The destination, the full set of questions Agent Karma will help developers answer, and how we sequence toward it without losing the validation wedge. Pairs with
product-strategy.md(positioning) andcompetitive-coverage.md(what we deliberately exclude).
Agent Karma's long-term aim is to be the trusted, local-first companion that helps an individual developer become genuinely self-aware and more effective across the entire arc of AI-assisted work — from how they ask, to what changed, to whether they validated it, to whether they're actually growing as an engineer.
From "Did I validate this?" (today) to "Am I becoming a more deliberate, capable AI-assisted engineer?" (destination).
It is holistic in the questions it helps you reflect on — and uncompromising in how it does so: private, coaching, self-comparative, never surveillance, never usage-policing.
Broad vision and a narrow MVP are not in conflict — they are different layers:
- The wedge (entry point): the one question we nail first, measure rigorously, and are known for → validation ("did you verify the AI's output?"). Nobody else owns it.
- The vision (destination): the full self-awareness arc across six layers (below).
The sequence: enter sharp → earn trust + adoption + data → expand deliberately, one layer at a time, never diluting the ethos.
The failure mode we refuse: delivering the whole breadth at once. That yields no focus, no moat, and a head-on fight with incumbents on their turf (see competitive-coverage.md). Breadth is the roadmap, not the MVP.
Rigor: 🟢 measured (objective signal) · 🟡 reflected (light/qualitative) · ⚪ aspirational (not built yet) Horizon: Now = shipped in MVP 0.1–0.7 · Next = near-term post-MVP · Later = vision
| Layer | Question | Horizon | How answered | Rigor |
|---|---|---|---|---|
| Intent | Did I ask the AI clearly? | Now | Prompt hygiene hint | 🟡 |
| Did I provide enough context? | Now | Dharma Card · context provided | 🟡 | |
| Did I set the right scope/intent? | Now | Intent capture (Dharma) | 🟡 | |
| Is my prompting improving over time? | Later | Prompt-hint trend across sessions | ⚪ | |
| Action | What files changed? | Now | File capture + git diff summary | 🟢 |
| What did this session consist of? | Now | Karma Trace timeline | 🟢 | |
| What did I do vs. the AI? | Later | Deliberately limited — no AI-vs-human attribution (no surveillance) | ⚪ | |
| Validation (the wedge) | Did I verify the AI's output? | Now | Objective Karma Score | 🟢 |
| Did I run tests / build / lint? | Now | Validation capture | 🟢 | |
| Is there test coverage for the change? | Now | Coverage component of the score | 🟢 | |
| Outcome | Is it ready to commit / review? | Now | Phal outcome | 🟢 |
| Did the result match my intent? | Now | Optional unscored reflection | 🟡 | |
| Did I commit AI code without validating? | Now | Opt-in pre-commit nudge | 🟢 | |
| Was my effort proportional to the risk? | Next | Dharma risk ✕ validation cross-check | 🟡→🟢 | |
| Effectiveness | Was this session effective? | Next | Session summary (validated + outcome-ready) | 🟡 |
| Where does AI help vs. slow me down? | Later | Per-tool / per-task-type outcome trends | ⚪ | |
| Growth | Am I getting better, or just more dependent? | Now (partial) → Later | Self-comparative score trend (now) → growth narrative (later) | 🟡 |
| Am I learning, or copy-pasting? | Later | Cross-session patterns (e.g. validation declining, risk rising) | ⚪ | |
| What's one thing to improve this week? | Next | Weekly reflection (one plain-language nudge) | ⚪→🟡 |
Read it this way: the MVP already touches every layer, but answers validation rigorously (🟢) and the rest lightly (🟡) or not yet (⚪). That is the sequencing — not the ceiling.
- NOW — MVP (0.1–0.7, shipped): intent capture, action capture, the validation wedge (rigorous), outcome/Phal, the opt-in pre-commit nudge, and a self-comparative score trend.
- NEXT — near-term post-MVP: weekly reflection (one true, useful thing from your own history); effort-vs-risk cross-check; session-effectiveness summary; per-tool / per-task-type trends; richer dashboard charts. (Maps to
roadmap.mdpost-MVP phases 1–4.) - LATER — vision: the longitudinal growth narrative ("am I improving or just more dependent?"); prompting-improvement trend; learning-vs-dependence signals; an optional local-LLM prompt coach (still on-device). (Maps to
roadmap.mdphases 5–10.)
Breadth never means abandoning the ethos. As we expand, each new question must still pass these:
- Local-first, no cloud, no login, no telemetry, no surveillance. No keystroke/scroll tracking, no AI-vs-human line attribution.
- No usage-volume metrics. We never add acceptance rates / lines-generated / tokens — the question is always "did I use AI well," never "how much."
- Coaching, never judgment. Self-comparative only; no leaderboards.
- Reflection before measurement. Add a new question as reflection (🟡) first; promote it to a measured signal (🟢) only when there's an honest, objective basis — never invent a fake metric (the prompt hygiene hint stays a hint, never a headline).
- The wedge test (from
competitive-coverage.md): does it help a developer reflect & improve, without surveillance, cloud, or usage-tracking? If not, it stays out.
A developer, after a few months with Agent Karma:
"I trust my AI-assisted code more, because Agent Karma turned validation into a habit — and I can actually see that I've gotten better at directing AI."
Validation is the door. Deliberate, self-aware AI-assisted engineering is the room. We lead with the door because it's the one only we can open — but the room is the vision, and the Question Map above is how we furnish it, one honest signal at a time.