Understanding reviews
How Hosho scores
The four macro factors, the sub-factors, and what the numbers mean.
The four macro factors
Every review evaluates the prompt against a set of research-backed checks, grouped into four macro factors, each broken into named sub-factors. A sub-factor's number is the share of its checks that pass - continuous, and it moves the moment you fix something.
| Macro factor | What it asks | Sub-factors |
|---|---|---|
| Scope | Is the prompt one job, sized for one agent - or several jobs welded together? | Focus · Volume |
| Structure | Is it organized so a model can follow it - sections, tools, and a clear output contract? | Layout · Tools & skills · Provider fit · Output & validation |
| Guidance | Does it say what to achieve, what it will be given, and how to reason about the work? | Goal · Inputs · Method & reasoning |
| Coherence | Does it hold together - unambiguous, decidable, consistent with itself, and free of dead weight? | Clarity · Criteria · Consistency · Bloat |
The overall score
A prompt scores out of 10, with a label - Excellent, Good, Needs Work or Critical - and, on a PR, a verdict: approve, approve with suggestions, or request changes. Sub-factor readings render as traffic lights: red for a failing area, amber for a partly failing one, green for clean. A 10 means every area is clean - the overall is a verdict about the whole prompt, not an average that lets one dead area hide inside a good one.
Findings
Every point deducted has a name. A finding quotes the problematic text, anchors it to its lines, says why it fails, and proposes a concrete fix - in the editor as a card you accept or dismiss, in the MCP inline in your IDE, on a PR as a before/after snippet. Nothing is a vibe; if Hosho cannot point at the line, it does not flag it.
Full review, check, benchmark
- A full review runs every check from scratch. It spends a review run.
- Check changes re-checks a prompt Hosho has already reviewed after you edit it. Free on every plan - iterating on a prompt is the work, and we do not charge for the work.
- The benchmark is the sealed baseline those checks compare against - one per prompt, moved only forward, and only by a human clicking "Set new benchmark". Your agent cannot silently move the goalposts.
Reviewed alone, or as it runs
A prompt can be reviewed standalone, or in its runtime context: bundled with the skills it loads, the sibling prompts beside it, and the shared references it names - so a component is judged by its place in the assembly, and never marked as missing something its parent supplies. Every report says which basis it used.
Where the checks come from
The engine is built on three lines of evidence.
Hosho's own research on real production prompts, published prompt-engineering research, and each provider's documented conventions - distilled into a core set of principles plus provider-specific checks tuned by model family and by whether the target is a standard or a reasoning model. The reasoning is written up in Hosho Research.