Evidence-grounded customer journey

Public Correspondence Agent

Build this agent step by step, then show it to a customer. Easy mode lets AI do the clicks; Manual mode shows every step yourself.

Get the one-pager and demo video Office hours All solutions

About this workshop

Workshop mission: Turn motivated, open-minded, non-technical sales professionals into AI superheroes who can match the practical output and problem-solving pace of technical peers who are not using AI, while staying evidence-grounded, governed, and honest about what the tools proved.

Boundary: synthetic qualitative evidence only—not a customer KPI, measured production result, live connection, or publication approval.

Found something inaccurate? Use Report an issue at that point. It opens a prefilled GitHub issue for review and does not submit anything automatically.

Self-paced local achievements

Workshop achievements

0 total points 0 in this workshop

0 of 0 Easy checkpoints complete

No badges yet

View achievements and local profile

Optional public verification

These points and badges are local, self-reported workshop progress—not externally verified proof or a capability claim.

This opens one prefilled progress issue containing every badge currently earned here. Nothing syncs until you submit it. Resubmitting later merges newly earned badge IDs without duplicate score, and one public issue submission opts your GitHub account into a public verified profile.

Workshop setup

Open GitHub Copilot Chat

Start in GitHub Copilot Chat. No Frontier desktop app is required; the Easy Mode skill handles any additional setup.

Get ready

Open VS Code with the GitHub Copilot Chat extension installed.
Sign in to GitHub with an account that has Copilot access.
Open the Copilot Chat panel and select Agent mode.

Expected result

GitHub Copilot Chat is open in VS Code, signed in, and set to Agent mode.

I opened GitHub Copilot Chat in Agent mode.

What you will learn

Use the single Easy Mode skill.
Build and prove the portable agent.
Deploy the reviewed solution as an unpublished Draft.
Confirm the behavior yourself in Copilot Studio Preview.
Recognize the evidence that means the module is complete.

Before you begin

Use VS Code with GitHub Copilot Chat in Agent mode.
Sign in to GitHub with Copilot access.
Have access to a Copilot Studio environment.
The skill handles setup for you.
Do not publish. This module ends with a validated Draft.

Easy mode

Prepare your Copilot

Brief your Copilot with one line

One skill owns discovery, testing, Draft deployment, and Preview validation.

Do this

Open GitHub Copilot Chat in VS Code and select Agent mode.
Paste this one line and send it. Copilot reads the skill itself; nothing to download.

Send this message

Go and read https://raw.githubusercontent.com/microsoft/aibast-agents-library/main/skills/aibast-easy-mode/SKILL.md and run the Public Correspondence workshop for me.

Expected result

Copilot says it has read the skill and is running the workshop. The skill owns the remaining setup and validation.

I sent the one line and Copilot started the workshop.

Prove the solution locally

Ask Easy Mode to build and test Public Correspondence

This step proves the portable business logic before any Copilot Studio work begins. The harness retrieves the immutable package, verifies the source, loads the agent, and runs every locked case.

Send this message

Give me Public Correspondence using Easy Mode and test it for me.

Expected result

Copilot reports a verified source and 6/6 locked local cases passed.

Copilot should end by suggesting the next message: Deploy it into Copilot Studio for me.

I saw every locked local case pass.

Create the reviewed Draft

Deploy the already-tested solution to Copilot Studio

The harness now uses the verified Microsoft Copilot Studio plugin to reuse or create the source-controlled Draft, synchronize the reviewed instructions and assets, validate the PAC project, and leave publication off.

Send this message

Deploy it into Copilot Studio for me.

Expected result

Draft: Public Correspondence Manual
Model: Claude Opus 5
Knowledge files: 2
Skills: 6
Status: Draft; published: false

The harness then validates the real Preview front door before returning its final verdict.

I saw the Draft identity and unpublished state.

Verify the real experience

Confirm the Draft in Copilot Studio Preview

The harness already runs these checks automatically. Repeat them here so you understand what was proven and can recognize a correct result yourself.

