R&D AI Helper — Newcomer Training (First Week)

R&D AI Helper — Newcomer Training

Your first week, day by day · one task, start to finish · what to expect on the score
Preset v0.4 ~45 min read Do-along
Welcome, Priya

By Friday you'll have signed a task, done it, and seen your first score.

This guide walks a new R&D engineer named Priya (joining the Kbot-Link team this week) through her first working task, end to end. You do the same steps in your workspace as you read. Nothing here is a demo — every screen shown is what actually happens.

What you'll do this week

  1. Before Day 1 · What the AI Helper is (and isn't)3 min
  2. Day 1 · T0 · Grant access so the AI Helper can measure fairly30 min
  3. Day 2 · Your first task · three-way agreement on required hours1 hour, includes 15 min discussion
  4. Day 2–4 · Doing the work · how "actual hours" are countedruns while you work
  5. Day 4 · Marking done · reading your delta5 min
  6. Day 5 · Your first weekly diary + score10 min
  7. When things go sideways · 5 common situationsreference
  8. One-page cheatsheetprint + pin
0

Before Day 1 · What the AI Helper is (and isn't)

3 min read

The AI Helper is not a boss and not a stopwatch. It's a scheduling teammate whose job is to make sure the number you're measured on is fair to you.

What it does

Reads the ticket you're about to work on, looks at similar past tasks and the code paths involved, and proposes a required-hours number — the honest amount of time this task should take. You either accept or counter. Once you and your manager sign, the number is locked and the clock is fair.

What it doesn't do

It doesn't watch your screen. It doesn't count keystrokes. It doesn't read your DMs. It doesn't compare you to other engineers behind your back. Your score is your actual hours ÷ the number you signed for — that's it.

You'll see the phrase "three-way agreement" everywhere. It just means the AI, you, and your manager all sign the hours before work starts. If any of the three doesn't sign, the task can't move to "In progress".
1

Day 1 · T0 · Grant access

Monday morning · ~30 min

Nothing else starts until you finish T0. The AI Helper can't propose fair hours if it can't see your work signals — so we spend 30 minutes granting scoped, revocable access to the things it needs to read.

1 Open the T0 wizard

From your Workforce workspace, click Start T0 access in the pinned banner. You'll see six sub-tasks. Each is one grant.

Workforce › Kbot-Link › OnboardingPriya S.
Task
T0 · Grant AI Helper access to your work signals
Estimated
3h · Not started
Subtasks (6)
0.1 AI subscription · 0.2 Repo · 0.3 Task board · 0.4 CI · 0.5 R&D Zoom · 0.6 Data-use scope

2 Grant access to your AI subscription (0.1)

Paste a read-only token from your own Claude / OpenAI / Gemini account. This lets the AI Helper see your past AI sessions on the code you're about to work on — so it can propose hours from your actual patterns, not a generic estimate.

!
Read-only. If you rotate the token later, access ends. If you leave the company, all grants revoke automatically.

3 Grant scoped repo access (0.2)

Sign in with your own GitHub through the Claude Code web session. You'll authorize read on src/public/kbot-link and the OTA routes — nothing else in the Kbot repo. You never hand over anyone else's GitHub credentials.

Scope requested
Read PRs, commits, review threads, CODEOWNERS within src/public/kbot-link and src/api/routes/ota.routes.ts
Not requested
Write · admin · other paths in the repo · other repos

4 Grant task board, CI, R&D Zoom read (0.3–0.5)

Three more one-click grants: GitHub Issues + Projects filtered by kbot-link label, GH Actions runs touching your paths, and read on the R&D Zoom channel only (DMs and other channels stay out of scope).

5 Sign the data-use scope (0.6)

One page. Says exactly what the AI Helper reads, stores, and never touches. You sign; your manager (in your case, the Kbot-Link lead) counter-signs. You can revoke any grant any time — the manager gets notified but can't block the revoke.

Data use — one page
Reads: scoped repo, board, CI, R&D channel, your own AI history. Stores: task events, agreed hours, actual hours, delta. Never touches: personal DMs, non-work repos, other channels, other engineers' AI history. Revocable any time.
When all six sub-tasks are done, T0 flips to Done and Workforce unlocks your first work task. If your lead hasn't counter-signed within 4 hours, the AI Helper pings them in R&D Zoom — you don't have to chase.
2

Day 2 · Your first task · three-way agreement

Tuesday · ~1 hour incl. 15 min discussion

The Kbot-Link lead assigns you KL-152 — "Add tenant filter to OTA release picker". Before you write any code, you'll agree the hours with the AI and your lead.

1 Open the task · read the AI's proposal

Open KL-152 in Workforce. The three-way agreement panel is already populated with the AI's first proposal.

◇ AI Helper
5h
Similar to KL-142 (rollout picker, 4h) + one extra dropdown state. Two files.
◆ You
Accept, or counter with a reason.
■ Lead
Approves after you and the AI converge.

2 Counter if you disagree — with a reason

You look at the code and notice the tenant filter has to talk to a permissions check the AI didn't reference. You counter with 7h and one sentence of reason. This is a normal round — most tasks take 1 or 2 rounds to converge.

Your counter (hours)
7h
Reason (one line)
Filter needs the per-tenant permissions check — extra service call + tests.

3 AI acknowledges · you converge

The AI reads your reason, agrees the permissions check is real, and updates its proposal to 7h. That's convergence.

