Project Delivery Framework (PDF)

Spec-driven, AI-assisted operating system for the "CEO of the engagement". A Claude Code skill library for senior delivery and engagement managers.

What this is

BMAD, Spec Kit, and the Anthropic Knowledge Work Plugins gave the engineering layer a mature, opinionated AI toolchain. The delivery-management layer had nothing equivalent.

PDF reuses BMAD's architecture — persona-mediated stage agents that dispatch to versioned workflow skills producing markdown artifacts — and points it at the delivery-management layer:

• 10 stage agents for the engagement lifecycle (shaping → mobilization → planning → execution → governance → risk → quality → commercial → people → closure)
• 39 workflow skills that produce structured, audit-ready artifacts
• 6 cross-cutting utilities (red-team, translate, decision-log, audit-log, elicit, diagnose-situation)
• Living engagement charter as the constitution every artifact reconciles to
• Audit-by-default: prompt+model+hash logged per workflow run
• Adversarial red-team gate on every outward-facing artifact
• Fully local-first; engagement data never leaves the user's machine unless explicitly shared

See ARCHITECTURE.md for the full design spec.

Who it's for

The senior delivery / engagement / programme manager who:

• Runs one or more engagements as the "CEO of the engagement"
• Is responsible for budget, delivery, people, client, and quality across an engagement
• Wants a private, opinionated, audit-defensible AI toolchain that operates at their level — not product-management tooling repurposed for delivery
• Is comfortable with Claude Code or a similar agentic CLI

Not for: product managers (use Anthropic's product-management plugin or BMAD); engineering leads in build mode (use BMAD); team platforms with multi-user state (PDF is single-operator-by-design).

Installation

# Install into the current directory
npx @systima/project-delivery-framework

# Or into a specific directory
npx @systima/project-delivery-framework --target ./my-engagement-workspace

# Force overwrite of existing PDF skills
npx @systima/project-delivery-framework --force

# See what would be installed without writing
npx @systima/project-delivery-framework --dry-run

# Check what's already installed
npx @systima/project-delivery-framework status

The installer copies skills into three locations so multiple AI tools work out of the box:

• .claude/skills/pdf-*/ — primary location; discovered natively by Claude Code and OpenCode (70 skill folders)
• .agents/skills/pdf-*/ — tool-agnostic mirror for any agent following the .agents/skills/ convention
• .opencode/commands/pdf-*.md — slash-command wrappers so PDF commands appear as /pdf-help, /pdf-agent-mobilizer, etc. in OpenCode's TUI

You can also use any other tool that supports the .claude/skills/ or .agents/skills/ discovery patterns (e.g. Codex CLI with the relevant plugin).

The installer also creates:

• _pdf/_config/pdf-help.csv — the skill index that drives pdf-help recommendations
• _pdf-output/engagements/ and _pdf-output/practice/ — empty output roots

It does not touch any existing skills from other frameworks (BMAD, Anthropic plugins, etc.). PDF coexists by namespace (pdf-*).

Quick start

Open the installed directory in Claude Code or OpenCode, then:

/pdf-engagement-init        # scaffold a new engagement
/pdf-help                    # show next required action

Or talk to an agent directly:

"hey Marcus"      # Mobilizer (charter, RACI, governance, comms, kickoff)
"hey Ronan"       # Run-Lead (weekly status, RAID, standups, velocity, blockers)
"hey Helena"      # Herald (steering, exec summary, stakeholder updates, escalations)
"hey Klaus"       # Risk Officer (risk deep-dives, change requests, mitigation, escalation decisions)
"hey Petra"       # Planner (plan, capacity, budget, estimate challenge)
"hey Theo"        # Treasurer (budget tracking, margin, change orders, commercial model, invoice prep)
"hey Quinn"       # Quality Steward (SDLC / QA / SysEng / Secure SDLC health)
"hey Iris"        # People Lead (team health, attrition risk, ramp, 1:1s, performance conversations)
"hey Sofia"       # Shaper (qualify, shape, ROM, SoW)
"hey Felix"       # Finisher (closure, retrospective, lessons, case study, handover)

Each agent greets, scans what's done, presents a menu of workflows, and dispatches. Most workflows support a dump intent for ingesting unstructured material (call transcripts, RFPs, status notes) and extracting structured artifacts from it.

