DeckForge

A presentation pipeline for coding agents. Six staged skills turn Claude Code, Codex, or OpenCode into a disciplined deck-building system: narrative structure, visual design, copywriting, diagrams, QA, and an auditable engineering build log — instead of one long "make me a presentation" prompt.

DeckForge governs what goes on the slides and whether it's true. Your agent (and your renderer of choice — HTML, PPTX, anything) does the building. Every deck ends with a machine-checked verdict (PASS / PASS WITH WARNINGS / FAILED) and an execution report that records what was claimed, what was verified, what was skipped, and why.

brief + sources
        │
        ▼
┌───────────────┐   ┌───────────────┐   ┌───────────────┐   ┌───────────────┐
│ 1 structure   │──▶│ 2 design      │──▶│ 3 copywriting │──▶│ 4 diagrams    │
│ outline +     │   │ tokens +      │   │ titles, body, │   │ escalation    │
│ fact bank     │   │ archetypes    │   │ fact fidelity │   │ ladder, specs │
└───────────────┘   └───────────────┘   └───────────────┘   └───────────────┘
                                                                    │
        ┌───────────────────────────────────────────────────────────┘
        ▼
┌───────────────┐   ┌───────────────┐   ┌─────────────────────────────┐
│ 5 render      │──▶│ 6 qa          │──▶│ 7 build report              │
│ your deck     │   │ SHIP/BLOCKED  │   │ confidence, telemetry,      │
│ system        │   │ verdict       │   │ PASS/FAILED — internal log  │
└───────────────┘   └───────────────┘   └─────────────────────────────┘

Seven pipeline stages; six are DeckForge skills — stage 5 (render) is your own deck system.

Quickstart

npm install -g deckforge-cli

deckforge install          # copies skills into your detected agent(s)
deckforge init my-deck     # scaffolds brief.md, sources/, build/
cd my-deck

(The npm package is deckforge-cli — the unsuffixed name was already taken by an unrelated project — but the command it installs is plain deckforge.)

Then open your agent in that directory and ask:

Build me a presentation from brief.md — follow the DeckForge pipeline.

Artifacts land in build/ as the stages run; the final build/execution-report.md tells you whether the deck is fit to ship and what to fix if it isn't. Full walkthrough: examples/quickstart/WALKTHROUGH.md.

The six skills

| Stage | Skill | What it enforces | |---|-------|------------------| | 1 | presentation-structure | Narrative outline from story frameworks; every claim traced to a fact bank; kill-filter cuts slides that don't earn their place | | 2 | presentation-design | Design tokens, 11 slide archetypes, evidence craft, 26 AI-tell checks | | 3 | presentation-copywriting | Title/body craft by register; a six-step fact-fidelity gate plus 18 copy-quality checks; escalates instead of inventing | | 4 | diagram-design | Escalation ladder (prose → list → table → diagram); every edge verb-labeled and provenanced; demotion is legal and always reported | | 6 | presentation-qa | Check catalog, consistency sweeps, accessibility annex; computes a SHIP / SHIP-WITH-ACCEPTED-MAJORS / BLOCKED verdict from pre-agreed gates | | 7 | presentation-build-report | Internal build log: per-stage + weighted-overall + weakest-stage confidence, telemetry, source coverage, prioritized recommendations. Deterministic Python aggregator — never part of the deliverable |

(Stage 5, render, is your deck system — HTML, PPTX, anything your agent can drive.)

Skills are Markdown instruction sets (progressive disclosure: a short SKILL.md, detailed references/). Only the build report ships code — a stdlib-only Python 3 module with its own test suite. Details: docs/skills-reference.md.

Agent compatibility

| Agent | Install target | Discovery mechanism | |-------|----------------|---------------------| | Claude Code | ~/.claude/skills/ or ./.claude/skills/ | Native skill discovery (SKILL.md frontmatter) | | Codex | ~/.codex/deckforge/skills/ | Managed block in AGENTS.md pointing at each skill | | OpenCode | ~/.config/opencode/deckforge/skills/ | Managed block in AGENTS.md pointing at each skill |

deckforge install auto-detects which agents are present; --agent and --scope user|project override. Uninstall removes exactly what install created (manifest-tracked). Honest caveat: on Codex and OpenCode the mechanism is an instruction block the agent reads, not native skill support — see docs/cross-agent.md.

CLI

deckforge install    [--agent claude|codex|opencode|all] [--scope user|project] [--force] [--dry-run]
deckforge uninstall  [--agent ...] [--scope ...]
deckforge list       [--json]
deckforge doctor     [--json]
deckforge init       [dir]
deckforge update

Zero npm dependencies. No network calls, no telemetry — ever. deckforge doctor checks node/python versions and agent detection.

Requirements: Node ≥ 18. Python ≥ 3.9 (stdlib only) — needed only for the build-report aggregator; the other five skills work without it. Windows: run Python with PYTHONUTF8=1 (the skills say this themselves; doctor reminds you).

The execution report

Every pipeline run ends with an internal build log — never a slide, never in the export bundle:

• Two confidence numbers, always: a weighted overall (is the build broadly healthy?) and the weakest stage (what's the risk floor?). A deck must not inherit its worst stage as its headline score — and a healthy average must not hide one shaky stage.
• Zero fabrication: unmeasured telemetry is "not captured", silent stages are "none reported". Numbers are never invented.
• Deterministic: same inputs, same report. Scoring rubric and status policy live in code, not vibes: docs/confidence-rubric.md.

A real generated sample (fixture package in, actual aggregator output out): examples/build-report-sample/.

Documentation

• docs/pipeline.md — stage-by-stage guide and artifact contracts
• docs/skills-reference.md — per-skill reference
• docs/confidence-rubric.md — scoring and build-status policy
• docs/cross-agent.md — installation and invocation per agent
• docs/faq.md — common questions
• CHANGELOG.md — including known calibration items planned for v1.1

Contributing

See CONTRIBUTING.md. Skill behavior changes bump the skill's version: frontmatter and get a CHANGELOG entry; the confidence rubric in skills/presentation-build-report/references/confidence-and-status.md is canonical.

License

MIT