# Research skill

> AI research with sources you can check. This free Claude skill tags every line observed, inferred, contested or unknown, and writes gaps down, not guesses.

- Canonical page: https://inspectedbymarko.com/skills/research/
- Site: Inspected by Marko (https://inspectedbymarko.com/)
- Last updated: 28 October 2026

## Research that says “I don’t know.”

AI research reads finished, even where it guessed. This skill stamps every line it writes: seen in a source, worked out, disputed or not found. What it can’t find, it writes down as a gap.

- [Get the skill](https://inspectedbymarko.com/skills/research/#free-skills)
- [See it mark a line](https://inspectedbymarko.com/skills/research/#one-line)

One skill · four guides · one check script · free, MIT License

From my thumbnails file, 12 Oct 2026. Lines cut for space, words as written. The printout: `research--youtube-growth-thumbnails--2026-10-12.md`, 12 OCT 2026.

```text
---
source_count: 34
confidence_distribution: observed: 172, inferred: 30, contested: 1, unknown: 17
coverage: 10 of 10 sub-questions resolved
---
## Key findings
1. [OBSERVED] YouTube Help's "Add video thumbnails" page recommends a custom thumbnail resolution of 3840 x 2160 pixels for videos. [S1] (as of 2026-10-12, page undated) (single-source)
11. [INFERRED] A video thumbnail file larger than 2 MB and up to 50MB can be uploaded only from desktop, because YouTube Help caps mobile video thumbnail uploads at 2 MB. Rests on findings 9, 10. [S1]
## Open questions and gaps
7. [UNKNOWN] Any peer-reviewed study measuring the effect of faces or facial expressions in YouTube thumbnails on click-through rate, views or watch time. Blocking reason: none surfaced in the searches run; …
```

Handwritten beside it: “use a face? no study found. the file says so.”

## 01 Confidently wrong: AI called my strategy bulletproof.

It missed the main competitor, which already did almost everything the product did. The answer read finished. Nothing in it said where it was guessing.

I believed it six times this year:

| # | Idea | What AI said | Verdict |
| --- | --- | --- | --- |
| 1 | The prompt tool | “Nobody has built this.” | false |
| 2 | The dog collar | “Nobody has built this.” | false |
| 3 | (name withheld) | “Nobody has built this.” | false |
| 4 | (name withheld) | “Nobody has built this.” | false |
| 5 | (name withheld) | “Nobody has built this.” | false |
| 6 | (name withheld) | “Nobody has built this.” | false |

> “I believed ‘nobody has built this’ six times this year. Six times it was false.”

Two of the six, the prompt tool and the dog collar, are in video 01. The other four stay unnamed.

- Receipt · AI search, tested (6 Mar 2025): **60%+ wrong**. 8 AI search tools, 1,600 citation queries. The premium versions got more wrong than the free ones. Source: [Tow Center for Digital Journalism, Columbia Journalism Review, “AI Search Has a Citation Problem”](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php).
- Receipt · Same test, ChatGPT search (6 Mar 2025): “signaled a lack of confidence just fifteen times out of its two hundred responses”. Source: [Tow Center for Digital Journalism, Columbia Journalism Review, “AI Search Has a Citation Problem”](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php).

Handwritten: “it doubted itself 15 times in 200. mine has to stamp every line.”

## 02 Four stamps: Every line gets one of four stamps.

So you, and the next AI that reads the file, know how far to trust each line. These are the counts from my 12 YouTube growth files.

| Stamp | What it means | Lines | Share |
| --- | --- | --- | --- |
| OBSERVED (source opened) | A source it opened says so. The only stamp that reads as fact. | 2,237 | 81.0% |
| INFERRED (working shown) | Worked out from its own findings, with the finding numbers cited. | 287 | 10.4% |
| CONTESTED (both sides kept) | Credible sources disagree. Each side stays in, with why they differ. | 30 | 1.1% |
| UNKNOWN (not found) | Nothing credible found. Written down as a gap, never filled with a guess. | 209 | 7.6% |
| Total |  | 2,763 | 100% |

> a short honest file beats a long fabricated one.
>
> The skill’s own rule, word for word.

Counted from the tags in my 12 files: observed 2,237 · inferred 287 · contested 30 · unknown 209. Total 2,763.

