mixtapetools

MixtapeTools is a collection of tools, templates, and philosophies designed for empirical economics research using Claude Code. The core philosophy treats Claude as a thinking partner rather than an autonomous agent or a simple code generator. The workflow emphasizes iterative steering, where the user thinks through the AI to maintain clarity.

Key principles of the workflow include:

• External memory via markdown: Using markdown files to provide institutional memory across Claude sessions to combat AI amnesia.
• Cross-software replication: Ensuring results are consistent across different environments (e.g., R, Stata, and Python).
• Adversarial review (Referee 2): Using fresh Claude instances to audit work to avoid self-grading bias.
• Verification through visualization: Prioritizing visual inspection of data to catch errors that numbers might hide.
• Documentation as first-class output: Treating documentation as an essential part of the research process.

The dashboard is organized into functional groups to prevent redundancy. Each tab serves a specific role in the research/audit lifecycle:

Output Group

• Narrative: The current best understanding of truth (present-tense). It is populated from narrative.md or auto-assembled from confirmed hypotheses and insights. It features Drift detection (warning if a narrative references a non-confirmed hypothesis) and Inline status badges for hypothesis IDs.

Structure Group

• Overview: Project management view showing operational health, hypothesis counts, and recent audit cards.
• Courtroom: Evaluates the five stages of proof (Show Bite, Event Studies, Falsification, Main Results, Mechanisms) to determine if evidence supports a causal claim.
• Checklist: A mechanical recipe of nine DiD (Difference-in-Differences) steps. Status is based on file existence, not quality.
• Code: A view-only unified code viewer. It distinguishes between Pipeline scripts (in the DAG, marked with freshness dots) and Stale scripts (not in the DAG).
• Data: A catalog of raw source datasets, not intermediate files.

Evidence Group

• Hypotheses: The audit trail of all claims (conjecture, testing, confirmed, complicated, rejected).
• Insights: A chronological log of empirical findings requiring exact numbers and hypothesis links.
• Decisions: A record of binding design choices that constrain downstream work.
• Figures: A gallery of PNGs from output/figures/ presented as flippable cards showing provenance.
• Tables: A viewer for LaTeX (.tex) and CSV files from output/tables/.

Reference Group

• Skills: A reference list of available Claude Code skills.

The weight of Ethos, Pathos, and Logos should shift depending on whether you are teaching or presenting research.

Classroom (Focus: Orientation)

Students are often less expert and more vulnerable to confusion. The goal is to guide them safely.

• Ethos: "I can guide you safely through this."
• Pathos: "You can learn this; the difficulty is manageable."
• Logos: "Each step follows, and you can see why."
• Strategy: Use more signposting, repetition, worked examples, and slower notation introduction.

Research Seminar (Focus: Credibility)

Peers are more expert and skeptical. The goal is to establish credibility under uncertainty.

• Ethos: "I know the literature, the method, and the threats."
• Pathos: "This problem matters, and the tension is real."
• Logos: "The evidence supports this claim under clearly stated assumptions."
• Strategy: Lead with the question/stakes, identify variation early, use assertion titles, and show one proof object per slide.

A slide deck is a form of sequential visual rhetoric designed to accompany a spoken argument. To use this medium effectively, understand its four constitutive properties:

• Sequential: Decks are ordered (Slide 1 $\rightarrow$ Slide 2 $\rightarrow$ Slide 3). This temporal order creates a narrative that imposes a specific sequence of ideas on the audience.
• Visual: Meaning is carried through visual composition (typography, color, imagery, white space, data visualization) rather than just text. The visual channel conveys patterns and relationships that text cannot.
• Rhetoric: Decks aim to persuade. They should utilize the Aristotelian framework: establishing credibility (ethos), engaging emotion (pathos), and presenting logical argument (logos).
• Designed to accompany spoken argument: A deck is not a document. The slides should act as visual anchors, evidence containers, or structural markers. If a slide contains everything the speaker says, it has failed as a slide and become a document pretending to be a presentation.

In the gtd (Get Things Done) system, hypotheses are documented using a structured Markdown format that includes metadata, a claim, a 'Courtroom' for causal identification, a 'Kills It' section for falsification criteria, and an 'Evidence' section for tracking results.

Metadata Schema

• id: Unique identifier (e.g., H01).
• title: A concise statement of the hypothesis.
• status: Current state (e.g., testing).
• parent/children: Links to hierarchical sub-claims.
• date_proposed/date_resolved: Temporal tracking.

Causal Identification (Courtroom)

To ensure scientific rigor, each hypothesis defines:

• Estimand: The specific effect being measured.
• Population: The target group.
• Variation: The source of causal identification (e.g., staggered rollout).
• Mechanism: How the variation occurs.
• Falsification: The design used to rule out null hypotheses (e.g., placebo tests).
• Sub-claims: Granular breakdowns (e.g., by urban vs. rural).

---
id: H01
title: Free tutoring increases attendance
status: testing
parent: null
children: [H01a, H01b]
date_proposed: 2026-04-01
date_resolved: null
---

## Claim
Districts that adopted the free after-school tutoring program experienced increased daily attendance rates...

## Courtroom
- **Estimand**: ATT of the tutoring program on daily attendance rate (percentage points)
- **Population**: Students in grades 3-8...
- **Variation**: Staggered district-level adoption...
- **Mechanism**: Districts opted in through a competitive grant application...
- **Falsification**: Same design in Fall 2025 (pre-program) should yield null
- **Sub-claims**: Urban districts (H01a) and rural districts (H01b)...

## Kills It
1. Pre-trends are not parallel...
2. Placebo estimate is statistically significant...
3. Effect disappears when controlling for confounders...

## Evidence
- [2026-04-10: Urban ATT = 2.3pp](../insights/2026-04-10_urban_att.md)

Effective slide decks are characterized by specific patterns that distinguish them from poor-quality or contextually inappropriate presentations. When evaluating or creating a deck, look for these three dimensions of patterns:

1. Domain-Specific Patterns

Different professional domains have evolved distinct norms:

• Management Consulting: Often follows rigid templates emphasizing recommendations and supporting evidence.
• Academic Economics: Typically leads with literature review and methodology before presenting conclusions.
• Product Marketing: Often leads with user stories and emotional hooks before presenting features.

2. Quality-Based Patterns

Effective decks share specific characteristics that separate them from ineffective ones:

• Effective Characteristics: Clarity of central message, economy of means, appropriate evidence, logical flow, and visual coherence.
• Ineffective Characteristics: Unclear purpose, overwhelming detail, missing evidence for claims, jumbled structure, and visual noise.

3. Context-Based Patterns

The level of detail and polish should match the intended setting:

• Board Presentations: Require high polish and specific information density.
• Team Working Sessions: May require more interaction and different levels of completion.
• Classroom Lectures/Academic Job Talks: Require specific structural orientations suited to teaching or formal defense.