Skip to content

About

Reusable Workspace Agent setup for evidence-backed project coordination: instructions, seven skills, templates, and fictional examples.

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Agentic TPM Starter for ChatGPT Workspace Agents

A reusable setup for a ChatGPT Workspace Agent that helps a TPM or project lead onboard projects, reconcile status, prepare reports, coordinate requested follow-ups, and assess closure. Project knowledge stays in durable external records so the agent can be rebuilt without moving the project's history.

Platform: Agentic TPM in this repository is a ChatGPT Workspace Agent, configured and used in ChatGPT. Codex helps prepare, review, and package the setup files; the hosted agent runs in ChatGPT with its own conversations and configured connections.

Version: 0.2.2. Status: portable starter, with fictional examples. This repository contains declarative instructions, seven skills, reference files, templates, and local packaging tools. It is not a hosted runtime, one-click deployment service, or claim of production qualification.

Why use a shared ChatGPT Workspace Agent?

Agentic TPM packages a repeatable operating process for work that requires multiple people to coordinate against maintained commitments and shared records. Each person keeps their own working context; the agent applies common methods to the same project sources. Human owners retain decisions and accountability.

The useful dividing line is shared continuity and accountability, not how long a task lasts. A three-day launch across teams may need this process; a months-long personal investigation may not.

Work or use case Suggested working mode Why
Explore an idea, prototype code, compare approaches, or draft a one-off analysis Direct Codex or ChatGPT use Flexible scope with little shared coordination overhead
Maintain personal research, notes, preferences, or role context Your private ChatGPT Project or Codex setup Keeps your individual perspective under your control
Run a cross-team launch or recurring delivery review ChatGPT Workspace Agent with registered project sources Consistent status, evidence, commitments, and follow-through across stakeholders
Take over an in-flight project or return in a fresh conversation ChatGPT Workspace Agent Recovers project context from verified external records
Prepare sponsor, TPM, and engineering views of one program ChatGPT Workspace Agent Different emphasis grounded in compatible source facts
Assess handoff or closure ChatGPT Workspace Agent Checks acceptance evidence and remaining obligations beyond a Done label

“Direct” describes the working pattern. It does not imply ChatGPT runs locally, or that Codex cannot support durable shared work. The same skills, connectors, and records can be used directly; the dedicated agent's value is packaging and sharing that setup.

Intended benefit What provides it How to check that it helps
Less repeated project explanation Registry, source mappings, and verified continuity records A fresh thread recovers the correct project with little guidance
Consistent project understanding Common source ownership and reconciliation rules Two authorized users receive compatible facts for the same source revisions
Repeatable reporting and handoffs Seven reusable skills and shared templates Results remain useful when someone other than the author runs them
More reliable follow-through Destination readback, action identity, and retry reconciliation Controlled repeat requests do not create duplicate actions
Reuse across projects and agent rebuilds Separate methods, private deployment bindings, and external project records Another project or replacement agent can reuse the methods without copying private context

These are design benefits to validate, not exclusive platform capabilities or established production guarantees. Measure total effort including setup, review, correction, and maintenance. More generated reports alone do not demonstrate value. Proactive follow-through and scheduled operation require separate runtime evidence.

Maturity levels and current position

This is a working model for Agentic TPM, not an official platform rating. Assess each deployed workflow separately: reporting can be mature while follow-up remains draft-only. Levels describe demonstrated independence and reliability, not permission tiers; they never grant access or bypass native controls.