## 03 One line, taken apart.

Each line stands on its own. Lift it out of the file and it still says who said it, when it was true and how sure to be.

The line, as it sits in the file (File line 39 · one of 220 stamped lines):

```text
1. [OBSERVED] YouTube Help's "Add video thumbnails" page recommends a custom thumbnail resolution of 3840 x 2160 pixels for videos. [S1] (as of 2026-10-12, page undated) (single-source)
```

| Part | What it is | Why it is there |
| --- | --- | --- |
| `1.` | its own number | So other lines can lean on it: “Rests on findings 9, 10.” |
| `[OBSERVED]` | the stamp | How far to trust it. This one: a source it opened says so. |
| `YouTube Help's "Add video thumbnails" page` | who says so | Named, so you can open it and check. |
| `recommends ... 3840 x 2160 pixels for videos.` | the claim, as written | Numbers and units exactly as the source states them. |
| `[S1]` | the receipt | Points to the source list at the end of the file, with the link. |
| `(as of 2026-10-12, page undated)` | when it was true | Specs, prices and counts change. Each one carries a date. |
| `(single-source)` | one source only | Only one source said it, so the line says so too. |

Open questions and gaps · Gap 7, from the same file (stamped UNKNOWN):

```text
7. [UNKNOWN] Any peer-reviewed study measuring the effect of faces or facial expressions in YouTube thumbnails on click-through rate, views or watch time. Blocking reason: none surfaced in the searches run; the SAGE journal article linked from a Tier 3 guide was retrieved in verification and is a lab study of visual temporal resolution that measures no thumbnails or viewing outcomes [S34]; the shared search budget was exhausted.
```

Handwritten: “a guide cited a study. it opened the study. nothing in it was about thumbnails.”

Contested · Finding 121, from the same file (stamped CONTESTED):

```text
121. [CONTESTED] How common exaggerated facial expressions are among high-performing YouTube thumbnails is disputed across vendor datasets: 5% of vidIQ's 2026 breakouts carried them, while VEED described exaggerated expressions as among the most common components of 2021 trending thumbnails; see the contested section (Tier 3, low-confidence). [S8, S16, S24]
```

Handwritten: “both sides kept. weak sources, flagged as weak.”

## 04 The rules: Five things it never does.

Each one is a way AI research looks finished when it isn’t. The skill would rather hand you a shorter file.

Rules · they override looking complete (From SKILL.md):

1. **Never: Cite a source it didn’t open.** No remembered links, no rebuilt URLs. If it didn’t open it this run, it doesn’t exist.
2. **Never: State a fact it can’t trace.** No source means UNKNOWN, written down as a gap.
3. **Never: Invent precision.** Numbers, units and dates exactly as the source states them. No extra decimals.
4. **Never: Promote a guess.** Passing INFERRED off as OBSERVED is the one fatal error.
5. **Never: Pad the file to look finished.** If the core can’t be sourced, it stops and tells you to narrow the question.

Three passes, then it stops:

1. **Break it down, search wide.** Each sub-question searched on its own, from more than one side, including the side that would prove it wrong.
2. **Fill the gaps.** Weak spots and disagreements get a second, targeted pass.
3. **Try to break it. Then ship.** Every load-bearing claim gets a query built to disprove it. What doesn’t survive is cut.

Handwritten: “pass 3 is the one that stings.”

The check it runs on its own file. Real output, 21 Oct 2026, on my thumbnails file. Header trimmed, path shortened.

```text
$ python scripts/validate_research.py research--youtube-growth-thumbnails--2026-10-12.md
NOTES
  Registry tier labels: Tier1=14, Tier2=10, Tier3=10
  Tag counts in body: OBSERVED=172, INFERRED=30, CONTESTED=1, UNKNOWN=17 (total 220)
FAIL (0)  provenance and structure, must be fixed
  none
WARN (0)  review each, fix unless justified
  none
RESULT: structural checks passed. WARN items still need a human pass, and
this script does not check fabrication, source quality, opinion, or
calibration. Run the self-audit gate in references/verification.md as well.
```

Handwritten: “even the check says what it can’t check.”

## 05 Receipts: I ran it on my own channel first.

Before the first video, I asked how YouTube growth actually works. Twelve questions, twelve files, every line stamped. My launch plan cites them.

