# Ideas

> Takes a topic idea, backs it with credible evidence, and develops concrete implementation proposals for a video from it. First checks whether the idea holds up at all. Use when a new idea exists and it needs to be clarified whether and how it holds up — or when new topics should be found from the current state of knowledge.

- Skill: `bendemartin97/ideas` (Agent Skill)
- Install (CLI): `npx skillmds@latest add bendemartin97/ideas`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bendemartin97/ideas/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: bendemartin97 (https://skillmd.com/u/bendemartin97)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bendemartin97/ideas

---


# Substantiate idea and translate into content

Two tasks, in this order:

1. **Does the idea hold up to the evidence?** If not, everything else is moot.
2. **How does it become a video?** Several paths, not one.

**This is not recycling.** It doesn't check what's already run, and nothing is derived
from existing posts — that's what `/recycling` is for. The source is outside
evidence, not the account's own archive.

## What comes in

An idea, as rough as it happens to be: a topic, an observation from daily work,
a question, a hunch, a sentence someone said.

If no idea comes with it, propose some — but **from the state of knowledge**: where
has the evidence shifted recently without common practice or public perception
catching up? Never from the account's own past posts.

## Step 1 · Sharpen the idea

Formulate in one sentence what the **testable claim** behind it is. "Make something
about sleep" becomes "caffeine after 2 p.m. measurably cuts deep sleep — its
half-life is longer than most people assume."

Without this sentence, nothing can be substantiated. If the idea is too vague to form
it, say so and ask exactly once.

## Step 2 · Search for evidence

**Actively research.** If web search is available, use it — don't rely on memory.
Search for the strongest sources the field has:

- for scientific claims: guidelines from professional societies, systematic
  reviews and meta-analyses, large original studies
- for craft, tech, and business claims: standards, primary data, manufacturer
  documentation, documented tests — never blogs citing blogs
- and explicitly also for **counter-positions**

### Binding rules

Credibility is the account's only capital. A fabricated piece of evidence costs more
than a hundred good ideas bring in.

- **A source is only named if it can be pinned down** — for a study: first author,
  year, journal; for anything else: origin and date. Otherwise `to be researched` —
  never a plausible-sounding source.
- **No number without a source.** No percentages, counts, or ratios from
  memory. Without a source, the number is left out, not estimated.
- **State the strength of the source:** for scientific claims the level of
  evidence (RCT, meta-analysis, cohort study, guideline, expert opinion), otherwise
  the source type (standard, primary data, documented test, single account). These
  are not the same and are not presented as if they were.
- **Name counter-positions**, even when they're inconvenient. A one-sidedly
  substantiated idea gets torn apart in the comments.

## Step 3 · Verdict

State clearly what the research found. Exactly one of these four answers:

| Verdict | Meaning |
|---|---|
| **Holds** | well substantiated, implementation follows below |
| **Holds with caveat** | substantiated, but only for a narrower claim — say which |
| **Contested** | studies contradict each other. **Usually the better idea:** the contradiction itself becomes the topic |
| **Doesn't hold** | the evidence doesn't support the claim |

**At "doesn't hold," the work stops.** Then don't propose an implementation, but
explain what's true instead — and ask whether that should become an idea. Burying an
idea is a full-fledged result, not a failure.

## Step 4 · Implementation

**Three variants with different angles**, not the same thing three times. They should
differ in *what* they use to enter the topic — say, via the gap between common
practice and the evidence, via a mistake the viewer recognizes in their own
routine, or via a widespread misconception.

Each variant consists of the triad that `FORMATS.md` requires, plus whatever's needed
for production:

```
### Variant <n> · <angle in two words>
ANCHOR:        <quote, striking phrase, or pointed thesis – not the topic>
MISCONCEPTION: <what is widely believed or practiced>
RESOLUTION:    <mechanism, number, or causal chain – concrete>
EVIDENCE:      <source, year · type and strength>
TO SHOW:       <what's displayed as proof – which screenshot, which spot>
STRUCTURE:     <the beats in one line>
PILLAR:        <from PILLARS.md>
STRENGTH:      <one sentence on what this variant can do that the others can't>
```

**On the TO SHOW field:** the proof image is the strongest building block of the
format — the evidence is shown, not claimed. Name concretely what's displayed and
which line on it needs to be legible. "Show the source" is not enough.

After that:

**Recommendation** — which variant, justified in two sentences.

**To be reviewed by a professional** — every claim that would appear in the video,
as a checklist. Claims in regulated fields — health, finance, law — need review by
someone qualified, even when they appear substantiated.

**Not proposed** — one or two obvious angles you deliberately left out, with a
reason. Makes the selection verifiable.

## What this skill doesn't do

- **No repurposing.** For a second cut or a story version, there's `/recycling`.
- **No topic lists.** "Ten ideas about sleep" is not a result. Better one idea
  that's substantiated and actionable.
- **No sugarcoating.** If the evidence is thinner than hoped, that goes in the
  verdict — not in a footnote.