Skip to content

Briefing

Parent: HL-TFW-38 Goal: Empirical validation that TFW agents skip RF §6-8, reviewers rubber-stamp, and researchers omit Findings Maps — across real projects.

Research Plan

Gather

  • Scan RF files from helpdesk, vlm-3, content marketing, and other active TFW projects
  • Check which RF sections (§6 Fact Candidates, §7 Strategic Insights, §8 Diagrams) are present vs absent
  • Scan REVIEW files — check if reviewers provide evidence or just checkmarks
  • Scan RES files — check if Findings Map sections are present

Extract

  • Analyze handoff.md Phase 3 (lines 71-107) — what does it explicitly mention vs what it omits
  • Analyze review.md Step 1 — is there any verification/audit instruction?
  • Analyze research/base.md Step 6 — does it mention Findings Map?
  • Cross-reference: do templates contain sections that workflows don't mention?

Challenge

  • Counter-hypothesis: maybe agents DO fill §6-8 when prompted well — check for positive examples
  • Challenge the "workflow > template" claim — search for external evidence on LLM prompt-following behavior
  • Check if audit-style review creates overhead that makes it impractical for trivial tasks

Hypotheses (from HL §10)

# Hypothesis HL Status
H1 Adding explicit §6-8 enumeration to handoff.md Phase 3 will stop executors from skipping them open
H2 An audit verification step in review.md will change reviewer behavior from "trust" to "verify" open

Scope Intent

  • In scope: Empirical analysis of real TFW artifact quality; current workflow text analysis; external research on LLM instruction-following patterns
  • Out of scope: Implementation of fixes (that's Phase A/B); CI/linting enforcement (TFW-34 scope); template format changes

Guiding Questions

  1. Which active projects have the most RF/RES/REVIEW files to sample from?
  2. Are there any projects where §6-8 ARE consistently filled — and if so, what's different about the prompt or context there?
  3. Is the diagram## User Direction

Q1 answer: Projects to scan: helpdesk (/d/projects/research/helpdesk), atamat (/d/projects/research/atamat), auto-schedule-prototype (/d/projects/research/auto-schedule-prototype), and steps-framework itself. Helpdesk is newest/most active. Atamat has most tasks but older TFW version. Scheduler has a few recent tasks.

Q2 answer: No projects where §6-8 are consistently filled — helpdesk is closest but only in the most recent phases.

Q3 answer: No — diagram collection has never been attempted.

User steering (post-Challenge): Review should be structured like the research flow — with explicit stages and cognitive mode transitions (Read → Verify → Challenge → Synthesize). This is a deeper insight than "add an audit step."

Stage complete: YES