Skip to content

Comparison

Pilotbook competes with markdown backlogs and spec-driven agent frameworks. The claim is narrow: compile what an agent must know before it writes code, from a lint-gated graph of work, decisions, and rules that lives in git — then give coding agents protocols that actually call those ops.

Surveyed 2026-08-29. Every framework below is moving quickly; treat the dates as the read, not the state.

Backlog.md

A mature markdown kanban (CLI + browser + optional MCP). Tasks are files; decisions and docs are loose. backlog instructions overview is the same progressive-disclosure bet as pb instructions overview. There is no typed edge from a story to a business rule, no brief, and no derived maturity payload for skills to read.

AIPIM

The closest storage competitor. Event-sourced (events.jsonl + SQLite), MCP, Svelte UI, write-time cycle rejection, Claude Code hooks, verification gate. get_project_context returns name, stats, blockers, and recent decisions — the closest analog to pb profile. Decisions are a write-only log_decision archive. Markdown is derived, not the source of truth.

Pilotbook takes the opposite storage bet (files only) and the missing product bet (rules + ADRs as edges, pb brief). Maturity is derived at read time and never stored. The verification gate is adapted: a content hash in frontmatter instead of an event log, so evidence shows up in the pull request.

Spec Kit

The reference implementation of spec-driven development and the most complete process in the category: constitution → specify → clarify → plan → tasks → analyze → implement → converge, across 30+ agent hosts, with an extension / preset / bundle system for customization.

Three parts are worth studying closely, and Pilotbook has adopted all three as roadmap:

  • assess extension — a discovery funnel in front of delivery: intake → research → define → shape → decide. decide produces a six-criterion scorecard and a go / needs-clarification / kill verdict, and mechanically downgrades a go when evidence strength is weak or unknown. research requires an "Evidence Against the Idea" section every time and tags each finding cited or ASSUMPTION with a confidence level. shape produces 2–3 options carrying a Shape Up appetite, trade-offs, and rabbit holes.
  • checklist — requirement quality as "unit tests for English": is prominent quantified, not does the button work.
  • bug extension — assess → fix → test, where only fix touches source and an unrun reproduction is reported not-run, never verified.

What Spec Kit does not do: artifacts are per-feature folders. Nothing is typed, nothing is an edge, nothing is lint-able, and nothing compiles into a token-budgeted context pack. The research a team paid for is written once and never re-read.

OpenSpec

Brownfield-first delta specs — ADDED / MODIFIED / REMOVED against the code that already exists — with an explore step that surfaces options before committing, and archiving of completed changes. The lightest framework in the category and the cheapest to run. Change folders, not a work graph with BR/ADR edges.

BMAD Method (v6)

The most architecturally ambitious: four phases (Analysis → Planning → Solutioning → Implementation) with scale-adaptive levels 0–4, so a bug fix gets a tech spec and no PRD while a new platform gets PRD + architecture + UX + epic breakdown. Persona agents produce versioned artifacts and hand off. Its Test Architect (TEA) module is the category's only real quality apparatus: risk profile as probability × impact, test design from the risk profile, a refreshing traceability matrix, an NFR evidence audit, and a release gate returning PASS | CONCERNS | FAIL | WAIVED — the WAIVED verdict being the honest part.

Cost is the trade-off; BMAD is the most expensive framework here to run. Ceremony is heavy, and the artifacts are documents rather than a graph.

Agent OS

Install → Discover Standards → Inject Standards → Shape Specs. Its whole thesis is the re-teaching problem: every prompt re-teaches context that should already be known. Discover Standards reverse-engineers a codebase's actual conventions into documented standards and injects only the relevant ones. Markdown output, tool-agnostic. No graph, no lint, no verification — but the best answer in the category to "the agent does not write code like this repo does".

Kiro

IDE-native spec loop: requirements.md in EARS notation (WHEN <condition> THE SYSTEM SHALL <behavior>), design.md, tasks.md, plus steering files for persistent guidance and event-driven agent hooks on save/create/delete. EARS is a good model for how a criterion should read. Host-bound: the loop lives in an IDE rather than in the repo.