Level Role and trigger What the agent does independently What demonstrates this level
1 — Reactive TPM assistant A person asks a question or requests a draft Reads permitted evidence, reconciles updates, identifies attention items, and drafts a response. The person reviews the result and carries out subsequent actions. Reviewed outputs preserve project identity, source facts, uncertainty, and human decisions.
2 — Supervised TPM coworker A person delegates a concrete assignment Carries out the defined task through permitted steps, saves progress in external records, verifies any requested writes, and returns exceptions to the person. It can complete an assignment without a prompt for every step. Actual task runs show fresh-conversation recovery, correct destinations, destination readback, duplicate prevention, and accurate reporting of partial or uncertain outcomes.
3 — Proactive project operator A verified schedule or supported event starts a bounded workflow Checks for relevant changes, identifies overdue actions or stalled dependencies, prepares next steps, and performs routine actions only where the task and native controls permit them. It brings business decisions and exceptions to a human. Observed trigger-driven runs show correct evidence coverage, useful alerts, suppression of resolved or duplicate follow-ups, quiet-hour handling, and successful pause and recovery.
4 — Delegated project operator within an agreed scope A human delegates a defined operating objective and constraints Repeatedly observes, plans routine coordination, acts where permitted, verifies outcomes, and adjusts within the agreed scope across runs. It tracks whether the intended result occurred, rather than stopping after a report or send. A representative live trial meets agreed quality and effort thresholds, handles missing access, failures and conflicting evidence, and preserves escalation, audit history, human intervention, and recovery.

An accountable project owner is a separate human role, not another agent maturity level. Across all four levels, people retain strategy, scope, commitments, priorities, risk acceptance, sensitive escalation, exceptions, and acceptance of closure. Level 4 means operating a delegated coordination loop; it does not mean independently owning every project decision or doing the delivery team's work.

Where the current ChatGPT Workspace Agent is

Current position: a supervised AI TPM coworker in intended role, with reactive assistance as the established current use and Level 2 execution requiring workflow-specific evidence. The maintainer reports using it mainly for tedious tracking and reporting: gathering updates, reconciling changes, identifying open items, and preparing reports for review. That supports a Level 1 usage description; it is not a formal runtime qualification.

The seven skills, external project records, action verification, and recovery instructions provide the design for Level 2. Confirmed completion of a particular delegated task can establish Level 2 for that workflow, but configuration alone cannot establish it for every capability. This public starter contains no runtime evidence establishing Level 3 or Level 4. It should not be described as running a whole project on its own.

Cloning the repository, installing skills, adding connectors, or selecting a model does not confer a maturity level. Record actual outcomes in a private validation receipt, with the installed version, workflow, user/source scope, observed results, and remaining gaps.

Next steps toward proactive operation

  1. Establish Level 2 for one useful workflow. Choose a daily snapshot or weekly report. Test evidence reconciliation, fresh-conversation recovery, an explicitly requested reversible save with readback, duplicate requests, and partial failures. Record results using the acceptance scenarios; measure total human effort including review and correction.
  2. Pilot Level 3 detection before automated follow-up. Configure one supported schedule or event for one registered project and verify actual runs. Begin with read-and-draft checks for overdue actions or stalled dependencies. Review whether alerts are accurate, timely, and worth the interruption.
  3. Extend only the demonstrated action path. Identify the exact routine task, sources, destination, audience, timing, and stop conditions under the effective native controls. Test stale evidence, recent replies, quiet hours, duplicate suppression, uncertain writes, pause, and recovery before relying on automatic coordination. A maturity label or saved preference is not authorization.
  4. Consider Level 4 only after Level 3 is reliable. Agree a narrow operating objective and human decision boundaries. Evaluate the full observe–act–verify–reassess loop over a representative live trial, with quality and effort thresholds agreed in advance. Verify human intervention and recovery; expand scope only after reviewing evidence.

The next development phase targets Level 3 proactive support after strengthening Level 2 execution. These are implementation and validation steps, not features enabled by this README or a promise of autonomous project ownership.

