krishamaze Skill Creator
Process
- Capture intent → write SKILL.md → test → eval → improve → repeat
- When satisfied: description optimization → package → return
Jump in at whatever stage the user is at.
1. Capture Intent
[DO] Extract from conversation first — tools used, steps taken, corrections made, observed I/O [DO] Fill gaps by asking: What enables the agent? When trigger? Expected output? Test cases needed?
<HARD-RULE: INTENT BEFORE STRUCTURE>
Iron Law: Never draw the skeleton until intent is confirmed by the user.
The skill must solve the user's actual problem — not what the agent assumes the problem is. Structure, pointers, and reference nodes are worthless if built for the wrong intent.
Mandatory questions — ask these, one at a time, wait for answers:
- What do you do today without this skill — and where does that break or feel wrong?
- What should change after the skill exists?
- What does "this skill is working" look like to you?
After answers received: [DO] Summarise intent back in 2–3 sentences. [DO] Ask: "Is this what you want the skill to do?" [DO] Wait for explicit user confirmation — "yes" or correction. Only after confirmation → proceed to SKELETON FIRST.
BANNED before intent is confirmed:
- Drawing any skeleton
- Creating any reference file
- Writing any pointer
- Assuming intent from domain knowledge alone
- Proceeding because "the request seems clear"
Self-check before touching any file: "Did the user explicitly confirm what this skill should do?" If no → STOP → ask the intent questions first.
Violation example: Agent understood "project memory" domain well. Built technically correct skill with ADR format, handoff modes, token budgets. User installed it. Intent was assumed, not confirmed. Skill solved the agent's interpretation — not the user's actual workflow gap. 368 lines of wrong direction.
</HARD-RULE: INTENT BEFORE STRUCTURE>
<HARD-RULE: SKELETON FIRST>
Iron Law: Never write content before the skeleton is confirmed.
Before writing any line of content in any .md file — output the full
SKILL.md skeleton first. Headers and pointers only. Zero content. Zero prose.
BANNED before skeleton is confirmed: Writing templates, formats, algorithms, examples, or any prose into SKILL.md or any reference file.
The skeleton looks like this:
## Section Name
[READ] references/file.md — WHEN <condition>
[DO] one line operational step
Self-check before writing any content: "Have I shown the skeleton and confirmed it with the user?" If no → STOP → output skeleton first.
Violation example: Agent wrote full ADR format, bash commands, and templates directly into SKILL.md — 368 lines — before any reference files existed. Map pattern was completely lost under content pressure.
</HARD-RULE: SKELETON FIRST>
2. Write SKILL.md
[READ] references/necessity-gate.md — BEFORE writing any line in any .md file [READ] references/pointer-syntax.md — WHEN writing any pointer or structuring the map [READ] references/skill-anatomy.md — WHEN deciding skill folder structure or progressive disclosure [READ] references/frontmatter.md — WHEN writing name, description, or compatibility fields
3. Test Cases
[READ] references/test-cases.md — WHEN writing eval prompts or assertions
4. Run and Evaluate
[READ] references/platform-notes.md — BEFORE running any evals [READ] references/eval-process.md — WHEN running test cases, grading, and reviewing results
5. Improve
[READ] references/improving.md — WHEN applying feedback and iterating on the skill
6. Blind Comparison
[SPAWN] agents/comparator.md — WHEN doing blind A/B comparison between two skill versions [SPAWN] agents/analyzer.md — WHEN analyzing why one version beat another Requires subagents.
7. Description Optimization
[READ] references/description-optimization.md — WHEN optimizing trigger description
Packaging
[RUN] scripts/package_skill.py <path/to/skill-folder>
[DO] Present the resulting .skill file to the user.