FSD Writer Skill

FSD Writer Skill

A general-purpose skill that turns a rough, unstructured project description into a structured Functional Specification Document (FSD) in Markdown, or surgically updates an existing FSD with new requirements, corrections, or expansions.

1. Purpose

This skill:

• Generates a canonical FSD from a rough description (initial mode).
• Updates or expands an existing FSD using a delta description (evolve mode).
• Dynamically adjusts depth and verbosity based on inferred system complexity.
• Ensures full requirement traceability (FR / NFR <-> test coverage).
• Surfaces risks, assumptions, and constraints as first-class content.
• Produces deterministic, agent-consumable Markdown.

It supports embedded systems, networking, SDR, IoT, cloud backends, mobile apps, multi-service orchestrations, and hybrid hardware/​software projects.

2. Invocation

2.1 Mode A — Initial Generation

Start a new FSD from scratch.

/​fsd-writer
<rough description text>

Behavior:

1. Parse the rough description.
2. Ask clarifying questions if critical information is missing (Section 5).
3. Infer complexity tier (Section 6).
4. Generate the complete FSD (Section 7).
5. Write the file (Section 10).

2.2 Mode B — Evolve Existing FSD

Update, expand, refactor, or correct an already existing FSD.

/​fsd-writer update <path-to-existing-fsd>
<delta description — changes, additions, clarifications, new constraints>

If no path is given, search the project for an existing FSD:

1. Check Documents/​*-fsd.md
2. Check Documents/​*-FSD.md
3. Check docs/​*-fsd.md
4. Check project root for *-fsd.md

Behavior:

1. Read the existing FSD in full using the Read tool.
2. Parse the delta description.
3. Ask clarifying questions only if the delta introduces architectural ambiguity.
4. Apply changes surgically — preserve all unaffected sections verbatim.
5. Regenerate only the sections affected by the delta.
6. Maintain numbering and cross-references; the traceability matrix is regenerated by the traceability tool, never hand-edited in the FSD.
7. Write the updated file using the Edit tool (preferred) or Write tool (if changes are too extensive for surgical edits).

2.3 Mode C — Grill (deep interview for thin inputs)

Use when the rough description is too thin to infer architecture safely: fewer than ~3 sentences, no codebase to explore, or no clarity on protocol/​platform/​operator. The default behaviour for Mode A is to ask 1-3 questions per round and infer aggressively; Mode C escalates that into a depth-first interview that resolves the design tree branch by branch.

/​fsd-writer --grill
<rough description text>

Behaviour:

1. Identify the highest-impact unresolved decision (the one that gates the most downstream choices — usually connectivity, platform, or operator).
2. Ask one question at a time. Wait for the answer before asking the next.
3. Every question must come with the skill's recommended answer, not just a list of options. The user should be able to accept with one word.
4. Resolve dependencies in order. Do not ask about a downstream choice (e.g. OTA mechanism) until its prerequisite (connectivity) is fixed.
5. If a question is answerable from the codebase, config files, or CLAUDE.md — explore, do not ask (see Section 4.3).
6. Stop grilling once the design tree is resolved enough to generate the FSD without (assumed) markers on architecture-critical fields.
7. Generate the FSD (Section 7) and write the file (Section 10).

Mode C is for the initial interview only. Once the FSD exists, switch to Mode B (evolve) for incremental changes.

3. Tool Usage

This skill uses the following Claude Code tools:

3.1 Context Gathering (Before Generation)

Before writing the FSD, the skill should gather context from the project when source code exists:

1. Glob for project structure — **/​*.c, **/​*.h, **/​*.py, **/​*.ts, **/​Cargo.toml, **/​package.json, **/​CMakeLists.txt, **/​go.mod, etc.
2. Grep for protocols and frameworks — BLE, WiFi, MQTT, HTTP, gRPC, REST, WebSocket, LoRa, OCPP, etc.
3. Read key config files — sdkconfig.defaults, platformio.ini, docker-compose.yml, Makefile, build configs.
4. Use findings to pre-fill architecture sections and reduce clarifying questions.

