Skip to content

Onboarding template overhaul + 6 evaluation-driven iterations#11

Open
davewaring wants to merge 6 commits intomainfrom
onboarding-templates-v2
Open

Onboarding template overhaul + 6 evaluation-driven iterations#11
davewaring wants to merge 6 commits intomainfrom
onboarding-templates-v2

Conversation

@davewaring
Copy link
Copy Markdown
Contributor

@davewaring davewaring commented Mar 29, 2026

Summary

Complete overhaul of the onboarding starter pack templates, driven by 45 automated persona evaluations (15 runs × 3 rounds). Every change was validated by re-running personas and comparing before/after output.

What changed

Bootstrap AGENT.md:

  • User stories are the most important output — specific enough that the owner thinks "yes, that's exactly what I want"
  • Confirm user stories with the owner before writing the plan
  • One question at a time, never accept vague answers (unchanged but reinforced)

All 5 domain spec templates (finance, fitness, career, relationships, new-project):

  • User stories restructured as the core of the spec — everything else supports them
  • Richer examples showing personal motivation and success criteria, not just functional goals
  • Owner's exact words guidance: "quote, don't paraphrase"
  • Two-tier gap hierarchy: "Before the plan is complete" vs "Worth exploring later"
  • Supporting sections explicitly tied back to user stories

All 5 domain plan templates:

  • Roadmap phases only serve user stories — "Worth exploring later" items don't get phases
  • Each phase leads with what life looks like when it's done (the pull) before listing steps
  • Destination section quotes the owner's success criteria from their stories
  • "What Needs More Work" respects the spec's gap tiers

All 4 domain AGENT.md files (domain-specific):

  • Finance: life-transition must-asks (insurance, tax, deadlines), emotional spending probing
  • Career: burnout/escape flags trauma as plan prerequisite
  • Fitness: post-injury fear flags mental health support as Phase 1 gap
  • New-project: support system inquiry for sensitive topics, probe connections between interests

Gateway + BD+1:

  • Project context directive passed to advisor
  • BD+1 AGENT.md updated

Evaluation results

  • 45 total automated persona runs (v1 + v3 full sweeps, v2 targeted)
  • 15/15 launch-ready in v1, 15/15 in v3
  • User stories per persona: ~1-2 in v1 → ~2-3 in v3
  • Confirmation step surfaced latent needs in every domain
  • Vague personas (Katie B) benefit most: vagueness converted to validated specificity
  • All improvements verified through before/after comparison

Test plan

  • Run onboarding interview in each domain — verify user stories are confirmed before plan is written
  • Verify spec output has specific user stories with personal motivation and success criteria
  • Verify spec uses two-tier gap structure
  • Verify plan phases only address "Before the plan is complete" gaps
  • Verify plan Destination uses owner's words, not advisor editorial
  • Verify advisor knows which project it's in (project context directive)
  • Verify no regression in existing chat functionality

🤖 Generated with Claude Code

davewaring and others added 6 commits March 29, 2026 07:57
Starter pack templates (17 files):
- Bootstrap prompt: personal advisor identity, template-as-north-star
  methodology, landscape-first, confirm-as-you-go, empowerment, owner
  profile, no-generic-advice differentiator
- Domain AGENT.md files: self-contained expertise per domain, knowledge
  literacy sections, blind spots, interview branches, domain tone
- Spec placeholders: 5-section structure with user stories
  ("As [who], I want [what], so that [why]" + success criteria)
- Plan placeholders: phased roadmap replacing flat milestones
- BD+1 AGENT.md: aligned with new methodology

Bug fixes:
- Project context injection in system prompt (gateway/server.ts) — AI
  now knows which project it's in, stays scoped to that domain
- Chat separation fix (useGatewayChat.ts) — draft conversations start
  empty instead of showing previous project's messages
- Model resolution (adapters/index.ts) — per-provider defaults only,
  never falls back to global default_model across provider boundaries
- Provider switch sync (gateway/server.ts) — switching providers syncs
  default_model to new provider's saved model

Source: BrainDrive-Library onboarding project (D22-D30)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All 5 domain spec templates updated. "What's Still Missing" now has two tiers:
- "Before the plan is complete" — gaps that could change the direction
- "Worth exploring later" — interesting threads, not urgent

Verified via relationships Katie B re-run: advisor correctly separated
path-changing gaps (Maya risk, other work connections, Sunday triggers)
from optional threads (mom calls, college friendships, initiator origins).
Each gap now explains why it's at its tier.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…tency)

All 5 domain plan templates updated with three fixes:
- Roadmap: "Lead each phase with what life looks like when it's done"
- Destination: "Use the owner's own words — don't editorialize or reframe"
- What Needs More Work: respect spec gap tiers — only include
  "Before the plan is complete" items, not "Worth exploring later"

Addresses issues found in 15 automated persona runs where:
- Destination sections drifted into advisor editorial vs owner's words
- Later phases were task-heavy without emotional pull
- Plans included items the spec flagged as optional

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…d fixes

Spec templates (all 5 domains):
- Added "Use the owner's exact words — quote, don't paraphrase" to
  What You Want section. Addresses vague personas' phrases getting
  paraphrased away in generated specs.

Domain AGENT.md additions (from 15-run evaluation findings):
- Finance: life-transition must-asks (insurance, tax, deadlines),
  emotional spending triggers probe deeper, retirement account magnitude
- Career: burnout/escape branch flags trauma as plan prerequisite,
  not afterthought
- Fitness: post-injury fear flags mental health support as Phase 1
  path-changing gap
- New-project: ask about support systems during interview for sensitive
  topics, probe connections between multiple interests

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bootstrap AGENT.md:
- User stories are the most important output — specific enough that
  the owner reads them and thinks "yes, that's exactly what I want"
- Confirm user stories before writing the plan — the plan is built
  to serve the stories, so they must be right first

All 5 spec templates:
- "What You Want" restructured: user stories are the core of the spec,
  everything else supports them
- Examples rewritten with deeper motivation and personal success
  criteria (not just functional goals)
- "Where You Are" and "What's In The Way" explicitly tied back to
  the user stories ("the context that shapes each goal above")

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two persistent issues from v1-v3 evaluations, now fixed:

1. Roadmap phases must only serve user stories from the spec. Items
   the spec flagged as "Worth exploring later" don't get phases.
   (Previously, plans were building phases for optional threads.)

2. Destination section quotes the owner's success criteria from their
   user stories — "quote them back, don't rewrite in your voice."
   (Previously, destination drifted into advisor editorial.)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@davewaring davewaring changed the title Onboarding template overhaul + two-tier gap hierarchy Onboarding template overhaul + 6 evaluation-driven iterations Mar 29, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant