review-summary-faithfulness
Skill49 starsChanged 3 months agoMITView on GitHub
Reviewer criteria for AI-generated summaries via with_review. Checks faithfulness to source, no hallucinated facts, appropriate length, key points preserved. Use as criteria.skill when summarising articles, documents, transcripts, or batch-task outputs.
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
Every factual claim in the summary must be directly supported by the source text.
- Distinctive: line 19
If the summary uses quote marks, the quoted text must appear verbatim in the source.
- Distinctive: line 27
If the source has 3 distinct points and the summary mentions only one, REVISE
- Distinctive: line 36
Don't add conclusions / recommendations the source doesn't make
- Distinctive: line 37
Don't soften or harden the source's stance — match the original register
≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Summary faithfulness review criteria
Review the worker's summary draft against these checks.
Checks
- Faithful to source (most important):
- Every factual claim in the summary must be directly supported by the source text. No invented numbers, names, or dates.
- Inferences (e.g. "this implies X") are OK if the implication is reasonable. Speculation presented as fact is REVISE.
- No hallucinated quotations:
- If the summary uses quote marks, the quoted text must appear verbatim in the source. Always REVISE.
- Key points preserved:
- For a 2-paragraph summary of a 3-page document: the main argument/decision/finding should be in the summary
- Specific numbers, names, deadlines that drive the source's conclusion shouldn't be stripped in favour of generic prose
- If the source has 3 distinct points and the summary mentions only one, REVISE
- Length matches task:
- "One-sentence summary" → one sentence. Not 3.
- "Bullet-point key takeaways" → bullets, not prose
- "Executive summary" → ~3 short paragraphs unless task says more
- No editorialising:
- Don't add conclusions / recommendations the source doesn't make
- Don't soften or harden the source's stance — match the original register
- Reads as a summary, not a paraphrase:
- Compresses, doesn't just rephrase sentence-by-sentence
When to APPROVE
- Faithful to source on every factual claim
- Captures the points needed to act on the source without re-reading
- Right length + format for the task
When to REVISE
- One specific hallucinated fact
- Missed a key point
- Wrong length (too long, too short)
- Editorialised where the source was neutral
When to REJECT
- Summarised the wrong document (e.g. context was article A, summary describes article B)
- Mixed multiple sources into one summary when the task was per-source
- Output is so far from a summary it's a different artefact
Verdict format
VERDICT: APPROVE — captures all 3 key findings without invention
VERDICT: REVISE — claims revenue grew 30% but source says 23%
VERDICT: REJECT — summarised article 2 instead of article 1In this repository32 other files it gives its agents.
- AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- agent-asks-tasksSkill
- caretakerSkill
- code-reviewSkill
- compare-optionsSkill
- create-reportSkill
- csv-analyseSkill
- document-qaSkill
- draft-emailSkill
- enrich-errorSkill
- extract-structured-dataSkill
- fact-checkSkill
- git-diff-summariserSkill
- librarian-curateSkill
- meeting-notesSkill
- morning-briefSkill
- plan-taskSkill
- reflectSkill
- remember-conversationSkill
- reverieSkill
- review-code-securitySkill
- review-email-toneSkill
- review-outputSkill
- rewrite-for-audienceSkill
- route-findingSkill
- routine-health-checkSkill
- save-research-docSkill
- score-importanceSkill
- summarise-urlSkill
- web-researchSkill
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

