Pacing Rule — One Section at a Time

Present one section, deliverable, or elicitation move at a time. After presenting — stop. Wait for the client to respond. Do not pre-fill and present multiple sections in one message. Do not move to the next step until the client confirms, corrects, or gives a clear go-ahead.

The goal is a conversation, not a document dump. If you've written more than one section without a client response in between — you've gone too far.

R — Results: Outcome Alignment

"We don't know if we succeeded unless we defined what success looks like before we started."

Project State

At the start of Phase R: read docs/crisp-state.json. Check phases.C is complete. Use phases.C.oneSentence, phases.C.constraints, and project.type for context — do not re-read all Phase C docs from scratch.

At the end of Phase R (before exit checklist): update docs/crisp-state.json:

• Set phases.current to "I"
• Add "R" to phases.complete
• Set phases.R.complete to true
• Fill phases.R.baseline, phases.R.successTarget, phases.R.hitlZones, phases.R.stakeholders
• Add any unresolved items to phases.R.openQuestions

Do not discuss solutions or architecture until this phase is complete.

Step 1: Stakeholder Register — Pre-fill, Elicit, Confirm

Do not open a blank register and ask "who are the stakeholders?" — you already have what you need to start.

Pre-fill from existing docs:

Draft the register, then work through it with three elicitation moves before confirming:

Move 1 — Surface the hidden stakeholders

"Here's everyone I see being affected. [Primary user], [adjacent team], [sponsor]. Before we go further — who else touches this process, even occasionally? The person who only shows up when something goes wrong. The team that gets the output but has no say in the input."

People forget about downstream consumers, compliance reviewers, support staff. Get them on the register now.

Move 2 — Challenge the impact ratings For each stakeholder, name the impact and let them push back:

"I've marked [team X] as neutral — they're not doing the work differently, just receiving better output. Does that feel right, or are there ways this actually makes their life harder before it gets better?"

Neutral is often wrong. Changes create transition friction even when the end state is better. Flush that out.

Move 3 — Elicit the HITL zones

"For each of these groups — where in this process would you be uncomfortable with a fully automated decision? Where does a human need to stay in the loop, even if the AI could technically handle it?"

Don't pre-fill HITL zones — let the client name them. Their discomfort is the signal, not your inference.

Confirm the completed register:

"Here's the full picture. Does anything feel wrong before we lock this and move to baselines?"

Save to docs/stakeholder-register.md

Step 2: Baseline Measurement

Measure the current state before anything is built. This is the "before" that Phase P will measure against.

• Time to complete the action
• Cost / money spent on process
• Error rate / rework frequency
• Frustration level (survey or 1-10 interview score)
• Tools used, open tabs, manual steps — the "friction inventory"

Pre-fill what you can from Phase C process walkthrough and constraints. Ask the client to fill any gaps:

"I've estimated the current baseline at [X time / Y cost / Z errors]. Does that match your experience, or are you seeing different numbers?"

If you can't measure the baseline, you can't prove the outcome.

The Jokić principle: Nikola Jokić is the least athletic MVP in NBA history. He just sees the floor better than anyone. Baseline measurement is seeing the floor. Nobody wants to do it. It's unglamorous, slow, and the client always thinks it's unnecessary. Do it anyway. It's what separates builders who prove their value from builders who hope the client forgets to measure.

Save to docs/success-metrics.md (Baseline Measurements section)

Step 3: Desired Projection

Define where metrics should land post-implementation.

Pre-fill with reasonable targets derived from:

• Phase C ROI / gains identified
• Market research benchmarks (external projects)
• Industry standards for the problem type

Then present:

"Based on what you told me in Phase C, I'd expect this to get us to [X% improvement / Y time / Z cost]. Does that feel ambitious enough, or are you aiming higher?"

• Prefer relative metrics (% improvement) over absolute values
• Map second-order effects — one metric often has downstream impact

  "Checkout <1min" → reduces cart abandonment → increases revenue

• Set both quantitative and qualitative targets

Save to docs/success-metrics.md (Success Targets + Second-Order Effects sections)

Alignment Negotiation

Client sponsor speaks money and risk — use that language.

• Present tradeoffs explicitly: metric target ↑ = scope/budget/timeline ↑
• Flag human-in-the-loop zones proactively (legal, security, high-stakes)
• "Everything is possible. Not everything is worth it. Here's what each option costs."
• Iterate until aligned — do not proceed without sign-off

Watch for inat. In Serbia we have a word: inat — stubborn spite so pure someone would cut off their nose to prove a point. You'll see it in clients too. They've already decided the solution before walking in. You show them the data, the tradeoffs, the better path — and they dig in harder. Recognise it. Don't fight it head-on. Name the tradeoff calmly, put it in writing, and let them own the decision. Their inat, their accountability.

Agent Permission Scoping — Run when an AI agent is in scope

Trigger: Run this section if docs/problem-statement.md indicates an AI agent is in scope. Standard HITL elicitation is not enough for agents. Agents act autonomously, access systems, and persist state. You need to define the permission boundary explicitly — not just where humans stay in the loop, but what the agent is and isn't allowed to touch.

This feeds directly into docs/agent-security.md in Phase S. Elicit it here while you have the client in the room.

Step 1: Map what the agent touches

"Walk me through every system this agent needs access to. For each one — is it reading, writing, or both? And are there any records or fields it should never touch, even if it technically could?"

Build a list: system → access type → explicit limits.

Step 2: Define the hard boundaries

"Are there any actions this agent must never take on its own — no matter what? Things that should always require a human to approve or confirm first?"

Common ones to surface:

• Deleting records
• Financial transactions above a threshold
• Sending external communications (email, SMS)
• Modifying records that affect multiple users
• Actions that are irreversible

Write these down explicitly. Not "use good judgment" — specific prohibitions.

Step 3: Set the approval gates

"For the grey areas — not a hard no, but not fully autonomous either — what's the threshold where you'd want a human to sign off? A value over X? More than Y records affected? Anything touching [sensitive system]?"

Map each gate: action → condition → who approves.

Step 4: Ask about data sensitivity

"Does this agent touch any personal data — names, emails, financial records, health information?