fact-check
Use when reviewing design documents, architecture specs, technical proposals, or plans that make claims about API names, file paths, enum values, counts, or mechanism feasibility — a lightweight claim-by-claim check against the codebase with evidence. Built into embedded-workbenc
Install
npx skills add https://github.com/AmethystLuna/embedded-workbench/tree/master/skills/fact-check
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install amethystluna-embedded-workbench@llmmart
git clone https://github.com/AmethystLuna/embedded-workbench.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole amethystluna/embedded-workbench collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Fact Check (fallback)
Lightweight claim verification built into this plugin. Use it when the full
logicprobe plugin is not installed — it checks every verifiable claim in a
document/plan against the codebase with evidence, without model escalation.
When to use
- Design docs, architecture specs, technical proposals, refactoring plans that name APIs, types, files, enums, counts, or mechanisms.
- Proactively when a plan will be approved (plan-mode gate) and logicprobe is unavailable.
What to verify (per claim)
| Claim type | Method |
|---|---|
| Numeric claims (counts, sizes, frequencies) | grep -c / grep -rn |
| API/type/enum names | Extract actual signatures/values from headers |
| File paths and line numbers | Confirm existence and content |
| Mechanism feasibility | Check language standard / compiler support |
Process
- Enumerate every verifiable claim (quote it verbatim).
- Verify each against the codebase with evidence (file:line + excerpt).
- Output per claim: claim → evidence → verdict.
Output format
- Claim:
- Evidence: <file:line + grep/read excerpt>
- Verdict: confirmed | refuted | unverifiable
- Correction direction (when refuted) — never fix inline.
Append a ## Plan Verification block:
## Plan Verification
- **Mode**: fact-check (fallback — logicprobe plugin not installed)
- **Claims checked**: [N] confirmed, [M] refuted, [K] unverifiable
- **Escalation**: model verification NOT run (requires logicprobe plugin)
Boundaries
- No executable model verification — no state-space exploration, no harness.
- State machines (≥3 states), ACK/NACK/retry/timeout logic, lock/unlock
ordering, and "always"/"never"/"guaranteed" claims require the
logicprobeplugin: install it (see.dsh/INSTALL.md) or tell the user verification is degraded to manual mode for those claims. - Checking a claim is not a security audit; being checked ≠ being safe.
Files (embedded-workbench)
-
SKILL.md 2.5 KB
--- name: fact-check description: "Use when reviewing design documents, architecture specs, technical proposals, or plans that make claims about API names, file paths, enum values, counts, or mechanism feasibility — a lightweight claim-by-claim check against the codebase with evidence. Built into embedded-workbench as the fallback when the full logicprobe plugin is not installed. NOT for state machines, protocol/retry logic, or 'always'/'never' behavioral claims that need executable model verification — those require the logicprobe plugin." --- # Fact Check (fallback) Lightweight claim verification built into this plugin. Use it when the full `logicprobe` plugin is not installed — it checks every verifiable claim in a document/plan against the codebase with evidence, without model escalation. ## When to use - Design docs, architecture specs, technical proposals, refactoring plans that name APIs, types, files, enums, counts, or mechanisms. - Proactively when a plan will be approved (plan-mode gate) and logicprobe is unavailable. ## What to verify (per claim) | Claim type | Method | |-----------|--------| | Numeric claims (counts, sizes, frequencies) | `grep -c` / `grep -rn` | | API/type/enum names | Extract actual signatures/values from headers | | File paths and line numbers | Confirm existence and content | | Mechanism feasibility | Check language standard / compiler support | ## Process 1. Enumerate every verifiable claim (quote it verbatim). 2. Verify each against the codebase with evidence (file:line + excerpt). 3. Output per claim: claim → evidence → verdict. ## Output format - **Claim**: <verbatim quote> - **Evidence**: <file:line + grep/read excerpt> - **Verdict**: confirmed | refuted | unverifiable - **Correction direction** (when refuted) — never fix inline. Append a `## Plan Verification` block: ```markdown ## Plan Verification - **Mode**: fact-check (fallback — logicprobe plugin not installed) - **Claims checked**: [N] confirmed, [M] refuted, [K] unverifiable - **Escalation**: model verification NOT run (requires logicprobe plugin) ``` ## Boundaries - **No executable model verification** — no state-space exploration, no harness. - State machines (≥3 states), ACK/NACK/retry/timeout logic, lock/unlock ordering, and "always"/"never"/"guaranteed" claims require the `logicprobe` plugin: install it (see `.dsh/INSTALL.md`) or tell the user verification is degraded to manual mode for those claims. - Checking a claim is not a security audit; being checked ≠ being safe.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.