Why it exists

Three failure modes that get worse as AI-assisted development accelerates.

01 — Tribal knowledge failure

Domain context lives in one person's head and never gets captured. When that person is heads-down or unavailable, the team stalls. Developers ask the same domain questions every sprint. The PM becomes a human search engine.

02 — Documentation drift

Governing docs are written once and go stale the moment work begins. Nobody trusts them, so nobody reads them. The gap between what the docs say and what the code does grows every sprint.

03 — Bi-directional blindness

Code evolves without informing docs. Docs change without reaching developers. The two never meet. As AI-assisted development accelerates, this gap widens faster than any team can manually close it.

Where dovetell fits

dovetell is the context injection layer in your agent harness. If you're not the model, you're the harness.

The agent stack

Model          Claude / GPT-4o / Gemini / Local
────────────────────────────────────
Agent Harness
  Tools        Cursor / Claude Code / Copilot
  dovetell  ←  context injection layer  ★
  Memory       filesystem / AGENTS.md / git
────────────────────────────────────
Your Codebase

Harness-as-a-Service platforms (Cursor SDK, Claude Code, GitHub Copilot) give you three things: which model, which tools, which task. They don't give you domain context — what your team's definitions are, which decisions were made, what the architecture constraints are. That's dovetell's job.

Three phases of agent development

• Phase 1 — Weights: bigger models, more data, fine-tuning
• Phase 2 — Context: prompt engineering, RAG — dovetell starts here
• Phase 3 — Harness: what environment should the model operate in — dovetell ends here

Team AI Maturity Model

Four levels describing how systematically a team uses AI. Every team is somewhere on this curve. dovetell helps you move up it.

Six capability areas

Teams are scored across six dimensions, 5 questions each, 0–3 points per question. Max 90 points.

How it works

5 minutes. 30 questions. Personalized score, gaps, and next steps. A unique link to track your team's progress over time.

Scoring

How points are calculated and how levels are assigned.

The math

6 sections
× 5 questions per section
× max 3 points per question
= 90 points maximum

Unanswered questions → 0 points
Skipped sections → 0 points for that section

Level thresholds

Answer scale

Tracking over time

Your unique link is your project's permanent home. Retake whenever your team's practices change and watch the trajectory.

How the ID system works

Hierarchy

email
└── uid  (you, across all projects)
    ├── pid-a  (Ops Analytics Team)
    │   ├── aid-1  Jan 2026 · v01 · 14/90 · Scattered
    │   ├── aid-2  Mar 2026 · v01 · 32/90 · Structured
    │   └── aid-3  May 2026 · v01 · 58/90 · Coordinated
    └── pid-b  (Data Infrastructure Squad)
        └── aid-1  Mar 2026 · v01 · 21/90 · Scattered

Your unique link

After completing the assessment you receive a link like:

dovetell.io/assessments/?pid=8f3c2a1b

Bookmark it. When you return, you'll see your last score and a retake button. Each retake generates a new aid while keeping the same pid — your project's growth trajectory builds automatically.

Phase 2 overview

The dovetell platform is coming. The prompt library and assessment are Phase 1 — validation that the problem is real and people will pay to solve it.

Phase 2 is the SaaS platform that makes the prompt library automatic — a living, queryable context layer that your team doesn't have to maintain manually.

The unvetted → reviewed → becomes truth loop

Everything the platform captures starts as unvetted. A PM reviews, accepts, edits, or rejects. Accepted items become vetted — authoritative, queryable, and surfaced to the team automatically.

No automation without human review. The human in the loop is a feature, not a limitation.

Capabilities

Four capabilities, built in dependency order. The queue is the connective tissue — nothing becomes truth without it.

1 — Context Console

The dashboard. A context health monitor, not a document editor. The PM opens it at standup — sees the drift score, unvetted count, stale definitions, decisions captured this sprint. Red/Yellow/Green in under 30 seconds.

2 — Context Query (MCP)

dovetell as an MCP server. A developer types @dovetell what's our defect rate threshold inside Cursor or Claude Code and gets the team's answer — vetted, sourced, with decision trace. "I want to be asked fewer questions" made literal.

3 — Context Import

Drop in existing docs — markdown, Notion exports, Confluence exports. dovetell extracts decisions, assumptions, and definitions. Everything starts unvetted. First value in under an hour.

4 — Unvetted Queue

The human-in-the-loop gate. Every automated update lands here first. The PM accepts, edits, rejects, or defers. Accepted items become vetted truth. This is where tribal knowledge stops being tribal.

Definitions

Shared vocabulary. When a term is used in dovetell documentation or code, it means what's defined here.

Core concepts

context base — The living collection of decisions, assumptions, definitions, and policies that governs how a team's AI-assisted work should behave.

context drift — The gap that grows between what a team's documents say and what the code or work actually does.

context orchestration — Assembling, aligning, and delivering the right context to the right person or tool at the right time.

tribal knowledge failure — Domain context that lives in one person's head and never gets captured.

unvetted — Status for any item captured automatically but not yet reviewed by a human.

vetted — Status for any item reviewed and accepted by a designated reviewer. Authoritative.

decision trace — A record of what was decided, when, by whom, under which policy.

drift score — A metric (0–100) indicating how far a team's documented context has drifted from their actual work.

dog food loop — Using dovetell to manage the context of building dovetell. Active from day one.

ID definitions

uid (userId) — Permanent per browser. Generated once on first submission. Ties all projects for one person together.

pid (projectId) — One per project or team. Persists across retakes. Enables growth tracking.

aid (assessmentId) — Fresh per submission. Tracks individual runs within a project.

vid (versionId) — Which question set. Currently v01. Bumped when questions change.

Decisions log

Key decisions that shaped the product. One file, scannable, honest about tradeoffs.

Assessment architecture

Infrastructure

Privacy

Privacy

Plain English. No surprises.

What we collect

Required: email address only.

Optional: project name, role, team size, company, industry, primary AI tool, how you found us.

Automatically: assessment scores, section breakdown, raw answers, skip count, unique IDs (uid/pid/aid/vid), timestamp.

How we use it

• To send you your results and unique tracking link
• To enable progress tracking over time (via pid)
• The dovetell product team sees your scores to improve the product
• Occasional product updates — opt out any time by replying "unsubscribe"

What we don't do

• Never sell your data
• Never share with third parties or advertisers
• No cookies (Plausible analytics is cookieless)
• No tracking pixels

Project name disclaimer

The project name field shown in the assessment gate says: "⚠ Don't include sensitive, confidential, or proprietary information here." We mean it. dovetell is not a secure document store.