TinyTroupe Feed Research Lab

Run bounded synthetic audience research for draft posts and X-style feed experiments inspired by TinyTroupe and public xai-org/x-algorithm architecture. Use when Codex needs to compare content drafts, persona reactions, conversation quality, link friction, safety/spam risk, or feed-loop hypotheses without claiming reach prediction, shadowban detection, live For You cloning, or private account analysis.

Use this skill to compare draft posts with synthetic audience personas and produce a research report. Treat outputs as qualitative pretesting and hypothesis generation, not live X ranking predictions.

Core Workflow

1. Collect 2-10 draft posts or content angles.
2. Clarify the target audience if available.
3. Run scripts/tinytroupe_feed_research_lab.py in deterministic mode.
4. Read feed_research_report.md, feed_research.json, and persona_reactions.csv.
5. Present the best draft, why it won, key objections, rewrite suggestions, and the boundary statement.
6. If the user asks for TinyTroupe proper, use the generated persona specs and experiment plan as the input to a separate TinyTroupe notebook or script.

Quick Start

SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/tinytroupe-feed-research-lab"
python3 "$SKILL_DIR/scripts/tinytroupe_feed_research_lab.py" \
  --audience "AI builders and creator-operators interested in X algorithm research" \
  --draft "I audited this viral X algorithm claim against public source. Verdict: misleading." \
  --draft "Replies are king. Here is what the public repo actually proves." \
  --output-dir /tmp/tinytroupe-feed-research

Use files:

python3 "$SKILL_DIR/scripts/tinytroupe_feed_research_lab.py" \
  --drafts-file /tmp/drafts.json \
  --personas-file /tmp/personas.json \
  --output-dir /tmp/tinytroupe-feed-research

The script writes:

• feed_research_report.md: human-readable comparison and rewrite guidance.
• feed_research.json: machine-readable drafts, personas, reactions, and warnings.
• persona_reactions.csv: row-level persona reactions.
• share_card.md: short public-safe summary.
• share_card.svg: visual summary card.
• tinytroupe_experiment_plan.md: optional bridge plan for a real TinyTroupe run.

Input Formats

--drafts-file accepts:

• JSON list of strings.
• JSON list of objects with id and text.
• Plain text blocks separated by ---.

--personas-file accepts JSON objects with:

• name
• segment
• interests
• dislikes
• reply_bias
• skepticism
• link_sensitivity
• safety_strictness

Missing persona fields fall back to conservative defaults.

Boundaries

Read references/research-boundaries.md before presenting results that mention algorithms, feed ranking, virality, reach, shadowbans, or account status.

Never say:

• "this predicts reach,"
• "this clones the X For You feed,"
• "this proves a shadowban,"
• "this optimizes for the live algorithm,"
• "this is what real users will do."

Prefer:

• "synthetic audience reaction,"
• "draft pretest,"
• "conversation-quality signal,"
• "X-style feed research sandbox,"
• "hypothesis to validate with real posting or user research."

TinyTroupe Bridge

The MVP script does not require TinyTroupe. It produces tinytroupe_experiment_plan.md so a later agent can create a TinyTroupe notebook with:

• the same personas,
• the same draft set,
• a structured reaction schema,
• a validation note that simulation outputs are research signals.

Companion Skills

Use x-algo-claim-auditor when the task is checking whether a viral algorithm claim is true. Use open-feed-recsys-lab when the task is verifying the public source repo, Phoenix artifact readiness, or architecture map.