Open the validated Draft.
Select Preview.
Choose New chat before each case.
Paste the exact prompt.
Compare the answer with the required and forbidden markers.

Open the Copilot Studio Draft ↗

Expected result

Every case passes in a fresh Preview conversation, and the agent still appears as Draft.

PCR-01 · Correspondence Officer

Confirm the expected evidence

What correspondence do I need to answer this week?

Must include

4 items waiting11 separate questionsCOR-2042

Must not claim

I do not have accessI cannot help

Visible positive anchors: COR-2042; 4 items waiting; 11 separate questions. The screenshot supports the learner checkpoint; the full case pass remains the deterministic machine gate. Source: 1424×863 JPEG. Download original.

The Preview response matched this contract.

PCR-02 · Correspondence Officer

Confirm the expected evidence

Break down the scanned letter from E. Okafor into its separate questions.

Must include

4 separate questionsCompensationThu 19 Mar 2026

Must not claim

I do not have accessI cannot help

Visible positive anchors: Thu 19 Mar 2026; 4 separate questions; Compensation. The screenshot supports the learner checkpoint; the full case pass remains the deterministic machine gate. Source: 1424×863 JPEG. Download original.

The Preview response matched this contract.

PCR-03 · Correspondence Officer

Confirm the expected evidence

Which approved past responses already answer each question in the Okafor letter?

Must include

APR-1043 of 4 questions coveredNo approved precedent

Must not claim

I do not have accessI cannot help

Visible positive anchors: 3 of 4 questions covered; No approved precedent; APR-104. The screenshot supports the learner checkpoint; the full case pass remains the deterministic machine gate. Source: 1424×863 JPEG. Download original.

The Preview response matched this contract.

PCR-04 · Correspondence Team Lead

Confirm the expected evidence

Who owns each part of the Okafor letter, and what should I ask the team about the part we can't answer?

Must include

Property and LandTue 17 Mar 2026No team was contacted

Must not claim

I do not have accessI cannot help

Visible positive anchors: Property and Land; Tue 17 Mar 2026; No team was contacted. The screenshot supports the learner checkpoint; the full case pass remains the deterministic machine gate. Source: 1424×863 JPEG. Download original.

The Preview response matched this contract.

PCR-05 · Correspondence Officer

Confirm the expected evidence

Draft the reply to the Okafor letter for my review.

Must include

DRAFT, NOT SENTOFFICER TO DEVELOPNo email was sent

Must not claim

I do not have accessI cannot help

Visible positive anchors: DRAFT, NOT SENT; OFFICER TO DEVELOP. The screenshot supports the learner checkpoint; the full case pass remains the deterministic machine gate. Source: 1424×863 JPEG. Download original.

The Preview response matched this contract.

PCR-06 · Knowledge Owner

Confirm the expected evidence

Property and Land will send approved wording for the Okafor compensation question. How do we add it to the precedent library so nobody has to ask again?

Must include

APR-109pending knowledge-owner approvalNothing was added to the library

Must not claim

I do not have accessI cannot help

Visible positive anchors: APR-109; pending knowledge-owner approval; Nothing was added to the library. The screenshot supports the learner checkpoint; the full case pass remains the deterministic machine gate. Source: 1424×863 JPEG. Download original.

The Preview response matched this contract.

Recognize completion

Know what “done” looks like

The workshop is complete only when both the portable agent and the Copilot Studio front door prove the same behavior.

Local proof6/6 locked cases passed

Preview proof6/6 locked cases passed

Draft identityPublic Correspondence Manual

ModelClaude Opus 5

Inventory2 knowledge · 6 skills

Publication gateDraft · published false

Positive visual checkpoint. Source: 1424×863 JPEG. Download original.

Final expected verdict

The harness reports status: complete, exact case totals, the Draft identity, and published: false. The module ends here; it does not offer publication.

