Writing a Sidekick skill that works first time

The five-part structure every skill on this site uses, why each part matters, and the mistakes that make Sidekick guess.

Updated 19 Sept 2026

Sidekick’s own skills follow a recognisable shape: a description that says when to use the skill, numbered steps, an output section, and a closing paragraph of things it must not do. The skills in this library copy that shape because it works.

The five parts

1. Description: what it does and when to use it. One or two sentences. Sidekick reads this to decide whether the skill matches what you asked. “Finds sales opportunities in the store’s own data. Use when the user asks how to boost sales or fix slow-moving products.”

2. Steps: what data to read, in what order. Name the reports or objects. “Pull the last 30 complete days and the preceding 30: total sales, orders, AOV, sales by product.” Ask for both periods in one request so comparisons line up.

3. Rules for reasoning. “From the returned rows only”, “never compute a percentage change from a zero baseline”, “do not recommend a product that is out of stock”. These stop Sidekick filling gaps with generic advice.

4. Output: fix the format and length. “Lead with the biggest opportunity, then at most four more, one sentence each, with a small table.” Without this, you get a different shape every run.

5. Guardrails: what it must never do without asking. Create, edit, delete, send. “Recommend and wait for the user to ask.”

Mistakes that make Sidekick guess

  • Vague timeframes. Say “last 7 complete days in the shop’s timezone.”
  • No output format. You will get prose one day and a table the next.
  • Asking for changes and analysis in one breath without a confirmation step.
  • Long preambles about your brand. Put that in /brand-voice once.
  • Placeholders that are not obvious. Use [LIKE THIS] and keep them to the few things that change each run.

Test it three times

Run the skill with different inputs. If the output shape changes, tighten the Output section. If Sidekick asks a question every time, add the answer to the steps. Then save it, and consider submitting it.

Skills mentioned

/boost-sales Analyzes Medium

Find Sales Opportunities

Finds the biggest sales opportunities in your own data and recommends up to five concrete actions for this week.

Analytics Not yet verified
/weekly-review Analyzes Easy

Weekly Sales Review

One-screen weekly review: revenue, orders, AOV, top and declining products, and three actions for the week.

Analytics Not yet verified