skill-forge-create

Name: skill-forge-create
Author: WrathZA

17stars0forks

How to install skill-forge-create?

wrathza/skillforge/skill-forge-create

$npx -y skills add WrathZA/skillforge --skill skill-forge-create

Installs into the current project.

Summary

Create a new SKILL.md via discovery recap, pattern selection, WebFetch spec fetch, knowledge delta discipline, skill-forge-judge + skill-forge-hitl quality gate. Triggers: create a skill, write a skill, new skill, SKILL.md, build a skill.

Files · 1

View on GitHub

SKILL.md

1	# Skill Forge Create
2
3	Write skills that score ≥B on skill-forge-judge out of the box. Every section must earn its tokens.
4
5	---
6
7	## Phase 1 — Discovery
8
9	Build understanding through a running recap. Ask questions to fill gaps; after each exchange, show the current recap:
10
11	```
12	## Discovery Recap
13
14	Domain: [what the skill covers, or "unknown"]
15	Decisions: [non-obvious choices the agent must make, or "unknown"]
16	Failure modes: [what breaks without this skill, or "unknown"]
17	Audience: [fragile/creative output, or "inferred: ..."]
18	Size: [reference files needed, or "inferred: self-contained"]
19	```
20
21	Then ask: `(a)ccept / (r)evise / (q)uit`
22
23	- `a` — recap is complete; proceed to Phase 2
24	- `r` — user adds or corrects; update recap and loop
25	- `q` — abort
26
27	Do not proceed to Phase 2 until the user accepts the recap. Domain, decisions, and failure modes must be filled before accepting — do not accept a recap with "unknown" in those three fields.
28
29	---
30
31	## Phase 2 — Pattern Selection
32
33	MANDATORY — READ [`references/pattern-selection.md`](references/pattern-selection.md) before selecting a pattern. Do NOT load this file during Phase 1, Phase 3, or later phases.
34
35	Select one pattern. State your choice and the one-line reason before drafting.
36
37	---
38
39	## Phase 3 — Draft
40
41	MANDATORY — WebFetch `https://agentskills.io/specification` before writing the description. Use the live spec to verify current frontmatter requirements and description field constraints. If unreachable, proceed with the spec knowledge already in context; note it in the draft.
42
43	Write the description before the body. Draft the body around what the description promises.
44
45	Write to earn tokens. For every paragraph, ask: "Does Claude already know this?"
46
47	- If yes → delete it
48	- If "it's a useful reminder" → one line max, then move on
49	- If no → expand it; this is the value
50
51	Before writing each section, ask: "What failure mode does this prevent?" If you can't answer, you haven't earned the paragraph.
52
53	Description requirements (THE most critical field):
54	- Answers WHAT (what does it do?)
55	- Answers WHEN (trigger scenarios — "Use when...", "Trigger phrases:")
56	- Contains searchable KEYWORDS (domain terms, file extensions, action verbs)
57	- Max ~200 chars for the core sentence; trigger phrases can follow
58
59	NEVER rules format — every NEVER must have:
60	```
61	- NEVER [specific construct/pattern]
62	Instead: [concrete alternative]
63	Why: [non-obvious failure mode this avoids]
64	```
65
66	Vague warnings ("be careful", "avoid errors") are prohibited.
67
68	Progressive disclosure thresholds:
69	- Body > 300 lines → move heavy content to `references/`
70	- Any section with a decision tree of 4+ branches → extract to `references/` regardless of total line count
71	- Add MANDATORY READ triggers at the exact workflow step that needs it
72	- Add "Do NOT load" guidance for files irrelevant to the current scenario
73
74	---
75
76	## Phase 4 — Self-Evaluate
77
78	Invoke `/skill-forge-judge` on the draft. (Requires skill-forge-judge to be installed — if missing, ask the user to install it before continuing.)
79
80	Target: ≥B (80%+). For a SKILL.md with no bash guidance → U + S (120 pts max) → ≥96 pts.
81
82	Always invoke `/skill-forge-hitl` on the numbered improvements list skill-forge-judge produced, regardless of grade. Do not apply fixes manually — that's /skill-forge-hitl's job.
83
84	If the /skill-forge-hitl loop stalls on a dimension (same item rejected after 3 revisions), surface it to the user: "Stuck on [dimension] — here's what I tried. Options: accept the current draft, revise the scope, or skip."
85
86	---
87
88	## NEVER
89
90	- NEVER write a section that restates Claude defaults ("write clean code", "handle errors", "be helpful")
91	Instead: Ask: "Would Claude do this without being told?" If yes, delete it.
92	Why: Default restatements dilute expert signal and train authors that padding is acceptable.
93
94	- NEVER add a NEVER rule without WHY and INSTEAD
95	Instead: Complete the three-part format before moving on.
96	Why: A prohibition without an alternative gets violated when the obvious path is blocked.
97
98	- NEVER dump all content in SKILL.md
99	Instead: Keep body under 300 lines; move detail to `references/` with MANDATORY READ triggers.
100	Why: A 700-line SKILL.md loads all at once on every invocation — the agent drowns in irrelevant content.
101
102	- NEVER skip Phase 4 (skill-forge-judge self-eval)
103	Instead: Run it even if the draft feels good.
104	Why: Skills that skip self-eval consistently have U1 or U3 gaps that aren't obvious to the author.
105
106	- NEVER manually apply skill-forge-judge findings one-by-one
107	Instead: Invoke `/skill-forge-hitl` on the numbered improvements list s