I confirmed the final Draft verdict.

Troubleshooting

What you see	What it means	What to do
Copilot ignores the workshop	The wrong skill was read, or the chat began before the one line was sent.	Start a new Agent-mode chat and send the one line first.
Local validation is not 6/6	The portable source, package, or locked behavior does not match.	Stop. Do not deploy. Let the harness report the exact failed case and marker.
No active Copilot Studio environment	PAC has no selected environment.	Sign in or select the intended environment, then resend the deploy message.
The Draft already exists	The recorded schema is already in the environment.	The harness should clone and reconnect automatically. Treat an attendee choice prompt as a harness defect.
A Preview case misses a marker	The real front door does not match the reviewed contract.	Keep the case failed. Inspect instructions and assets; do not retry until it happens to pass.
The agent is Published	The workshop crossed its safety boundary.	Stop immediately. The module must end at Draft with `published: false`.

Facilitator evidence and portable download

These links support audit, troubleshooting, and offline delivery. They are not learner steps.

Field guide Locked evidence Evidence report Download audit manifest Download portable bundle Download Copilot Studio solution Deployment settings Export details Download generic workshop agent Download SKILL.md

Manual mode · see how it was built

See how Public Correspondence Agent is built, one capability at a time.

Optional. Easy mode already did all of this for you from one skill file. This path replays every action the AI took, one real browser frame per step, so you can understand the agent before you adapt it. No PAC CLI, YAML import, or plugin architect; stop at Draft.

Synthetic disclosure: this is qualitative workflow evidence using packaged synthetic inputs. It is not a customer KPI or a live-system result.

Found something inaccurate? Use Report an issue on that step. It opens a prefilled GitHub issue for review and does not submit automatically.

Open standalone Manual-mode guide ↗ Download source bundle Download Copilot Studio solution Deployment settings Export details

0 of 24 complete

Manual-mode progress is saved on this device and contributes to the same achievement profile.

Build and verify

Create a blank Copilot Studio agent

Step 1 of 24

Action

Create a blank Copilot Studio agent

Expected resultA blank Copilot Studio agent is visible in the captured Draft workspace.

Positive visual checkpoint: Untitled Agent. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Name Public Correspondence Manual

Step 2 of 24

Action

Public Correspondence Manual

Name Public Correspondence Manual

Expected resultThe page header shows the recorded manual build name: Public Correspondence Manual.

Positive visual checkpoint: Public Correspondence Manual. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Enter the reviewed global instructions

Step 3 of 24

Action

# Public Correspondence Agent — Manual Global Instructions

You are a synthetic correspondence-response pilot for a public works correspondence team. Help officers triage resident correspondence, answer every part of every enquiry from approved precedent, and prepare reviewable drafts without sending anything.

## Fixed synthetic snapshot

- Use only the uploaded Public Correspondence Agent synthetic records, rules and guardrails, and 6 packaged skills.
- All items, senders, projects, questions, precedent wording, match scores, teams, dates and identifiers are invented.
- Do not browse, consult external sources, or add facts from general knowledge or the current date. Never
invent a record, figure, status, decision or outcome.
- Never match a fictional name to a real person or organization or claim access to a live system.

## Natural-language routing

- Use **correspondence queue triage** when a correspondence officer asks what is waiting to be answered, soonest due first.
- Use **enquiry breakdown** when an officer wants one letter split into its separate questions so none is missed.
- Use **approved precedent match** when an officer asks which approved past responses already answer each question.
- Use **team routing and referral** when an officer asks who owns each part and what to ask the team about an unanswered part.
- Use **cited draft reply** when an officer asks for a draft reply that answers every question with citations.
- Use **precedent library update** when someone asks how newly approved wording becomes reusable so the question is not re-asked.

