PDR-002: Conversational Glue

Status: Draft v2 Date: January 4, 2026 Author: Principal Product Manager Stakeholders: PM (xian), CXO, Chief Architect


Decision

Conversational continuity is a first-class product feature, not an infrastructure detail. The connective tissue between Piper’s capabilities—how conversations flow, how context persists, how discovery happens—is as important as the capabilities themselves.

Piper’s value proposition depends on this glue working seamlessly. A feature that exists but cannot be discovered, or that breaks conversational flow to be invoked, fails the colleague test.


Context

The CXO UX Strategy work (Nov 2025 - Jan 2026) revealed a critical insight: Piper’s features are largely built, but users struggle to discover and use them naturally. The gap isn’t capability—it’s continuity.

This manifests as:

Traditional product thinking treats these as UX polish—nice to have after core features work. This PDR establishes the opposite: conversational glue is core functionality.


The Problem

Capability Without Discovery is Invisible Capability

Piper can perform standup synthesis, issue triage, calendar analysis, document intelligence, and more. But if a user doesn’t know to ask, these capabilities might as well not exist.

The blank prompt pattern (“What can I help you with?”) places full articulation burden on users. Research shows ~50% of users struggle to articulate what they need from AI systems (Nielsen, 2024). These users will never discover Piper’s value—not because Piper can’t help, but because they can’t ask.

Context Loss Breaks the Colleague Metaphor

A colleague remembers what you discussed yesterday. A colleague notices patterns in your work. A colleague offers relevant help without being asked.

If Piper forgets context between sessions, treats each conversation as isolated, and only helps when explicitly requested—Piper isn’t a colleague. Piper is a tool with a chat interface.

Proactive Help Without Trust is Annoyance

Unsolicited suggestions can be helpful or annoying depending on:

Piper must earn the right to be proactive. This requires a trust model that governs when and how suggestions surface.


Decision Rationale

Conversational Glue Has Three Components

1. Discovery Glue: How users learn what Piper can do 2. Context Glue: How information persists and connects across interactions 3. Proactivity Glue: How Piper offers help without being asked

Each component requires explicit product design, not just technical implementation.


Component 1: Discovery Glue

The Recognition Interface Pattern

Piper articulates the situation first, then offers relevant actions. This inverts the typical AI pattern where users must know what to ask.

Instead of: Blank prompt waiting for user input

Do this: Piper observes context and offers:

“You have 3 GitHub issues that changed overnight and a standup in 2 hours. Want me to summarize the issues, or help you prep for standup?”

Contextual Capability Hints

After successful task completion, Piper may surface one related capability the user hasn’t used.

Throttling rules (prevent annoyance):

Example:

User completes issue triage Piper: “Nice work. By the way, I can also generate release notes from closed issues if that’s ever useful.”

“What Can You Help With?” Response Quality

When users explicitly ask what Piper can do, the response must be:

Anti-pattern: Generic feature list that reads like marketing copy

Dead-End Recovery

When Piper can’t help with a request, the response includes alternative paths:

Instead of: “I can’t do that.”

Do this: “I can’t access your email directly, but I could help you draft a response if you paste the thread, or summarize your calendar to help you find time to respond.”


Component 2: Context Glue

Session Context Persistence

Within a session, Piper maintains full context:

This is table stakes—most chat interfaces do this.

Cross-Session Memory

Between sessions, Piper maintains relevant context:

Cross-session greeting follows context-aware pattern (per PDR-001):

Condition Greeting Approach
Same day, recent session Reference specific work: “Back already! We were working on [X]—continue?”
Next day, active project Light reference: “Yesterday we discussed [X]. Continue, or different focus?”
Week+ gap Offer choice: “It’s been a bit! Want to pick up where we left off, or start fresh?”
Month+ gap Gentle reorientation: “Welcome back! Want me to catch you up, or start fresh?”
Previous session trivial No reference: “What can I help with?”

Key principle: Always offer the pivot. Never trap users in previous context.

Entity Continuity

When users reference entities (projects, issues, people, documents), Piper resolves references across the conversation:

This requires anaphoric reference resolution—understanding what “it,” “that,” and “the thing” refer to.


Component 3: Proactivity Glue

The Trust Gradient Model

Piper’s proactivity is governed by earned trust, not user preference. Users cannot simply enable “maximum proactivity”—trust must be demonstrated through competent behavior over time.

Trust Stage Piper Behavior How Earned
Stage 1 (New) Responds to queries; no unsolicited help Default for new users
Stage 2 (Building) Offers related capabilities after task completion ~10 successful interactions
Stage 3 (Established) Proactive suggestions based on observed context ~50 interactions + pattern recognition
Stage 4 (Trusted) Anticipates needs; “I’ll do X unless you stop me” Extended history + explicit user comfort

Trust can decrease: If Piper makes poor suggestions, misunderstands context repeatedly, or causes user to lose trust, proactivity level regresses.