Data sources: purpose, ownership, and access

Source Purpose and owned information Boundary
Individual ChatGPT Project or Codex setup Private role context, goals, working notes, experiments, and personal preferences User-private by design; verify the actual sharing settings. Never copy automatically into shared records
Jira Authoritative project tracking and execution status: workflow state, assignee, priority, and execution dates Reports and caches cite Jira; agent delivery-risk assessments remain separate from recorded status
Confluence Shared human-facing context, plans, evidence links, documented decisions, reports, and history Distinguish proposals, approved decisions, superseded information, and evidence; it is not a competing tracker
SharePoint Shared registry, project metadata/configuration, source mappings, runbooks, checkpoints, provenance, and action history Operational continuity, not personal memory or a second authoritative execution-status store
Slack, email, calendar, and approved meeting records Supporting evidence and requested stakeholder coordination A message is not automatically a commitment; a calendar invitation is not proof of a meeting or decision
This GitHub repository Reusable agent instructions, skills, public templates, and fictional examples Contains no live project records, private deployment bindings, or deployed-agent state

Shared means available to authorized users, not everyone who can invoke the agent. Verify the effective connector identity and destination audience on each path; an agent must not expand source access. Avoid copying restricted evidence into broader reports or shared caches. Different access rights can legitimately produce different coverage, which must be disclosed without exposing restricted material.

For example: Jira records “Blocked,” a message says “looks done,” and a Confluence report says “on track.” Preserve Jira's recorded status, cite and date the conflicting evidence, and flag the reconciliation needed. Do not silently redefine the status. Likewise, a newer proposed launch date does not replace an approved baseline.

These products are the reference deployment, not employer-specific dependencies. Other trackers, wikis, and record stores can substitute when source ownership, access, identity, revision handling, and write verification are explicitly mapped and retested.

How personal work and shared project work fit together

  1. Keep individual context in your own private ChatGPT Project or Codex setup. A personal workspace can be durable even when the current task is short.
  2. For coordinated project work, invoke the shared agent from a Project thread if your environment supports that workflow. Verify which instructions, files, and context actually reach it; invocation does not prove full context transfer.
  3. Supply the registered project identity, your role, the task, relevant preferences, and sharing constraints. If invocation is unavailable, use a direct agent conversation with that explicit context.
  4. The agent resolves shared project configuration in SharePoint, reads Confluence context and Jira tracking, and uses permitted supporting evidence. It tailors the response to you without changing shared facts.
  5. When an experiment becomes an accepted project decision or commitment, deliberately publish the approved content to its owning system and verify the result. Personal opinions and exploratory notes stay private unless explicitly selected for sharing.

An engineering lead might ask for technical dependencies while a TPM asks for delivery exposure. Their views should differ in emphasis, with compatible facts where their permissions and source versions overlap. Private context must not leak through reports, action logs, caches, or another user's response. See the context and isolation test procedure.

Start here

  1. Read the architecture and build/install guide.
  2. Inspect the agent instructions, operating contract, and seven skill directories.
  3. Run the local checks and package builder below. They need Python 3.10+ and only its standard library; no credentials or network access.
  4. Paste the creation prompt into your supported ChatGPT Workspace Agent creation assistant. Codex can also prepare the artifacts and guide installation. Supply a tag/commit or the generated archive.
  5. Connect your own approved sources, then test the fictional first run before onboarding real project data.
python3 scripts/validate.py
python3 scripts/package.py

The build writes dist/agentic-tpm-workspace-starter-0.2.2.zip, seven ZIPs in dist/skill-packages, and a SHA-256 manifest. Source files are human-readable and can be installed manually if the creation assistant cannot retrieve GitHub.

What is included

Path Purpose
agent/instructions.md Standing role, routing, evidence and reliability instructions
agent/operating-contract.md Source ownership, persistence and native controls
agent/acceptance-tests.md Behavioral acceptance scenarios; not an executed test result
skills/ Seven reusable skills with metadata and supporting files
deployment/deployment.example.json Empty organization-specific bindings; fill a private copy
docs/ Architecture, connectors, installation, SharePoint and Confluence layouts, personal-context testing
templates/sharepoint/ Registry, configuration, context, source map, preferences, state, evidence, actions, runs and report manifests
templates/confluence/ Project home, daily, weekly and meeting templates
examples/example-launch/ Complete fictional project context, connected records, evidence and sample project pages
prompts/ Creation, first exercise and optional schedule prompts
scripts/ Offline validation and reproducible packaging
templates/validation-receipt.json Record actual installation and test outcomes

