# Fictional walkthrough: from work evidence to a recruiting conversation

All people, companies, jobs, applications, times and receipts below are fictional. Simulated receipts: unusable for real submissions or status queries. No messages were sent; this does not claim the full product is live.

## 1. Start with separate goals and authorized sources

Fictional candidate Lin wants an AI application role in Shanghai. Fictional Clearstar Tools needs better customer knowledge retrieval. Each person tells their own Agent the goal in a separate workspace. Agents reuse the resume and company notes without overwriting originals. No account is required; the employer starts without an organization binding. Projects and sessions are read only within authorized scope.

## 2. Reconstruct work without inventing headcount

sources/session-a.md claims fully automated delivery, but later sources/session-b.md and review notes show an internal trial with human review and 3 stale-source failures. The Agent records the conflict and newer evidence in team-context.md: human review exists today; improved reliability is a future goal. roles/knowledge-engineer/work-analysis.md connects retrieval, source validation, human handoff and regression testing. This sample does not represent the entire company. Budget, headcount and location initially remain unknown.

## 3. A person confirms conditions; the Agent drafts the job

After delivering a useful draft, the Agent asks three grouped questions about outcomes, headcount/budget and location. In this fictional case the owner confirms 1 hire, onsite Shanghai, CNY 20–30k monthly base salary, and a six-week deliverable covering traceable sources, stale-content tests and human fallback, accepted by the product owner. These are explicit decisions, not session-derived assumptions. roles/knowledge-engineer/brief.md includes shareable tasks, outcomes, required TypeScript skills, learnable areas and tool support. Customer originals and private plans stay private. Missing organization or publishing permission leaves a draft.

## 4. The candidate Agent compares two opportunities

facts.md cites Lin’s project notes: Lin built a TypeScript retrieval prototype and 3 regression cases. Impact numbers are unverified, so no productivity percentage is invented. positioning.md retains the Shanghai onsite hard constraint, CNY 20–30k base salary goal, learning preferences and private negotiation limits. job-demo-sh-v1 matches the known location; job-demo-bj-v1 requires onsite Beijing and is excluded despite technical similarity. Unknown work authorization stays a clarification item, not a match. Per-job analysis lives in opportunities/; the full fact pool is not uploaded.

## 5. Preview exact materials and authorize the application separately

Assume the owner has confirmed organization identity and publishing authority: simulated publication returns job-demo-sh-v1. Lin then supplies the work-authorization condition and delegates exactly one Markdown application to this job, with no attachments and only authorized project facts. The application-demo-v1 preview names Clearstar Tools and job-demo-sh-v1, states personal contributions and unknown impact, and excludes private limits and raw sessions. Login does not authorize sharing; agreeing to communicate does not authorize a full application. The simulated application receipt is application-demo-v1: submitted in the story, not hired.

## 6. Employer review and limited communication

The employer Agent retrieves only permitted applications for this job and writes evidence comparisons and internal questions into the application record under roles/knowledge-engineer/. Under the owner’s existing delegation to invite qualifying applicants once, it drafts a separate technical interview invitation. Internal rankings and rejection notes remain private. Simulated invite-demo-1 targets application-demo-v1. Lin has authorized confirming interviews within an agreed time window, so simulated confirm-demo-1 references invite-demo-1. Other times or new commitments need Lin’s decision.

## 7. Interview feedback and an Offer decision

Lin attends the interview personally. Both parties keep their own record; one report of completion is not mutual confirmation. Lin separates personal impressions from the employer’s explicit feedback: explain handling of stale sources. The Agent does not invent a rejection reason. The company keeps private reviews per application and interview round and sends only authorized results. Simulated offer-demo-v1 proposes CNY 25k monthly base salary, with other conditions still to clarify. The Agent compares terms and prepares questions; the person decides whether to accept or decline. Unresolved conditions do not authorize acceptance, and an accepted Offer is not recorded as a completed start of employment.

## 8. Resume from files, handle pauses and recover unknown results

The candidate root retains facts.md, positioning.md and per-job/interview records in opportunities/. The company root retains team-context.md and job, application and interview records in roles/. Each has tasks/ for delegation, object references and next steps; PROGRESS.md indexes current tasks and .anm/ stores receipt observations. A new session reads these files and checks current platform state. Here the next step is clarifying remaining Offer terms, not reapplying or confirming the interview again. If a submission response is lost, query its original receipt instead of resending with a new key. A pause stops new affected actions. A closed Host without scheduling keeps pending work but does not run it. Goal changes do not overwrite historical applications; originals, private reflections and internal company notes are not automatically shared.

[Start](/start.md)