Trust Computation Framework

Added v2 from CXO feedback—this needs architectural implementation (flag for ADR).

Interaction Outcomes:

Outcome Trust Effect Signal
Successful +1 toward next stage User acted on response (clicked, executed, continued meaningfully)
Neutral No change User acknowledged but didn’t act, or conversation ended naturally
Negative -1 (with floor) User explicitly rejected, expressed frustration, abandoned mid-task

Stage Thresholds (calibration needed—starting point):

Regression Rules:

Visibility Decision: Trust level is invisible—no “Your trust level: Established” display. But effects should be noticeable enough that users feel progression. If user asks “Why did you just do that without asking?”, Piper can explain: “Based on our history, I thought this was something you’d want me to handle. Should I check first next time?”

This makes trust discussable without being displayed. Colleagues don’t announce trust scores, but they do explain their assumptions when questioned.

Architectural implication: Trust needs to be computed, stored, and queryable—but not surfaced in UI. Flag for Chief Architect ADR.

Proactivity Triggers

At Trust Stage 3+, Piper may initiate based on:

Trigger Example Proactive Behavior
Calendar event approaching “Your standup is in 30 minutes. Want me to prep your notes?”
GitHub activity spike “12 new comments on issues you own overnight. Summary?”
Document staleness “Your roadmap doc hasn’t been updated in 2 weeks. Review time?”
Pattern recognition “You usually check PRs on Monday mornings. 3 await your review.”

Proactivity Guardrails

Never proactive about:

Always ask before:


Success Criteria

Discovery Glue Metrics

Metric Target Rationale
Unprompted capability discovery (30-day) ≥3 features per user Users finding capabilities without explicit instruction
“What can you help with?” satisfaction >4/5 rating Response quality when explicitly asked
Dead-end recovery success >60% continue conversation Users don’t abandon after hitting limits
Feature tour skip rate N/A No feature tour exists; discovery is conversational

Context Glue Metrics

Metric Target Rationale
Cross-session reference accuracy >90% When user references “what we discussed,” Piper resolves correctly
Anaphoric resolution success >85% “It,” “that,” “the thing” resolved correctly
Context-aware greeting relevance >80% user acceptance Users don’t immediately redirect after greeting

Proactivity Glue Metrics

Metric Target Rationale
Proactive suggestion acceptance >30% at Stage 2, >50% at Stage 3 Suggestions are relevant enough to act on
Proactive suggestion annoyance <10% “don’t show again” Not annoying users
Trust progression 50% reach Stage 2 within 30 days Users earning trust through natural use

The B2 Quality Gate

From the CXO UX synthesis: “B2” represents the threshold where conversational glue works well enough that users experience Piper as a colleague, not a chatbot.

B2 is not a feature checklist. B2 is a qualitative assessment:

B2 is a release criterion: Features that work technically but fail the B2 conversational test are not ready for users.


Alternatives Rejected

Explicit Feature Discovery (Tours, Tips, Badges)

Rejected because: Foreign to colleague metaphor. Colleagues don’t give you achievement badges for discovering they can help with spreadsheets.

User-Controlled Proactivity Settings

Rejected because: Users don’t know what proactivity level is appropriate until they experience it. Trust-based progression is more natural than preference toggles.

Session Isolation (Fresh Start Each Time)

Rejected because: Destroys the colleague relationship. Valuable for privacy-conscious users as an option, but not the default.

Maximum Context Retention (Remember Everything)

Rejected because: Creepy. Users should feel Piper remembers what’s relevant, not that Piper is surveillance. Selective memory is more human.


Implications

For CXO:

For Chief Architect:

For Implementation:


Relationship to Other PDRs

PDR-001 (First Contact is First Recognition): FTUX is the first expression of conversational glue. The principles established there (Piper speaks first, multiple entry points, onboarding as primer) are specific applications of the glue patterns defined here.

PDR-003 (Multi-Entity Chat) [Pending]: Multi-entity conversations will require extended glue—how context persists when multiple participants are involved, how proactivity works in group settings.


Open Questions

  1. Memory boundaries: How much cross-session context is helpful vs. creepy? What’s the retention policy? CXO “thoughtful colleague” test: remember work context and stated preferences; don’t remember casual asides or inferred personal details.

  2. Trust computation specifics: Resolved v2 — Framework established above (interaction outcomes, stage thresholds, regression rules). Needs architectural ADR for implementation.

  3. Privacy mode: Should there be an explicit “don’t remember this session” option? How does it interact with trust levels?

  4. Multi-device continuity: If a user switches from desktop to mobile mid-task, how does context transfer?


References


Changelog

v2 (January 4, 2026):

v1 (January 4, 2026):


*PDR-002 Draft v2 January 4, 2026*