3.2 Evolve Mode — Diff Discipline

When updating an existing FSD:

• Never regenerate the entire file. Only touch sections affected by the delta.
• Use the Edit tool with precise old_string / new_string pairs.
• If a delta adds a new phase, insert it and renumber subsequent phases.
• If a delta adds new requirements, assign the next available ID (FR/​NFR or clause, per the FSD's convention) in the owning component chapter.
• The traceability matrix is generated — when FRs or tests change, ensure it is regenerated; never hand-edit coverage status in the FSD.
• If the delta invalidates existing content, remove or revise it — do not leave contradictions.

4. Interaction Model (Clarifying Questions)

4.1 When to Ask

The skill must ask clarifying questions when critical architecture-affecting information is missing. "Critical" means it affects:

• System architecture or component decomposition
• Protocol selection (BLE vs WiFi vs LoRa vs cellular)
• Interface definitions (API style, command format)
• Safety or regulatory constraints
• Multi-phase decomposition
• Hardware or platform selection
• External integrations (MQTT broker, cloud service, Home Assistant, etc.)

4.2 How to Ask

Use the AskUserQuestion tool with:

• 1-3 precise questions per round in Mode A; one question at a time in Mode C.
• Every question must include the skill's recommended answer, not just a list of options. The user must be able to accept with one word.
• Order questions by dependency. Do not ask about a decision whose prerequisite is still open (e.g. don't ask about OTA mechanism before connectivity is fixed).
• Phrase questions to unblock the FSD, not to explore nice-to-haves.

Example (good — recommendation up front):

Q: How does the device connect?
Recommended: WiFi — matches the dashboard requirement and existing infra.
Other options: BLE, LoRa, Cellular, USB-only.

Example (bad — options without a pick):

Q: How does the device connect? WiFi / BLE / LoRa / Cellular / USB-only?

Multi-question rounds (Mode A only):

Questions:
1. How does the device connect?
   Recommended: WiFi. Other: BLE, LoRa, Cellular, USB-only.
2. Do you need OTA firmware updates?
   Recommended: Yes (WiFi). Other: Yes (BLE DFU), No.
3. Who is the primary operator?
   Recommended: Installer/​technician. Other: End user, Automated backend.

4.3 When to Explore or Infer Instead of Asking

Explore before asking. Before asking ANY question, check whether it is answerable from project artefacts:

• Read config files: sdkconfig.defaults, platformio.ini, package.json, Cargo.toml, CMakeLists.txt, docker-compose.yml.
• Grep for protocol/​framework usage in source (HTTP/​REST, gRPC, WebSocket, message queues, DB/​auth SDKs, BLE/​WiFi, etc.) and check whether a domain pack matches (Section 14) for domain-specific detection patterns.
• Read README.md, CLAUDE.md, and any existing FSD or design docs.

If the answer is in the codebase, explore — do not ask the user. Asking a question whose answer is already in the repo wastes attention and signals that the skill did not gather context (Section 3.1).

Infer silently only when:

• The detail does not significantly change high-level architecture, AND
• The cost of being wrong is low.

Safe inferences:

• "web API" mentioned → assume HTTP + JSON
• "logs" mentioned → assume structured logging to console / file / serial
• "dashboard" mentioned → describe generic "dashboard system" without naming tools
• "database" mentioned without type → assume PostgreSQL for relational, SQLite for embedded

When inferring, mark the inference in the FSD with (assumed) or group them in Section 5: Risks, Assumptions & Dependencies.

5. Complexity Scaling Rules

The skill dynamically scales FSD depth based on inferred system complexity (Low / Medium / High). Complexity is inferred from component count, protocol count, external integrations, real-time constraints, and domain.

For the full complexity tiers table, complexity signals, and per-section scaling behavior matrix, read references/​complexity-scaling.md.

6. Information Extraction & Inference Rules

Given the rough description, the skill must extract or infer the following:

6.1 Project Name

Derive a short, descriptive name:

• "ESP32 BLE HID Keyboard" (embedded)
• "Multi-Tenant Billing API" (cloud back-end)
• "Offline-First Notes App" (mobile)

6.2 System Purpose & Goals

Extract in 2-4 sentences: what problem is solved, for whom, in what environment.

