How it works

From "build the X feature" to a shippable PR — without losing the plot.

Most AI coding tools start with the prompt. fastpace starts with the spec — and feeds the spec, the domain context, and the linked Jira ticket into your AI agent so it produces code that fits your codebase the first time.

Layer 1 — domain context

The L1 → L5 ladder.

AI agents are stateless. Every prompt starts cold. Domain context is the thing you'd brief a new senior dev with on their first day — and most teams have none of it captured anywhere the agent can read. fastpace gives you five layers of progressively richer ground truth, each one readable by the agent before it writes a line of code.

  1. L1

    Code only

    The agent reads files in your repo. The Cursor / Copilot baseline. Useful for syntax, useless for "should this even exist?"

  2. L2

    + Architecture & patterns

    Bounded contexts, conventions, layering rules, terminology glossary. fastpace/context/architecture.md + fastpace/context/patterns.md. Now the agent knows that "service" means microservice in your codebase, not Spring service bean.

  3. L3

    + PRD / ERD

    Product requirements + entity-relationship-decisions docs in fastpace/prds/ and fastpace/erds/. The agent sees the why and the what, not just the how. Slash commands write these for you.

  4. L4

    + Execution plan

    Waves → epics → stories with hour estimates in fastpace/execution-plan.md. The agent picks up the next story and executes against a sized spec, not a guess.

  5. L5

    + ADRs & learnings

    Architectural Decision Records and accumulated team learnings in fastpace/context/decisions.md. Three months in, the agent stops repeating the same mistake the team already debated and resolved.

Everything lives in fastpace/ in your repo. The repo is the source of truth. Any model, any runtime, any future tool can rehydrate the full institutional knowledge from a git clone.

Layer 2 — spec-driven workflow

PRD → ERD → execution plan, in three slash commands.

Three commands, ten minutes per layer, each one builds on the last. Your AI agent does the heavy drafting; you do the senior-dev review pass.

/fp-write-prd

Interactive: "what feature, who's it for, what success looks like." Agent drafts the PRD; you redline. Output: fastpace/prds/<slug>.md.

/fp-prd-to-erd

Reads the PRD; produces entities, relationships, key decisions, edge cases. Catches the missing-edge-case before code gets written. Output: fastpace/erds/<slug>.md.

/fp-erd-to-execplan

Reads the ERD; sizes the work as waves → epics → stories with hour estimates. You file Jira / Linear tickets straight from this. Output: fastpace/execution-plan.md.

The agent now executes against a real spec — not "build the X feature." Reviewers can compare merged code to PRD acceptance criteria. PMs can check progress against the execution plan without an interview.

Layer 3 — backlog sync

Jira, Linear, Confluence, Google Docs — bidirectional.

Your real backlog is already where you keep it. fastpace pulls the issue body into context when the agent picks up a story, and writes the AI work back to the issue (commit IDs, provenance trailer summary, link to the run manifest) so traceability survives.

Jira

Atlassian Cloud + Server. Stories pulled into context; commit IDs + AI summary written back as comments.

Linear

GraphQL native. Issue → context flow + AI commit ↔ issue linking via auto-comment.

Confluence

Atlassian docs as L2 / L3 context. Page bodies cached locally; updated on-demand.

Google Docs

Drive + Docs API. PRDs and design docs already in Docs become L3 context without copy-paste.

Layer 4 — runtime extensibility

Skills, agents, and hooks shape your AI to your stack.

The bundled defaults work out of the box. The customization surface is what makes fastpace fit your team specifically — TypeScript-only repo, Python-only, monorepo, microservices, per-language linting, your own conventions.

25+ Skills

Slash commands like /fp-write-prd, /fp-discover, /fp-context-score. Drop your own in .claude/skills/; they show up immediately.

7 Agents

onboarder, code-reviewer, security-reviewer, architect, evidence-curator — each with explicit scope manifests (allowed tools, paths, command patterns).

16 Hooks

PreToolUse + PostToolUse: scope guards, runtime guards, audit appenders, run-manifest writers. Add your own per-team policies — e.g. "block all secrets in commit messages."

Layer 5 — provenance & signed receipts

Every AI call leaves a signed trail.

This is the layer the trust edition emphasizes — but it's not only useful for compliance. For day-to-day dev: any reviewer can trace any line of merged code back to the AI call that produced it. Bisect any incident; explain any change in code review.

Five layers. One npm install. Same agent, dramatically better output.