The drawer · 12 files, compiled 12 Oct 2026:

| # | File | Sources opened | Unknown |
| --- | --- | --- | --- |
| 01 | Algorithm and discovery | 48 | 17 |
| 02 | Analytics and benchmarks | 53 | 17 |
| 03 | Channel page and branding | 41 | 22 |
| 04 | Descriptions and metadata | 81 | 20 |
| 05 | Engagement and promotion | 75 | 20 |
| 06 | Monetization and policies | 58 | 15 |
| 07 | Retention, hooks and length | 38 | 16 |
| 08 | Search and SEO | 54 | 18 |
| 09 | Shorts and long-form | 50 | 17 |
| 10 | Thumbnails (stamped “On top of this page”) | 34 | 17 |
| 11 | Titles | 40 | 16 |
| 12 | Upload cadence and timing (circled: “this one set my launch day. see below.”) | 44 | 14 |

My YouTube growth research (Compiled 12 Oct 2026 · 12 files):

| Total | Count |
| --- | --- |
| Sources opened | 616 (367 tier 1 · 139 tier 2 · 110 tier 3) |
| Observed | 2,237 |
| Inferred | 287 |
| Contested | 30 |
| Unknown | 209 (Written down as gaps. Not guessed.) |
| Stamped lines | 2,763 |
| Sub-questions | 107 of 108 |

The one left open says so in its file.

One line, one decision:

- From the cadence file · conclusion 6: `6. [OBSERVED] YouTube sends a viewer a maximum of 3 upload and live stream notifications from each channel in a 24-hour period. [S2, S1] (as of 2026-10-12, pages undated)`
- My launch day · cadence plan, 20 Oct 2026: Three videos go public at once. Nothing else that day: no Short, no trailer.

Handwritten: “one line, one decision. that’s what the stamps are for.”

## 06 Get the skill: Want research you can check?

One skill, four guides and the check it runs on its own file, in one zip. Read every file before you run it. I would.

```text
research-skill.zip
└─ inspected-by-m-deep-research/
   ├─ LICENSE  (MIT, word for word)
   ├─ README.md  (how to install it)
   ├─ SKILL.md  (the rules and the loop)
   ├─ references/  (4 guides)
   │  ├─ search-tactics.md  (how it searches)
   │  ├─ source-directory.md  (where good sources live)
   │  ├─ output-spec.md  (what the file must hold)
   │  └─ verification.md  (how it attacks its own claims)
   ├─ assets/template.md  (the file’s skeleton)
   └─ scripts/validate_research.py  (the check)
```

- **Works in:** Claude Code, in the Claude desktop app, with web search on.
- **Needs:** Python 3 for the check script. Nothing else.
- **Yours to:** Copy, change and share, under the MIT License.

How to get it: enter your email in the form on the page, “Get the research skill.” ([Email me the skill](https://inspectedbymarko.com/skills/research/#free-skills)). The zip comes by email (version of 28 October 2026).

Free to copy, change and share, under the MIT License. It comes with no warranty: read a skill before you run it.

Box not ticked? You get the skill and nothing else, and your address is deleted after 30 days.

[How I use your email](https://inspectedbymarko.com/privacy/#what-we-collect)

## 07 Work with me: Research done. Now build it right.

I build MVPs, do rebuilds and run fix sprints after an inspection. Quoted on a 20-minute call. I do 80 to 90% of the work with my AI skills. My team polishes and tests.

- **Builds:** Quoted on a call. MVP builds · rebuilds · fix sprints after an inspection. [Ask about a build](https://inspectedbymarko.com/call/?need=build#pick-a-time)
- **The inspection:** from $3,500 (USD · starting price). One app or repository · report pack in 5 working days. [See what the inspection covers](https://inspectedbymarko.com/work-with-me/#scope)
- [Request a 20-minute call](https://inspectedbymarko.com/call/): Free, 20 minutes, no code access. For apps with paying users.
- [Not ready? Watch the inspections on YouTube](https://www.youtube.com/@InspectedByMarko)

---

Provider and data controller: OneCodeToRuleAll d.o.o. (OCTRA) · Seat: Sinđelićeva 6, 19000 Zaječar, Serbia · Registered with the Serbian Business Registers Agency, no. 21564516 · Tax ID (PIB) 111898813 · marko@inspectedbymarko.com