The 10 agents

| Stage | Agent | What they handle | |---|---|---| | 01 — Shaping (pre-sales) | Sofia | Qualification (MEDDIC), opportunity shaping, ROM estimates (cone of uncertainty), SoW drafting | | 02 — Mobilization | Marcus | Engagement charter (constitution), RACI, governance plan, comms plan, kickoff deck, stakeholder map (influence/interest + salience + profiles) | | 03 — Planning | Petra | Plan (Mermaid Gantt + phase table), capacity plan (matrix + rows), budget baseline (monthly + milestone), estimate challenge (pre-mortem + red-team) | | 04 — Execution | Ronan | Weekly status (Teams/Slack-paste-friendly), RAID updates, standup digests, velocity checks, blocker triage | | 05 — Governance | Helena | Steering packs (MARP + speaker notes), exec summaries (strict 1-page), stakeholder updates (9 archetypes), escalation memos (SCQA), relationship-health cards (RAG; trust-equation dimensions) | | 06 — Risk & Change | Klaus | Risk deep-dives, change requests (granular taxonomy → classic inference), mitigation plans, escalation decisions | | 07 — Quality | Quinn | SDLC / QA / SysEng / Secure SDLC health cards (RAG-rated; with compliance regime cross-check) | | 08 — Commercial | Theo | Budget tracker (weekly + monthly), margin analysis (period + milestone), change orders, commercial-model reviews, invoice backup packs | | 09 — People | Iris | Team health, attrition risk, ramp plans, 1:1 prep (cumulative per person), performance conversations (Radical Candor) — all confidential by default | | 10 — Closure | Felix | Closure checklist, retrospective (internal + client-shareable), lessons learned (propagated to practice library), case study (internal + public), handover pack |

Core design principles

1. Charter as constitution. Every artifact reconciles to the engagement charter; charter changes are append-only with a visible change-log.
2. Spec-driven, not template-driven. Each artifact has structured frontmatter linking it to its sources and the model used.
3. Stage-orchestrated, persona-mediated. Agents route; workflows do the work.
4. Local-first. Engagement data lives in _pdf-output/engagements/<slug>/ and is gitignored by default. Confidential 09-people/ data is excluded even when other engagement data is committed.
5. Audit-by-default. Every workflow invocation writes a JSONL audit record.
6. Adversarial before stakeholder. Every outward-facing artifact passes pdf-red-team before delivery.
7. House-style separable. Persona, validation thresholds, and templates are all overridable via per-engagement .pdf-config.toml or per-skill customize.toml.

Repository layout

project-delivery-framework/
├── ARCHITECTURE.md          # full design spec (20 sections, decisions log)
├── CHANGELOG.md             # per-stage build history
├── README.md                # this file
├── LICENSE                  # MIT
├── package.json             # npm metadata; `bin` = installer CLI
├── bin/cli.js               # npx installer
├── .claude/skills/          # 70 skill folders (pdf-*)
├── _pdf/_config/            # pdf-help.csv (60-row skill index)
└── _pdf-output/             # output root (engagements + practice library)
    ├── engagements/         # one folder per engagement (gitignored content)
    └── practice/            # cross-engagement lessons library

Compatibility

• AI tooling: Claude Code is the primary target. Skills are file-based markdown + TOML and portable to other Claude-Code-compatible agents (Codex CLI, Cursor, OpenCode) with minor adaptation.
• Coexistence: PDF's pdf-* namespace doesn't collide with BMAD's bmad-* skills. Both can be installed in the same workspace.
• Node: ≥ 16.7.0 for the installer (uses fs.cpSync).

Status

v0.1 — feature-complete. 55 workflows + 10 agents built across 11 development stages with explicit consultation gates per stage. See CHANGELOG.md.

Known future work (out of scope for v0.1):

• API sync providers (Jira, ADO, GitHub, Confluence, Calendar, Slack/Teams, time-tracking) — architecture is forward-compatible; see ARCHITECTURE.md §20.
• Cross-engagement portfolio view (pdf-portfolio).
• Practice-library search during engagement init.
• Real-world hardening via dry-runs and feedback from first uses.

Influences

Acknowledged debts:

• BMAD-Method — the skill anatomy, agent-vs-workflow split, central CSV index, and per-stage orchestration pattern.
• Spec Kit / GSD — the constitution.md pattern that became PDF's living charter.
• Anthropic Knowledge Work Plugins — the operations / productivity / engineering / product-management plugins informed several workflows; PDF uses the file-based, locally-runnable, Apache-2.0 design philosophy.
• ArcKit — audit-ready artifact structure; Orange Book risk taxonomy informing Klaus's design; MADR ADRs.
• PMI's framing of delivery roles and prompt-engineering guidance — for the role taxonomy in customize.toml files (not the PMBOK content itself).

PDF is the delivery-management-layer sibling these projects were missing.

Contributing

This is currently a single-operator project. If you find it useful and have improvements, open an issue or PR; PRs should:

• Match the existing skill anatomy (SKILL.md frontmatter + customize.toml + optional references / assets)
• Include a CHANGELOG entry
• Pass npx @systima/project-delivery-framework status after install in a clean directory

Licence

MIT — see LICENSE.