Temporal Reasoning Sleuth

What this Claude Code skill does

Engineers temporal reasoning capabilities for AI agents — enabling them to trace decision chains, reconstruct causal sequences, and reason over event timelines spanning months or years of organizational history. Use when an agent must answer questions like 'what decisions led to X' or 'how did this situation evolve.'

Temporal Reasoning Sleuth is a packaged skill for Claude Code: a SKILL.md plus references and a runnable test that drops into .claude/skills/, so your agent can run this workflow reliably instead of you re-explaining it each time. Install it via MCP, run it as a CLI, or fork it into your own.

Inside the SKILL.md

<h1>Engineering Temporal Reasoning</h1>
<h2>Why LLMs Struggle with Temporal Reasoning at Scale</h2>
<p>LLMs have two temporal reasoning failure modes:</p>
<p><strong>1. Attention degradation</strong>: When a long context contains hundreds of events in chronological order, attention distributes across the entire sequence. The model cannot reliably identify which events are causally linked vs. merely adjacent in time. "Lost in the middle" is the symptom.</p>
<p><strong>2. Context poisoning</strong>: Events retrieved without their causal context contaminate reasoning. If you retrieve "auth service migrated to OAuth2" without "auth breach incident that caused it", the model may draw wrong conclusions about the migration's purpose.</p>
<p>The fix is not bigger context windows — it is structured temporal storage and targeted retrieval that feeds the model a <em>curated causal slice</em>, not a raw timeline dump.</p>
<pre class="mermaid">flowchart LR
    subgraph wrong[&quot;Raw Timeline Dump (fails)&quot;]
        direction TB
        E1[Event A] ~~~ E2[Event B] ~~~ E3[Event C] ~~~ E4[Event D] ~~~ E5[Event N...]
    end

    subgraph right[&quot;Curated Causal Slice (works)&quot;]
        direction LR
        C1[&quot;Auth Breach\n2024-02-01&quot;] --&gt;|CAUSED| C2[&quot;Decision: OAuth2\n2024-03-15&quot;]
        C2 --&gt;|TRIGGERED| C3[&quot;Implementation\n2024-04-10&quot;]
        C3 --&gt;|CAUSED| C4[&quot;Mobile Update\n2024-05-20&quot;]
    end

    wrong --&gt;|&quot;model gets lost&quot;| FAIL([Attention\ndegrades])
    right --&gt;|&quot;model follows chain&quot;| WIN([Correct\ncausal reasoning])

…