write-skill

name	write-skill
description	The teaching guide for authoring a single high-quality Day AI skill prompt. The thinking process, the structure, worked examples, Day AI language conventions, and the revision test. Read this before writing any skill. Usage: /write-skill (read as a guide) — or invoke to craft one prompt interactively.

/write-skill

This is the craft document. A skill is a standing job you give a teammate's agent — work it does on a schedule, on an event, or on demand, without anyone having to ask. The difference between a skill that gets switched off in a week and one that becomes indispensable is almost entirely the prompt.

Read this before writing any prompt. If invoked directly with a target person and intent, walk through the process below interactively and produce one finished prompt.

The standard: would this be something you'd be proud to have the whole company see? If not, it's not done.

---

Before you write: understand the person

You cannot write a good skill for a role you've reduced to a title. Answer these four questions first — from workspace/PEOPLE.md, the workspace graph, and the plan — and write nothing until you can:

1. What is this person's job, really? Not their title — their daily reality. A "VP of Sales" at a 12-person company is doing the selling. A "VP of Sales" at a 500-person company is coaching and forecasting. The same title, two completely different skills.

2. What would be embarrassing for them to miss? A stale deal that's about to slip. A promise they made on a call and forgot. A relationship going quiet. An email they never replied to. The best skills are built around these specific moments, not around generic "insights."

3. What data does their agent actually have — and what's missing? Don't just build around what exists. Notice what's absent and turn it into value: "I'd have prepped you for that call if their account were in your customer memory." Gaps become coaching, not silence.

4. Where are they in their Day AI journey? A teammate with a sparse, new workspace needs the skill to coach them toward value — the workspace-health content is the value early on. A teammate with a rich workspace needs pure intelligence and zero hand-holding. The prompt must know which one it's talking to.

---

The structure of a great prompt

Once you understand the person, structure the prompt this way:

1. Open with identity and context. Tell the agent who it is, who the user is, what the company does, and what matters. Be specific — generic context produces generic output. "You are Ada, Pici's agent. Pici runs 5–8 demos a week and lives in his calendar; he's the first human a prospect meets." beats "You are a helpful sales assistant."

2. Structure around what the person needs, not data sources. This is the single highest-leverage move.

   • Bad: "Check email. Check the calendar. Check the pipeline."
   • Good: "Promises coming due. Relationships going quiet. Unfinished business from yesterday's calls. Prep for who he's meeting today."

3. Name the patterns explicitly, with examples. Not "surface relevant insights" — instead: "A promise coming due: three weeks ago on the Trust & Will call he said he'd send pricing by end of month. That's tomorrow and no email has gone out."

4. Set the quality bar with a number. "2–5 items, not 15. Each item should make them think, 'I'm glad someone caught that.'" Without a number, agents pad.

5. Account for an empty workspace. A prompt that assumes a fully instrumented workspace produces empty or misleading output for someone just starting. The best skills say what they would have caught if the data were there.

6. Pull the human into the loop early. On a skill's first runs, have the agent show its reasoning and invite correction: "I flagged these three as going quiet — were those the right ones?" As the skill matures, this fades.

7. Specify anti-patterns. What it must not do is as important as what it does: don't summarize meetings the person attended; don't repeat stale items; don't pad; don't manufacture urgency.

8. Give explicit permission to say nothing. "If there's genuinely nothing worth flagging, say so. Don't invent filler to look busy."

9. End with delivery format. Channel (Slack/email), shape (bullets/table/prose), and length. Match the channel: long analysis doesn't belong in a Slack DM.

---

Worked example

The ask: "A morning skill for our AE, Sarah."

Weak version (what not to write):

Every morning, check Sarah's email, calendar, and CRM. Summarize what's important and surface any relevant insights to help her have a productive day. Send it to Slack.

Why it fails: no identity, organized around data sources, "relevant insights" means nothing, no quality bar, no anti-patterns, would produce identical output for any salesperson on earth.

Strong version:

You are Sarah's agent. Sarah is an AE at Acme selling to mid-market RevOps leaders; she runs ~6 active deals at a time and her biggest risk is letting a warm deal go cold while she's heads-down on the next call. Every weekday at 7:30am, before she opens her laptop, give her the one screen she needs.

Lead with the single most important thing — name the person, the company, the dollar figure, the specific risk. Then, only if they earn the space:

• Promises coming due: something she committed to on a call that's now due or overdue. Quote what she said and when.
• Going quiet: an open deal with no touch in 7+ days where the last signal was positive. Draft the re-engagement email so she can send it in one click.
• Today's prep: for each external meeting today, the one thing she needs to remember — last conversation, open question, who else is in the room.

2–4 items, never more. If a deal is genuinely quiet because it's dead, don't resurface it. Don't summarize meetings she ran. Don't tell her things she already knows. If it's a quiet morning, say so in one line and stop.

Deliver as a Slack DM: bold the names, one line per item, the draft email in a collapsed block.

Notice: it would produce different output for a different person, every section earns its place, and you could not delete the context paragraph and still know who it's for.

---

Day AI language conventions

Every line a teammate reads from a skill must use Day AI's language:

• The AI is "your agent" or its name — never "your assistant" or "the bot."
• The data layer is "your customer memory" or "what I know" — never "the CRM" or "the database."
• Frame value as what becomes possible, not time saved.
• Name capabilities by what they do: draft an email, search the pipeline, send a Slack message, update a contact.
• No category labels, no competitor comparisons, no leading with "AI."

---

The revision test

After drafting, before deploying, check every one:

1. Would this produce different output for different people? If swapping the target person gives the same briefing, it's not specific enough.
2. Does every section earn its place? Cut any section that would rarely produce something useful.
3. Is there a single line that says "surface insights" or "provide relevant information"? Delete it. Replace with a named pattern.
4. Could you delete the identity/context paragraph and still know who this is for? If yes, it isn't grounded enough.
5. If the workspace is half-set-up, does the skill handle that gracefully? Or does it produce empty output?
6. Did you assume any reporting relationships? Never infer who manages whom from titles — only state relationships confirmed in the data.
7. Is the quality bar a number? "2–5 items" not "a few."
8. Does it have permission to say nothing? And explicit anti-patterns?
9. Is the channel/format/length specified and matched to the channel?
10. Does every line obey the Day AI language conventions?

---

Config that ships with the prompt

When you hand a finished prompt to agent-implementor (or write the manage_skills call yourself), specify:

• name and slashCommand (unique per agent, lowercase-hyphens).
• triggerType: SCHEDULE (with a cron triggerValue + timezone), EVENT (with event types), or NEITHER (on-demand).
• notificationType: ["slack"], ["email"], or both — and slackNotificationChannels if Slack. If you set no notification target, the skill defaults to delivering by email — so a skill never goes nowhere, but don't rely on the default by accident: choose the channel deliberately to match the output (long analysis → email; a short nudge → Slack), and set Slack explicitly when that's where the person actually works.
• scope: a per-agent skill (targetScope: "agent" + the teammate's targetAssistantId) when the prompt is tailored to one person; a workspace-library skill (targetScope: "workspace_library", deploymentMode: "MANAGED") when a whole team needs the same capability.
• Remember the tier budget: automated skills consume the target agent's slots. One excellent scheduled skill beats three thin ones.

A scheduled skill defaults well at 0 13 * * 1-5 (8am US Eastern, weekdays) unless you know the person's timezone — then adjust.