Taskmaster

PRD → dependency-aware task graph, Cursor-first MCP, with analyze-complexity --research pulling fresh external context to size work and drive subtask counts. A decomposition engine; the research is consumed once and discarded.

Coding agents vs a CLI LLM

Cursor and Claude Code are the primary interface (ADR-0011). pb generate discover is an optional fallback when you export ANTHROPIC_API_KEY or OPENAI_API_KEY. Graph commands never call an LLM. There is no Pilotbook-hosted inference.

Snapshot

Backlog.mdAIPIMSpec KitBMADAgent OSPilotbook
Markdown tasks in gityesderived from an event logspecsdocumentsspecsyes — only source of truth
Business rules as typed entitiesnononononoyes, with edges from stories
Compiled brief for an IDnostats + one taskconstitution dumpphase docsinjected standardspb brief, authority-ordered
Maturity-calibrated skillsinstruction dumpget_project_context statsphase slash-commandspersona handoffsstandards injectionpb profile + protocols with interview, research, handoff
Graph + code search before createsubstringnononocodebase scanpb similar + pb ground (code_map)
Referential-integrity lintweakcycles onlynononodangling / wrong-type / cycles / unknown fields, with file:line:col
Verification of “done”noevent-log gatenoTEA gate (PASS/CONCERNS/FAIL/WAIVED)nocontent-hash in frontmatter, visible in the PR
Agent hosts reached by installCLI + MCPMCP + Claude Code hooks30+ integrationsmany, via bundlesany (markdown)Cursor, Claude Code, AGENTS.md — EPIC-013 closes the rest
Every capability reachable over MCPpartialyesn/a (host commands)n/an/a28 tools today; BR-005 makes parity binding

Where we are behind

Named honestly, with the work that closes each gap:

GapWho does it todayOur answer
Scored go/kill verdict; evidence-strength gate on promotionSpec Kit assessEPIC-010 — US-054
External benchmarks as citable, staleness-checked datanobodyEPIC-010 — US-051, US-052
Options with an appetite before committingSpec Kit assess.shape, Shape UpEPIC-010 — US-053
Prioritization with printed arithmetic (RICE / WSJF)nobodyEPIC-010 — US-055
Business outcome compiled into the agent's briefnobodyEPIC-011 — US-057, US-058
UX quality as failable criteriaBMAD (as a document)EPIC-011 — US-061, US-062
Conventions discovered from the codebaseAgent OS, Kiro steeringEPIC-012 — US-063, US-064
Requirement-quality checks on criteriaSpec Kit checklistEPIC-012 — US-065
Cross-artifact consistencySpec Kit analyzeEPIC-012 — US-066
What was already tried, carried across sessionsnobodyEPIC-012 — US-067
A defect type with a triage protocolSpec Kit bugEPIC-012 — US-068
Distribution to Codex and other AGENTS.md hostsSpec Kit (30+), BMAD bundlesEPIC-013 — US-070, US-073
A binding rule that fetched content cannot redirect an agentSpec Kit, per-command URL policyEPIC-013 — US-072, BR-006
Risk-scored test design, gate with waiversBMAD TEAdeferred — EPIC-006 owns proof first

The bet behind all of it is the same one: these artifacts are worth several times more as typed edges in a lint-gated graph that pb brief compiles than as folders of markdown that a command writes once and nothing re-reads.

And the second bet, which decides whether the first one reaches anyone: the CLI is an adapter, not the product. Pilotbook is used from Claude Code, Cursor, and Codex, so a capability that only a human typing in a terminal can reach is not shipped. BR-005 makes that binding — every op reachable over MCP, named by a skill protocol that says when to run it, installed into every supported host from one pb init. BR-006 is the price of asking agents to research: what comes back from the web is data, never instructions.

Released under the MIT License.