?
42	- Are there concrete examples of input/output?
43
44	### Interview and Research
45
46	Proactively ask about edge cases, input/output formats, example files, success criteria, and dependencies. Check existing skills in `.claude/skills/` — spawn a haiku subagent to scan for overlapping or related skills. Come prepared with context to reduce burden on the user.
47
48	Do not proceed to drafting until the functionality is clearly understood.
49
50	## Step 2: Plan Resources
51
52	For each concrete example from Step 1, analyze:
53	1. How to execute from scratch (what steps, what tools)
54	2. What scripts, references, or assets would help when executing repeatedly
55
56	Build a list of reusable resources to include. Common signals:
57	- If all test cases would independently write similar helper scripts, bundle that script in `scripts/`
58	- If the skill needs detailed reference docs (API schemas, large examples), put them in `references/`
59	- If the skill produces output from templates, put templates in `assets/`
60
61	## Step 3: Initialize or Locate the Skill
62
63	### New skill
64
65	Scaffold with:
66
67	```bash
68	python ${CLAUDE_SKILL_DIR}/scripts/init_skill.py <skill-name>
69	```
70
71	This creates the directory structure at `.claude/skills/<skill-name>/` with a template SKILL.md. Follow `hk-` prefix convention for this project.
72
73	### Existing skill
74
75	Read the existing SKILL.md and understand its current structure before making changes.
76
77	## Step 4: Write the Skill
78
79	The skill is written for another instance of Claude to use. Focus on non-obvious procedural knowledge — things the model wouldn't know or would get wrong without guidance.
80
81	### Implementation order
82
83	1. Bundled resources first (scripts/, references/, assets/)
84	2. Delete unused example files from initialization
85	3. Complete SKILL.md last (it references the resources)
86
87	### SKILL.md Structure
88
89	Read `references/writing_guide.md` for the complete writing guide including all frontmatter fields, dynamic context injection, and subagent execution patterns. Key points:
90
91	Frontmatter — The `description` is the primary triggering mechanism. Descriptions longer than 250 characters get truncated in the skill listing, so front-load the key use case. Include both what the skill does AND specific trigger

People who liked this also viewed

6 picks

Type