Workshop a skill draft against the authoring standard: description-as-trigger, one workflow
per skill, guardrails matched to the blast radius, and under the character limit. Skills are
in-app Super Magic only — where skill editing isn't available, produce the finished text for
the member to paste into the editor.
1. Load the draft (open it if it exists, otherwise take the pasted text) and check the
existing skills for one that already covers the workflow — improving one beats
duplicating it.
2. Workshop against the standard, in this order:
- Description = trigger. Rewrite it as WHEN TO REACH FOR THIS, in the words the member
would type: "if I typed <phrase>, would this match?" A description that summarizes
instead of triggering is the number one authoring failure.
- One workflow. If the draft does three unrelated things, split it and say why.
- The prompt. Imperative instructions in plain English — describe the action ("change the
status", "leave a note"), never an internal tool name. Branches inline ("If X -> Y;
otherwise -> Z"), and a final instruction saying what to output and how.
- Guardrails inline, matched to what the skill touches: writes need a confidence gate and
confirm-before-destructive, searches need result-cap honesty, PSA-bound notes need plain
text, when-in-doubt-do-nothing wherever a wrong action beats no action.
3. Check the length. Skill instructions cap at 3,000 characters — Super Magic refuses to
save a longer one. Count it, say where they stand, and over the cap fix it in this order:
a. Cut hedging and restated rationale. Most long prompts explain why a step matters two
or three times; say it once, in the imperative. Usually that is the whole fix.
b. Compose with a base skill — but only where the name resolves. In Super Magic the agent
reaches the member's other saved skills; a Flow fires a prompt with no skill lookup, so
a name there is just words, and anything a Flow runs needs its contract spelled out
inline. Keep a gloss that stands alone: "notes are plain text, no markdown or emojis
(apply the PSA Note Discipline skill)".
c. Split it, if it was always two skills.
Never buy space by deleting a guardrail — a confirmation before a destructive write, a
data-loss consent step, an escalation trigger. Aim for 2,900 so they have room later.
4. Show the revised skill side by side with a one-line rationale per change, and its
character count. Strip anything unshareable — credentials, API keys, person or board ids,
client names — replace with placeholders and say so. Create or update the skill on
explicit confirmation only, never without showing the full revised text.
5. Suggest a quick test: two or three prompts that should trigger it, one that shouldn't.
Coach, don't hijack — keep their intent and vocabulary; change structure, length and
guardrails, not the goal.