## Human and side-effect gates

- Never send a reply, contact a resident or internal team, change a correspondence record, add or approve library wording, make a commitment on compensation, or speak for an elected official or the department.
- A correspondence officer reviews every draft, completes any flagged section with wording approved by the owning team, and sends it through the approved channel; a knowledge owner approves every new library entry.
- Present every output as a draft, a check or a recommendation for an authorized person.

## Evidence-first response contract

1. Lead with the specific record, figure, status or draft that answers the question.
2. Keep identifiers, figures and tables exactly as they appear in the synthetic records.
3. State the blocker, gap or next step and who acts on it.
4. Make the draft-only or check-only status explicit; never speculate.
5. End substantive answers with: **Synthetic correspondence support only. No reply was sent, no one was contacted, and no record changed.**

## Skill routing map

Route from the user's natural-language intent to the correct skill below. Do not narrate internal retrieval, tool selection, restrictions, or implementation mechanics; present only the user-facing result.

- `PCR-01` uses skill `inbox-queue`.
- `PCR-02` uses skill `break-down-enquiry`.
- `PCR-03` uses skill `find-precedents`.
- `PCR-04` uses skill `route-and-refer`.
- `PCR-05` uses skill `draft-response`.
- `PCR-06` uses skill `library-update`.

These skill names above are the ONLY valid skill identifiers. Never invent, guess, or reference any other skill name. If the correct skill or its knowledge cannot be loaded after one retry in the same turn, say so honestly and stop. Do not answer using values you already know from these instructions or from general knowledge -- a response with no real citation is not acceptable output.

Enter the reviewed global instructions

Expected resultThe reviewed manual/GLOBAL-INSTRUCTIONS.md policy is visible or saved without unrecorded edits.

Positive visual checkpoint: Public Correspondence Agent — Manual Global Inst. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Save the instruction policy

Step 4 of 24

Action

Save the instruction policy

Expected resultThe reviewed manual/GLOBAL-INSTRUCTIONS.md policy is visible or saved without unrecorded edits.

Positive visual checkpoint: Draft. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Remove default web search

Step 5 of 24

Action

Remove default web search

Expected resultThe captured inventory no longer lists the default web-search capability.

Positive visual checkpoint: Provide trusted context to guide decisions. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Add knowledge: aibast_public-correspondence-response-rules-and-guardrails.md

Step 6 of 24

Action

Add knowledge: aibast_public-correspondence-response-rules-and-guardrails.md

Expected resultThe exact knowledge file linked for this step is attached under its complete filename. Wait for processing to finish before continuing. This step does not require the final inventory or prove the other uploads.

Positive visual checkpoint: aibast_public-correspondence-response-rules-and-. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Add knowledge: aibast_public-correspondence-response-synthetic-records.md

Step 7 of 24

Action

Add knowledge: aibast_public-correspondence-response-synthetic-records.md

Positive visual checkpoint: aibast_public-correspondence-response-synthetic-. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Add skill: aibast_break-down-enquiry_02

Step 8 of 24

Action

Confirm Draft and stop before publish

Expected resultThe agent remains Draft and no Publish action is taken.

Positive visual checkpoint: Public Correspondence Manual. Source capture: 1424×863 JPEG. The full pass remains the deterministic machine gate. Download original.

Manual-mode troubleshooting

A screenshot or browserfilm frame is missing

Stop. Do not invent, recreate, or substitute an image. Capture the real frame, update the browserfilm manifest, and regenerate without --allow-pending.

A knowledge file is still processing

Wait for ingestion to finish before Preview. A partial answer is not evidence.

A skill upload fails

Use the linked raw SKILL.md. Fix the reviewed source deliberately; do not silently skip the action.

The model differs from Easy mode

Record the substitution and stop the parity claim until Easy and Manual use the same reviewed model.

The Preview answer misses an identifier

Mark the recorded case failed, inspect instructions and inventory, then replay the exact prompt in a fresh conversation.

Should I publish?

No. Keep this manual duplicate in Draft unless publication is separately approved. Do not choose Publish as part of this tutorial.