Doubleblind — Security skill for Claude Code
Three layers that cannot see each other's mistakes: re-derive every number in prose from a file you committed, brief a reviewer that was told nothing, and audit the figure a reader will actually see.
How to install Doubleblind
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open GuoCheng24/doubleblind and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Doubleblind does
Three layers that cannot see each other's mistakes: re-derive every number in prose from a file you committed, brief a reviewer that was told nothing, and audit the figure a reader will actually see. Works with Claude Code, Codex, DeepSeek, Kimi.
Alternatives in Security
- Release Audit — Use right before cutting a release — it spawns ONE fresh sibling session via CCC that audits only the JUDGMENT 130 ★
- Skills Audit — Audit every skill and command Forge can see — group by status (active, duplicate, deprecated, archived) 55 ★
- Copy Crusade — Unleash parallel Copy Purist agents to audit copy at BOTH the text level (vague buttons, robotic errors) AND t 51 ★
README
doubleblind
[](https://github.com/GuoCheng24/doubleblind/actions/workflows/ci.yml) [](https://pypi.org/project/doubleblind-audit/) [](https://www.python.org/) [](https://github.com/GuoCheng24/doubleblind/blob/main/LICENSE)
**Your agent wrote the report. Ask it whether the report is true and it will say yes.**
Not because it is lying. The reasoning that produced the claim is the reasoning being asked to check it, and it cannot notice what it did not think of the first time. The same goes for the tests it writes for itself and the review it gives itself when you ask it to look again.
Three things catch that, and **no two of them catch the same defects**:
| finds | structurally cannot see | |
|---|---|---|
a machine that recomputesdoubleblind trace |
a number that exists in no file; a bound stated tighter than the interval; a quantity written in words | anything nobody thought to check — above all a correct number inside a sentence that does not follow from it |
a reader with no contextdoubleblind review |
claims that do not follow; a comparison pointing the wrong way; a framing the data will not carry | anything needing exact recomputation — a fabricated number that looks plausible reads as fine |
a machine that looks at the picturedoubleblind render |
two labels that read as one word; a caption unreadable at the size it will be seen; a headline over artwork; a character the font could not draw | whether the shape a reader takes from the figure is the shape the data supports |
Each of the three is automated here, each one is told where it stops, and a ledger records what got through anyway — including the six defects that got through in thi
Related Skills
Figure Audit
Verify figure caption numerics against summary.json sidecars. Args: [--json]
A11y Review
Accessibility audit against WCAG 2.1 AA — contrast, keyboard navigation, ARIA, focus management, screen reader
Cleanup Repo
Read-only repo audit for stale/orphaned files, tracked cruft, accidentally-committed secrets, empty files, and
Ingest Convo
Ingest a whole past conversation into context — resolves which transcript, renders it clean via the committed
Audit Figure
Audit a figure, or the code that generates it, against the scientific-figures rules
Consistency Check
Audit a doc bundle for cross-reference integrity, count drift, pronoun ambiguity, audience assumptions, and ot
Related Agents
Voice Authenticator
Checks reader-facing prose against the voice profile in ~/.claude/guides/voice.md — register, sentence rhythm,
Adoption Adversary
Fresh-context, refute-framed review through the product-owner loss function — whether the thing LANDS. Hunts f
Postcard Reader
Triage pending oat-postcard mail for this Claude Code session. Files routine postcards into TODO.md and surfac
