add-entry - SKILL.md Agent Skill

name: add-entry description: >- Add a single new element to a process wiki section, AI-drafted. Read the section's context, ask the SME in the chat what they want to add, research it in the wiki and — where useful — on the web, draft an initial proposal, refine it with the SME, and write it on approval. Use this whenever the user wants to add an entry or element to a section — a market trend, innovation idea, competitor, CX benchmark or any other element type.

Add Entry

You add one new element to a section of a process wiki — AI-drafted, researched, and refined with the SME before it is written. You are invoked with a process <slug> and a <section> id.

You do the judgement — ask, research, draft, refine — and the native tools do the mechanical write. This is the interactive, single-entry counterpart to the bulk source-innovation / source-cx skills.

Step 1 — Read the section's context

use the getSectionContext({ slug, section }) tool once — it returns everything Step 1 needs in one call, so you never re-recall the element shape:

types — the element type(s) this section holds, each as a fill-in-the- blanks skeleton: the blocks (the ## headings to write), the fields (type-specific frontmatter, with hints), the relations (each with the element type(s) it may point at), and required (the frontmatter keys that must be present — drop one and conformance fails). Draft against this skeleton; do not recall the shape from the schema by memory.
existing — the section's current elements (id + title), so the new entry fits and does not duplicate one already there (see the duplicate check in Step 3).
overview — the process overview (domain, what it does), for context.

If types holds more than one element type, note them (you pick the right one from the SME's description in Step 3). expandElement({ type, id }) is still available if you need a specific existing element's full body.

Step 2 — Ask what to add

Ask the SME in the chat, substituting the section label:

What would you like to add to {section label}? Describe it in a sentence or two — I'll research it and draft a proposal you can refine.

Wait for their answer.

Step 3 — Research, then draft

Duplicate check first. Compare the SME's description against the existing titles from Step 1; if it looks close to one already there, also use the searchProcesses({ query }) tool to confirm. If it is a near-duplicate, say so and ask whether to extend the existing element instead of drafting a new one — don't silently create a second copy.

Research the entry before drafting — never draft from a blank guess:

Always read the related wiki elements — for an innovation idea, the pain- and friction-points it might address; for a control, the process steps; for a competitor move, the trends and the As-Is. The entry must connect to what the wiki already holds.
Use web search when the entry is an external signal — a market trend, a competitor move, a CX benchmark — or when the SME's request needs facts you do not have. Cite the real source; never invent one. When the entry needs several external signals (e.g. the trend, a competitor and a benchmark), dispatch the searches concurrently as read-only sub-agents in one message rather than running them serially, then merge the findings.

Then draft the element against the Step-1 skeleton for the chosen type — fill every block, every field, and every required relation it lists (for a web-sourced type that means sourceUrl, asOf and the rest — do not leave them off), an honest confidence, and a source (the SME, the wiki element, or the study/URL the research found). Because you draft straight into the skeleton, no required field is dropped and the post-write conformance check passes first time.

Pre-fill the relations. Each relation in the skeleton names the element type(s) it may point at; match the SME's description (and your research) against those target sections and pre-populate the relation id-lists for the SME to confirm in Step 4 — relations are the field most often left empty, and grounding the entry is the point.

If the section holds several element types, pick the type from what the SME described; if it is genuinely ambiguous, ask which.

Step 4 — Refine with the SME — Y / E / R

Present the draft and offer exactly three choices:

[Y] Yes — accept the draft. Write the element as status: draft; the SME approves it later in the app, not here.
[E] Edit — apply the SME's corrections, show the result, ask again.
[R] Rewrite — the draft missed; redraft (sharper questions, more research) and re-present.

Loop until [Y]. Always offer all three.

Step 5 — Write it

Do not call createElement during Steps 3 or 4. The element exists only in the chat until the SME sends [Y]. Writing before confirmation is forbidden — no exceptions, even if the draft looks final.

While drafting and presenting (Steps 3–4) the element has no id yet — refer to it by description, never guess an id. The id is assigned here, at write time.

On [Y]:

Assemble an element object (title, confidence, source, the scalar frontmatter fields, the relation id-lists and the blocks) and use the createElement({ type, element }) tool — status: draft. Do not set an id: the backend assigns it and returns it in the result.
use the checkConformance({ slug }) tool — fix any flag.

Then confirm with this exact line, substituting the id, title and section:

Added {id} — {title} — to {section label}. Review it on the card.

Scope

You add one element per run, to the section you were invoked for. Everything you write is status: draft — the SME reviews it on the card (approval, or the relevance triage for web-sourced types); you never set approved or relevance yourself. You never write content you cannot ground in the wiki, the SME, or a real source. To add to a different section, the SME runs you again from that section.