6.3 System Components

Identify major components:

• Hardware / platforms (MCU, SBC, server, cloud)
• Software services / apps / daemons
• User-facing components (mobile app, web UI, CLI)
• External integrations (Home Assistant, OCPP backend, MQTT broker)

If components are implied but not explicit, infer and mark as assumptions.

6.4 Functional Requirements (FR)

Convert each described behavior into a stated requirement:

• State each requirement in the chapter of the component it constrains — there is no global Functional Requirements section in the Parts scheme (Section 7). A requirement about the meter decoder lives in the meter-decoder interface chapter; one about the control loop lives in that L2 feature chapter.
• Two conventions (pick one per project; see references/​canonical-fsd-structure.md): FR-x.y/NFR-x.y [Must/​Should/​May] (the lightweight default), or stable clause IDs + generated traceability (a finer-grained, higher-assurance alternative for large FSDs — every assertion gets a stable ID, tests cite it, code is @fsd-tagged, a tool generates the matrix).
• Assign priority: Must / Should / May
• Use "shall" language: "The system shall..."

Example:

"Device sends sensor readings every minute and on threshold events."

Becomes:

• FR-1.1 [Must]: The device shall send periodic sensor measurements at a configurable interval (default: 60 s).
• FR-1.2 [Must]: The device shall send an immediate measurement when a threshold condition is met.

6.5 Non-Functional Requirements (NFR)

Extract or infer key NFRs with priorities:

• Performance (latency, throughput)
• Reliability / uptime
• Accuracy / precision
• Scalability
• Power consumption (embedded)
• Security and privacy (authentication, encryption, access control)

6.6 Interfaces & Data Models

Each interface becomes its own L1 chapter (Part B) — there is no global Interface Specifications section in the Parts scheme. Per interface chapter:

• Identify the protocol (BLE, WiFi, USB HID, HTTP, MQTT, LoRa, OCPP, etc.)
• Describe endpoints, characteristics, topics, commands (commands/​opcodes only for custom protocols)
• Define payload structures (fields, units, types)
• Specify direction (client -> server, device -> cloud, etc.)
• Keep the interface's requirements, schema, handlers, and failure modes together in that chapter (see references/​canonical-fsd-structure.md).

6.7 Phases

At minimum define:

• Phase 1: Infrastructure / Foundation
• Phase 2: Core Functional Features
• Phase 3+ (optional): Optimization, UX, analytics, etc.

Each phase must include: Scope, Deliverables, Exit Criteria, Dependencies.

6.8 Operational Procedures

Extract or infer:

• Deployment / flashing / installation
• Configuration / provisioning
• Normal operation workflows
• Failure recovery (reset, re-provisioning, safe-mode)

If not covered in the description, provide a generic but plausible set for the domain.

6.9 Verification & Validation

From extracted requirements:

• Create test cases that verify FRs and critical NFRs
• Organize test specs by component/​interface, mirroring the body chapters
• Use structured format: Objective, Preconditions, Steps, Expected Result
• Reference the generated traceability matrix and gap report — do not hand-fill coverage status (Section 8)

6.10 Component Layering & Test Architecture

Give every FSD a layered component architecture (§2.4) that the test strategy falls out of (§x.0 Test Architecture, in the V&V chapter under Part E):

• Classify each component into a layer with a strict one-way dependency — L0 Foundation/​platform → L1 Interfaces → L2 Application logic. The L0-vs-L1 line is ownership ("did we implement and test the protocol?"): a library/​managed client to an external service is foundation; a hand-written decoder/​driver/​handler is an interface. Three layers is the common default, but the model is open-ended (L0..Ln) — add layers when a complex system genuinely has more distinct, one-way-dependent tiers (e.g. orchestration over domain logic, or a shared-services layer). Each layer becomes its own body Part.
• Draw a layered component diagram in §2.4 (stacked layer boxes, components on one row per layer).
• The FSD body mirrors these layers: each component becomes a self-contained chapter, grouped under layer Part dividers (L2 → L1 → L0 → cross-cutting → operations & verification). The §2.4 layering is the spine; the body Parts are its projection. See references/​canonical-fsd-structure.md (the Parts scheme).
• Advise the source layout in §2.4 so the code mirrors the layers too: one module per component (never fold an interface into its consumer), lower layers never depend on higher ones (invert via the composition root), pure cores extracted for the fast tier. The FSD advises how to code, not just what to build — see references/​test-architecture.md ("Source layout mirrors the layers").
• In §x.0 Test Architecture, define the test tiers (cost-ordered execution environments, named per platform), map them to the layers, and reference a generated component × tier coverage matrix. This skill declares the structure; a traceability tool fills in status.