People who liked this also viewed

6 picks

Type

skill-forge-create

17stars0forks

View on GitHub

How to install skill-forge-create?

wrathza/skillforge/skill-forge-create

$npx -y skills add WrathZA/skillforge --skill skill-forge-create

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Skill Forge Create
2
3	Write skills that score ≥B on skill-forge-judge out of the box. Every section must earn its tokens.
4
5	---
6
7	## Phase 1 — Discovery
8
9	Build understanding through a running recap. Ask questions to fill gaps; after each exchange, show the current recap:
10
11	```
12	## Discovery Recap
13
14	Domain: [what the skill covers, or "unknown"]
15	Decisions: [non-obvious choices the agent must make, or "unknown"]
16	Failure modes: [what breaks without this skill, or "unknown"]
17	Audience: [fragile/creative output, or "inferred: ..."]
18	Size: [reference files needed, or "inferred: self-contained"]
19	```
20
21	Then ask: `(a)ccept / (r)evise / (q)uit`
22
23	- `a` — recap is complete; proceed to Phase 2
24	- `r` — user adds or corrects; update recap and loop
25	- `q` — abort
26
27	Do not proceed to Phase 2 until the user accepts the recap. Domain, decisions, and failure modes must be filled before accepting — do not accept a recap with "unknown" in those three fields.
28
29	---
30
31	## Phase 2 — Pattern Selection
32
33	MANDATORY — READ [`references/pattern-selection.md`](references/pattern-selection.md) before selecting a pattern. Do NOT load this file during Phase 1, Phase 3, or later phases.
34
35	Select one pattern. State your choice and the one-line reason before drafting.
36
37	---
38
39	## Phase 3 — Draft
40
41	MANDATORY — WebFetch `https://agentskills.io/specification` before writing the description. Use the live spec to verify current frontmatter requirements and description field constraints. If unreachable, proceed with the spec knowledge already in context; note it in the draft.
42
43	Write the description before the body. Draft the body around what the description promises.
44
45	Write to earn tokens. For every paragraph, ask: "Does Claude already know this?"
46
47	- If yes → delete it
48	- If "it's a useful reminder" → one line max, then move on
49	- If no → expand it; this is the value
50
51	Before writing each section, ask: "What failure mode does this prevent?" If you can't answer, you haven't earned the paragraph.
52
53	Description requirements (THE most critical field):
54	- Answers WHAT (what does it do?)
55	- Answers WHEN (trigger scenarios — "Use when...", "Trigger phrases:")
56	- Contains searchable KEYWORDS (domain terms, file extensions, action verbs)
57	- Max ~200 chars for the core sentence; trigger phrases can follow
58
59	NEVER rules format — every NEVER must have:
60	```
61	- NEVER [specific construct/pattern]
62	Instead: [concrete alternative]
63	Why: [non-obvious failure mode this avoids]
64	```
65
66	Vague warnings ("be careful", "avoid errors") are prohibited.
67
68	Progressive disclosure thresholds:
69	- Body > 300 lines → move heavy content to `references/`
70	- Any section with a decision tree of 4+ branches → extract to `references/` regardless of total line count
71	- Add MANDATORY READ triggers at the exact workflow step that needs it
72	- Add "Do NOT load" guidance for files irrelevant to the current scenario
73
74	---
75
76	## Phase 4 — Self-Evaluate
77
78	Invoke `/skill-forge-judge` on the draft. (Requires skill-forge-judge to be installed — if missing, ask the user to install it before continuing.)
79
80	Target: ≥B (80%+). For a SKILL.md with no bash guidance → U + S (120 pts max) → ≥96 pts.
81
82	Always invoke `/skill-forge-hitl` on the numbered improvements list skill-forge-judge produced, regardless of grade. Do not apply fixes manually — that's /skill-forge-hitl's job.
83
84	If the /skill-forge-hitl loop stalls on a dimension (same item rejected after 3 revisions), surface it to the user: "Stuck on [dimension] — here's what I tried. Options: accept the current draft, revise the scope, or skip."
85
86	---
87
88	## NEVER
89
90	- NEVER write a section that restates Claude defaults ("write clean code", "handle errors", "be helpful")
91	Instead: Ask: "Would Claude do this without being told?" If yes, delete it.
92	Why: Default restatements dilute expert signal and train authors that padding is acceptable.
93
94	- NEVER add a NEVER rule without WHY and INSTEAD
95	Instead: Complete the three-part format before moving on.
96	Why: A prohibition without an alternative gets violated when the obvious path is blocked.
97
98	- NEVER dump all content in SKILL.md
99	Instead: Keep body under 300 lines; move detail to `references/` with MANDATORY READ triggers.
100	Why: A 700-line SKILL.md loads all at once on every invocation — the agent drowns in irrelevant content.
101
102	- NEVER skip Phase 4 (skill-forge-judge self-eval)
103	Instead: Run it even if the draft feels good.
104	Why: Skills that skip self-eval consistently have U1 or U3 gaps that aren't obvious to the author.
105
106	- NEVER manually apply skill-forge-judge findings one-by-one
107	Instead: Invoke `/skill-forge-hitl` on the numbered improvements list s

People who liked this also viewed

6 picks

Type