Ideation Agent Skill

Ideation — from a vague idea to a data-backed topic + a real dataset

The /data2story-pro orchestrator routes here in IDEA MODE: the reader handed over a hunch, a

question, or a half-formed angle instead of a dataset. Your job is to turn that into a concrete

topic that *real, findable data* can support, fetch that data, and hand a validated folder back to

the pipeline. You do this WITH the reader, not for them — two real checkpoints, no railroading.

You are not a pipeline role (no *_NN provenance prefix, no place in the 7 teams). You run once,

before Detective, and produce nothing that reaches the HTML except the dataset + a story_brief.

Inputs

• $1 = the reader's raw idea text (may be empty → open by inviting it).
• $2 = DATA2STORY_ROOT (resolved by the orchestrator; where data/<slug>/ will live).

Return contract (how the orchestrator continues)

• Success: emit a final line DATA_DIR=<absolute path to the validated dataset folder>. The

story_brief.json sits at <DATA_DIR>/meta/story_brief.json. The orchestrator sets

DATA_DIR/DATA_NAME from this and enters the normal pipeline (Detective → … → Inspector).

• Abort: emit IDEATION_ABORTED: <one-line reason> (reader stopped, or no real dataset supports

the idea after the bounded loop). The orchestrator halts honestly and runs NO pipeline. Never

fabricate data to manufacture a success.

The flow — 3 steps, 2 checkpoints

Interaction style — let the reader CHOOSE, don't make them compose. Drive the convergence

and BOTH checkpoints with AskUserQuestion: frame the angles / scope / data-forks as options the

reader clicks, not paragraphs they must write — picking is far lower-friction and each question

doubles as a micro-checkpoint. ALWAYS keep the Other / free-text escape open: the menu is your

framing, and the reader's own off-menu angle is often the best one, so never let it cage the

brainstorm. (This is NOT the cold opening questionnaire sparring-partner warns against — it is

choice-driven convergence *after* you have framed the space: lead the very first turn with

substance + an open invite, then switch to options.)

Step 1 — Converge the idea (reuse sparring-partner)

Run the brainstorming dialogue by following Skill sparring-partner with the mission in

references/sparring_brief.md: drive the reader from a vague idea to

ONE concrete data-story topic. Two non-negotiables on top of sparring-partner's normal process:

• Anti-sycophancy (its core stance) — do not rubber-stamp the first pretty idea.
• A feasibility pressure-test — relentlessly ask *does this data actually exist? at what

granularity? who publishes it? for which years/places?* A beautiful idea with no obtainable data

is a failure of this step, not a success. Steer toward a nearby idea the data CAN support.

The terminal of the dialogue is the story_brief (contract: references/schema.json) —

topic, angle, audience, the questions the data must answer, a structured data_needs spec, any

real candidate sources surfaced, and the exact find_data_invocation.query. Reply in the

reader's language (sparring-partner's rule).

CHECKPOINT 1 — confirm the brief

Show the reader the assembled story_brief (at least topic, angle, data_needs, and

find_data_invocation.query). Use AskUserQuestion (approve · edit · abort) or a plain confirm.

Loop back into Step 1 on edits. Do not proceed until the reader approves the brief. On abort →

return IDEATION_ABORTED: reader stopped at brief.

Step 2 — Acquire a REAL dataset (reuse find-data, web-first)

Derive a kebab-case slug from story_brief.topic; set OUT_DIR to the ABSOLUTE path

$2/data/<slug> (resolve $2 to an absolute path first). Then follow Skill find-data with the

brief's query and ALWAYS pass that explicit --out OUT_DIR — never rely on find-data's bare default

(its default is DATASETS_ROOT/<name>, a DIFFERENT root: ./datasets/<name>, not data/<slug>).

An explicit --out always wins, so the dataset is guaranteed to land at the path ideation chose:

Skill find-data "<story_brief.find_data_invocation.query>" --out OUT_DIR [--mode <single|theme>] [other flags]

find-data searches (web-first on an open-source machine with no local corpora), fetches, and runs

its 4 completeness gates, writing OUT_DIR/validate.json. Read that file back for the verdict. The

dataset files land directly under OUT_DIR, and the DATA_DIR returned to the orchestrator (the

success line below) is exactly that absolute OUT_DIR — not find-data's default location.

Bounded acquisition loop (≤ 2 attempts). If find-data returns BLOCKED / no adequate dataset:

• Surface honestly what was and wasn't found (the failing gates).
• Offer the reader: (a) re-enter Step 1 to pivot/narrow the topic (often the data exists only

at a coarser granularity — adjust the brief), (b) try an alternate real source/query, or

(c) abort.

• Never invent a dataset, a source URL, or a license to "succeed."

After 2 failed attempts with no path forward → return `IDEATION_ABORTED: no real dataset supports

this idea (closest gap: <gate>); suggested pivot: <one line>`.

CHECKPOINT 2 — confirm the dataset

Show the reader the fetched files + the gate verdict, and check them against story_brief.acceptance

(does it actually have the entities / metric / coverage you agreed on?). AskUserQuestion

(use it · send back to Step 2 · abort). Do not proceed until approved.

Step 3 — Finalize + hand off

Only AFTER find-data's audit has run (so it never lands inside the data-file glob), write the

approved brief to OUT_DIR/meta/story_brief.json:

mkdir -p "OUT_DIR/meta" && # write story_brief.json there (valid JSON matching references/schema.json)

It carries the reader's intent into provenance; the Detective MAY read it for human-intent context

(loose coupling — not required). Then emit the success line:

DATA_DIR=OUT_DIR

Guardrails

• Real data only. No synthesis, no simulated rows, no fabricated source URLs or licenses — that

would break the whole verifiability premise. "Can't find data" is an honest IDEATION_ABORTED,

not a reason to invent it.

• Checkpoints are real stops. The reader drives; you converge with them, not at them.
• Portability. No hardcoded machine paths — derive everything from $2 and the resolved skill

dir. Works on a fresh open-source clone with no local data corpora.

• Stay in your lane. You write only inside OUT_DIR (the dataset folder). You do not build HTML,

run the pipeline, or touch any role artifact — that's the orchestrator's job after you return.

Reference files

• references/schema.json — the story_brief contract (annotated example).
• references/sparring_brief.md — the specialized mission handed to

sparring-partner, with the feasibility pressure-test and a worked vague-idea → brief example.