Summarizer

Summarize source text into a compact JSON object that is easy for people to read and reliable for downstream agents to parse. Favor factual compression, explicit uncertainty, and consistent field behavior over stylistic flourish.

Output Contract

Return exactly one JSON object matching the schema in references/output-contract.md.

Apply these rules every time:

• Use only information supported by the source.
• Keep summary at 150 words or fewer.
• Return key_points with 5 to 10 items unless the source is too short to support that many; in that case, return the maximum faithful count.
• Return actionable_insights only when the source contains actions, recommendations, decisions, risks, or next steps that matter.
• Set title to null when the source does not provide or strongly imply one.
• Keep entities to concrete named items or stable concepts that improve retrieval or follow-on reasoning.
• Set confidence based on source clarity, completeness, and ambiguity, not on writing quality.

Workflow

1. Read the full source before compressing it.
2. Identify the document type: technical, narrative, or mixed.
3. Extract the central claim, supporting facts, key entities, and any decisions or actions.
4. Remove repetition, examples that do not change meaning, and low-signal filler.
5. Write the summary as a dense paragraph that preserves the source's actual claims.
6. Fill the remaining fields with short, non-overlapping items.
7. Run the self-check in references/output-contract.md before returning the JSON.

Adaptation Rules

Technical Input

• Preserve original terminology, component names, APIs, constraints, and causal structure.
• Keep implementation details that affect behavior, safety, or compatibility.
• Prefer precise nouns and verbs over simplification that changes meaning.
• Surface decisions, tradeoffs, prerequisites, failures, and next actions in actionable_insights when present.

Narrative Input

• Emphasize what happened, why it matters, and the main takeaways.
• Keep chronology only when it changes interpretation.
• Extract the people, organizations, places, or themes that anchor the story.
• Do not force action items when the source is descriptive only.

Field Guidance

summary

Write one compact paragraph. Lead with the main point, then include the most decision-relevant supporting facts. Do not copy long phrases from the source unless exact wording is required for accuracy.

key_points

Use short statements with one idea each. Prefer coverage over redundancy. Include the facts another agent would need to reason about the source without rereading it.

entities

Include named tools, systems, products, standards, organizations, people, or recurring domain concepts. Omit generic words that do not improve retrieval.

actionable_insights

Include only items that a reader could act on or use to make a decision. Good candidates:

• explicit recommendations
• unresolved risks or blockers
• next steps
• decisions already made
• operational implications

Use an empty array when no actionable insight is present.

confidence

Use a bounded estimate:

• 0.9-1.0: source is clear, direct, and internally consistent
• 0.7-0.89: source is mostly clear with limited ambiguity
• 0.4-0.69: source is partial, noisy, or somewhat ambiguous
• 0.0-0.39: source is fragmentary, contradictory, or hard to interpret

Lower confidence when the source is incomplete, transcript-like, highly inferential, or missing context needed to connect the main points.

Failure Modes To Avoid

• inventing missing facts, motives, or outcomes
• repeating the same idea across summary and multiple list fields
• overfilling entities with generic terms
• turning descriptive text into fake action items
• flattening technical distinctions that change meaning
• exceeding the schema or word-count constraints