Seven capabilities

Skill Use it for
tpm-runtime-onboarding Resolve missing project identity and verified per-user defaults
tpm-onboarding-planning Register new work or preserve an in-flight handover
tpm-status-attention Find changes, blockers, stale evidence, risks and decisions
tpm-daily-snapshot Reconcile an exact time window into one dated snapshot
tpm-reporting Prepare weekly, milestone or audience-specific reports
tpm-stakeholder-followup Draft or execute specifically requested coordination
tpm-closeout Assess acceptance, handoff and remaining obligations

A new project gets configuration and a runbook. It does not need a new copy of every skill or a separate agent.

Prerequisites and platform boundaries

You need access to a workspace that supports creating/editing the intended agent, plus the apps you choose to connect and source-system permissions for the exact project. This package does not provide those entitlements. User interface labels and connector actions may change; inspect the current surface. Official ChatGPT Workspace Agent documentation is the product reference; this repository documents a particular setup pattern.

The main deployment uses SharePoint for durable metadata, Confluence for the team-facing project record, Jira for execution fields, and optional Slack/Outlook evidence. Other systems can be mapped deliberately, with ownership and verification contracts retested. See connector summary.

Native platform and connector controls remain effective. This starter does not implement custom authority modes, persisted blanket grants, or approval bypasses. Task authorization, tool permission and accountable business decisions are separate. Draft-only work stays draft-only. A configured connector is not a proven connection for every user.

Your project data belongs outside this public repository

Use a private deployment copy and your approved record store. Do not commit actual tenant domains, emails, project IDs, source extracts, account mappings, credentials, tokens, cookies, internal screenshots or operational audit history. Example Launch is entirely fictional; example.invalid links intentionally do not resolve.

Templates contain nulls and named placeholders by design. Resolve required bindings before live access. Store classification and retention choices in private configuration. Never replace an incomplete real record with a blank template.

A useful first task

Ask: “For this registered project, what changed since the previous cutoff, what needs attention today, and which claim is uncertain?” Start with the fixture and draft-only behavior. Then test one exact approved read, fresh-conversation recovery, and an explicitly requested reversible write with destination readback.

A daily snapshot should lead with current status, due work and the top risk. The appendix holds traceability and coverage. A weekly report uses verified daily snapshots and admits missing days. An invitation does not prove a meeting occurred; a chat suggestion does not create a commitment; a Done label does not prove closure acceptance.

What the checks prove

The local validator checks required files, JSON parsing, linked fixture identity, local Markdown links, skill layout, and common accidental private-data patterns. The packaging script validates first and builds deterministic archives with SHA-256 hashes. These checks do not prove hosted reasoning, live permissions, connector behavior, schedule execution, or fault recovery.

Use the read-only implementation audit to compare an installed agent with this model. Run acceptance scenarios against the actual agent and record outputs. Personal-project context transfer must be tested independently. Configure and observe schedules separately; a saved cadence is not a running automation.

Maintaining your version

Fork or copy the starter, keep private bindings separate, review changes, increment VERSION, update CHANGELOG.md, rerun checks, and build a new release. Upload instructions, references and skills as a matched version. Verify the installed inventory and a fresh chat before broader use. GitHub version control does not automatically deploy or roll back the hosted agent; see the release procedure in the install guide.

Reuse and contribution

Licensed under MIT. See CONTRIBUTING.md. Keep examples fictional and changes reproducible. A useful contribution includes the failing scenario, expected behavior, revised instructions or templates, and the limits of the validation performed.

About

Reusable Workspace Agent setup for evidence-backed project coordination: instructions, seven skills, templates, and fictional examples.

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages