1---
2name: cold-read
3description: Review anything you are about to hand to a person (screenshot, GIF, chart, UI state, user-facing string, error message, PR or commit body, report) as a reader with no context, before it ships. Use this before sending an image or chart, before reporting a UI change done, before filing a PR or writing a commit message, and before any string a user will read. It catches what the author cannot see after working with full context. Cut-off text, jargon, dead ends, demo-only data and claims nobody can judge without the thread. Lighter than the whole-app tester in `ux-testing`.
4---
5
6# Cold read
7
8After hours in a task, every gap in the artifact is filled in by what you know.
9You read "needs ACP adapter" and see a precise term; the owner sees nonsense.
10Re-reading with that context passes the same flaws, so the check has to run in a
11reader who never had it.
12
13`AGENTS.md` says when: anything that leaves the thread. This skill says how.
14`ux-testing` ("fresh eyes") drives a whole app as a first-time user; this is the
15per-artifact version of the same idea.
16
17## 1. Make the artifact the thing the reader sees
18
19- Review the output, not its source: the PNG, the rendered page, the final PR
20 body, the string as it appears in the UI.
21- Use real data. A fixture that happens to fit hides truncation, `0.00`, odd
22 ordering, and entries the owner did not expect to see.
23- Capture every state a person can land in (empty, long, error, narrow), not just
24 the happy one.
25
26## 2. Self pass, every time (cheap)
27
28A screenshot script that exits 0 has been tested, not reviewed: the feature
29works and the image can still be cropped, blank, stale, or showing the wrong
30state. Capture and send must be separate steps, never one script or one call,
31and no image goes to the owner (or into a PR, report, or `html_render`) that you
32have not opened in this turn.
33
34Open the file with the image viewer and write down, in plain words, what is on
35it before judging. Check that it shows the state you meant to prove, not just
36some state. Then check:
37
38- text ending in `…` (a finding even when intentional, since it reads as cut
39 off), clipped at an edge, or overlapping something;
40- uneven spacing, empty regions, content hugging one side;
41- marks with no label: icons, badges, `+1`, colours, ticks;
42- a chart's title, axis names, units and legend. Can its claim be read at a
43 glance?
44
45## 3. Cold pass (anything past a one-liner)
46
47Launch a fresh sub-agent on Claude Sonnet at low effort. It gets the artifact,
48one line naming the audience, and the questions for its kind. It never gets the
49diff, the thread, the code, or what you meant it to say. When Claude usage is
50low, use `gpt-6.1-terra` at low effort instead, and check geometry and
51affordance yourself in the self pass (see "Choosing the reader").
52
53```
54You are <audience, e.g. "a developer who has never seen this app">. You have
55only the attached <artifact>. Do not read any code or docs. Say in two
56sentences what you think it shows or does. Then list at most three things a
57real person would stop or misunderstand at (confusing, cut off, undefined,
58misleading, missing), worst first, one line each: what you saw → what you
59expected. Skip taste and polish. Do not say whether you like it. If it is fine,
60say "nothing" and stop.
61
62Questions: <rows from the table below>
63```
64
65A reader that cannot restate the artifact correctly in its first two sentences
66is the finding.
67
68| artifact | ask |
69|---|---|
70| screenshot, UI state | Is any text cut off or any gap uneven? What can I do next, and can I undo it? Which things are buttons? |
71| UI copy, labels | Is every word and symbol defined for someone who just arrived? Does anything use the project's internal vocabulary? |
72| chart | What is the claim? Are axes, units and legend present? What would a skeptic say is missing? |
73| generated or computed content | Does anything look wrong for a real input (rounding, ordering, casing, an item that shouldn't be there)? |
74| form of the content | What does the reader decide from this, and is this the shape that serves that decision? |
75| PR, commit, issue body | Can a reviewer judge each claim without the thread? Is the size justified in the first lines? Is anything stated that the reader cannot verify? |
76| error or status | What happened, whose fault is it, what do I do? |
77
78## Choosing the reader
79
80Measured on 18 screenshots (9 with one planted flaw each, 9 fixed controls) at
81low, medium and high effort. Single runs, blatant flaws only; subtle ones such
82as a 2px misalignment are untested.
83
84| reader | planted flaws caught | findings per image | misses |
85|---|---|---|---|
86| Claude Sonnet, any effort | 9/9 | 3.3–4.1 | none |
87| gpt-6.1-sol | 6–8/9 | 1.1–1.4 | unclickable menu, uneven padding |
88| gpt-6.1-terra, any effort | 7/9 | 0.9–1.1 | unclickable menu, uneven padding |
89
90- More effort never caught more; on Sonnet it only added nitpicks, so use low.
91- Sonnet sees layout and affordance the others miss, and its extra findings are
92 mostly polish, which the three-finding cap in the brief removes.
93- Terra is solid on meaning, cut-off text, rounding and missing labels, and
94 stays quiet on clean images. It misses pure layout and affordance, so as the
95 fallback the self pass must check spacing and "what looks clickable" itself.
96- Sol offers nothing over Terra here. Do not use opus- or fable-class models.
97
98## 4. Act on it
99
100- Fix every real confusion and re-render. Re-run the cold pass once if the fix
101 changed layout or wording substantially.
102- Findings about taste or direction go to the owner as a question. The cold
103 reader cannot judge taste, so never ask it whether something is good.
104- Report the fix in one line after the artifact, never before it.
105
106## Where it has already failed
107
108Owner complaints this skill exists to prevent: a "Use reset link" cut off to
109`Use…`; "needs ACP adapter makes no sense to me"; "unclear where you get this
110key from"; "I can't tell what is a button"; "a learning rate reads 0.00";
111"deepseek is showing up here"; "displaying a table is kind of wrong"; "+4800,
112what the heck".