◇ AI
7h
Agreed. Permissions check was missed on the first pass.
◆ You
7h
Signed 10:14 IST.
■ Lead
Notified. Approves within 4h.

4 Lead approves · task can move to In progress

Your lead sees the agreed 7h and one line of reasoning. They approve in 15 seconds. The task auto-flips to In progress and the "actual hours" timer starts tracking.

If you and the AI can't converge in 2 rounds, the discussion escalates to your lead — not you vs the AI. That is the point of the three-way flow: no side wins alone.
3

Day 2–4 · Doing the work · how actual hours are counted

Runs quietly while you work

You don't press "start" and "stop". Actual hours are inferred from two signals you already produce.

1 Commit-authoring windows

Every time you commit code that references KL-152 (in the branch name, PR link, or commit message), the AI Helper looks at the timestamps and infers active work windows. Two commits 20 minutes apart on the same branch = 20 minutes of work.

2 Task-timer (only if you want it)

If you prefer to be explicit — some engineers do — hit T to start a timer on the open task and T again to pause. It's additive: the timer sums with the commit windows, deduped.

3 Watch actual grow next to required

The task detail shows required and actual side by side, live. You can see whether you're on plan without opening a dashboard.

Required
7h
Actual (live)
4.5h
Delta
0.64
on plan · running under estimate
!
Chat DMs, browser tabs, and non-work commits are not counted. If you take an hour lunch mid-task, that hour isn't in your actual — the commit-window heuristic naturally skips it.
4

Day 4 · Marking done · reading your delta

Thursday · 5 min

Task shipped. PR merged. Time to read the score.

1 Move to Done

Either flip the state to Done yourself, or let the auto-close fire when the linked PR merges to main. Both do the same thing.

2 Read the delta

Actual came in at 6.5h against a required of 7h. Delta = 0.93. That's "on plan" — the score band is 0.9–1.2.

Required
7h
Actual
6.5h
Delta
0.93
on plan
✓ Done — signed

3 Understand what the delta means

Three bands, no punishment for being close:

Under 0.9
Good — but check
You beat the estimate. AI updates its baseline so future proposals are tighter.
0.9 – 1.2
On plan
Nothing to review. This is the target band.
Over 1.2
Retro
Not a red flag. Just a signal to look at the estimate or the execution together with your lead — learn, not punish.
5

Day 5 · Your first weekly diary + score

Friday · 10 min

Friday afternoon, your first weekly diary lands. It's yours — you see it before your lead does.

1 Open the weekly diary

Workforce sidebar → Diary → this week. The layout is identical to the daily view, aggregated across your 5 working days.

Priya S. · Kbot-Link · Week 38Sep 15 – Sep 19
Score
87
out of 100
Tasks
1 done · 1 in progress
Delta (median)
0.93
on plan
Active hours
28h
of 40 target

2 Read your score breakdown

Score = weighted sum across six categories (task-time, code quality, delivery, collaboration, context assets, reliability). Week 1 is mostly task-time and delivery — your context-assets and collaboration scores fill in as you work more with others.

3 No surprises for your lead

Your lead sees the same diary you see, at the same time. Nothing is hidden. If they want to talk about anything on it, they'll ask — and you'll both be reading the same page.

Week 1 score of 87 is a strong start. Most engineers land in the 75–90 band while their context-asset coverage catches up.
?

When things go sideways · 5 common situations

Reference · use as needed
1 · Task is bigger than we agreed

Open the task, click Re-estimate. AI proposes a new required-hours. New three-way sign starts. Original agreement stays in history — nothing hidden.

2 · Blocked by someone else

Flag Blocked with the name of the blocker. Clock keeps running on the actual, but the delta calc ignores blocked time. Your lead sees the block within minutes.

3 · You disagree with the AI two rounds in a row

Hit Escalate. Your lead joins the thread and the three of you land the number together. This is a normal path, not a failure.

4 · You forgot to hit T on the timer

Doesn't matter — commit-authoring windows will pick most of it up. If you want to correct, use Adjust actual with a note. Adjustments are visible to your lead.

5 · Task needs to split in two

Split creates a sibling task and asks the AI to propose hours for each half. The original required is retired; the two new signatures replace it.

Bonus · You want to revoke access

Settings → Grants → click Revoke on any T0 grant. The AI Helper stops reading that signal immediately. Your lead is notified.

One-page cheatsheet · print + pin

The whole system in 12 lines
Every task
  • AI proposes required hours
  • You accept, or counter with reason
  • Manager approves
  • Task moves to In progress
  • Actual hours track from commits + timer
  • Delta = actual ÷ required
Delta bands
  • < 0.9 — beat estimate (good)
  • 0.9 – 1.2 — on plan
  • > 1.2 — retro with lead
  • Blocked time — excluded from delta
Shortcuts
  • C — new task
  • T — start / pause timer
  • A — accept AI proposal
  • N — counter
  • G then D — go to Diary
  • ? — full shortcut list
Rules that don't change
  • Access is scoped and revocable
  • DMs, other channels, other repos — never read
  • You see your diary before your lead does
  • Escalation is normal, not a failure
  • All hour agreements are auditable
R&D AI Helper · Newcomer training v1 · walkthrough uses Priya S. (new hire, Kbot-Link team, Week 38) as the sample. Same shape works for any team on any Workforce subscriber. Questions → #rd-onboarding in Zoom.