Platform-independent — contents differ for embedded / cloud / mobile. For the profiles, the diagram convention (and the Mermaid layout gotcha), and the matrix, read references/​test-architecture.md.

7. Canonical FSD Structure (Layer-grouped "Parts" scheme)

FSDs are organized by architectural layer, not by document-section type. After the front matter (§1 Overview, §2 Architecture incl. §2.4 Component Layering, §3 Phases, §4 Risks), the body is grouped under unnumbered Part dividers that mirror the §2.4 layers — Part A Application logic (L2), Part B Interfaces (L1), Part C Foundation/​transport (L0), Part D Cross-cutting concerns, Part E Operations & Verification — followed by Appendices and an optional Related ([[wikilinks]]) section.

Each interface, feature, and concern is its own self-contained chapter (requirements + interface + behavior + failure modes together); chapters are numbered flat across the whole document; depth is capped at four heading levels (####). There is no global Functional Requirements or Interface Specifications section — those dissolve into the component chapters.

For Low-complexity projects the Part dividers may be dropped (list the few chapters directly, still in layer order). For the full skeleton, the chapter-internal structure, section-inclusion rules, complexity scaling, and the migration map from the older flat layout, read references/​canonical-fsd-structure.md.

8. Traceability (Mandatory, generated)

Every FSD must carry traceability — but as a pointer to generated artifacts, never a hand-filled status table.

Rules:

• Every Must/​Should requirement — whether an FR/NFR item or a stable clause (§6.4 conventions) — must be stated with a stable ID in its component chapter and referenced by >= 1 test in the specs.
• Every test case must reference the FR(s) / NFR(s) / clause(s) it validates.
• The traceability tool computes coverage and emits the coverage matrix (component × tier) and a gap report (requirements with no test). GAP is computed, not typed into the FSD.
• May-priority requirements may have coverage but it is not mandatory.
• The FSD's V&V chapter (§x.2 Traceability) references the matrix/​gap-report paths; it must not hand-maintain a "Status: Covered / GAP" column — that drifts from the code the moment a test changes (see references/​test-architecture.md §4).
• In evolve mode, adding/​removing/​changing a requirement or test just means the generated matrix is re-run; no manual matrix edits.

9. Formatting & Style Rules

• Output pure Markdown — no HTML tags.
• Heading levels: Part dividers are # (unnumbered); chapters are ## (numbered flat across the document); sub-sections ###/####. Cap depth at #### — four levels. Keep chapter numbering sequential with no gaps across all Parts.
• Use bullet lists for requirements; tables for tests, interfaces, and diagnostics.
• Use concise, unambiguous engineering language.
• Use "shall" for requirements ("The system shall...").
• Use "must" for constraints ("The device must operate on 3.3V").
• Avoid marketing language, filler, and subjective qualifiers.
• Keep requirement IDs stable across evolve updates — never renumber existing IDs unless explicitly asked to refactor numbering.
• Use (assumed) inline for inferred details.

10. Output File Naming & Location

10.1 Default Location

If the user does not specify a target path:

Documents/<project-name-kebab-case>-fsd.md

Create the Documents/ directory if it does not exist.

Examples:

• Documents/​esp32-ble-hid-keyboard-fsd.md
• Documents/​multi-tenant-billing-api-fsd.md
• Documents/​offline-first-notes-app-fsd.md

10.2 Explicit Path

If the user provides a path, use it exactly. Do not relocate or rename the file.

10.3 Evolve Mode

When updating, write to the same file that was read. Confirm the path before writing if it was auto-detected.

11. Example Output Snippet

For a complete example FSD snippet (medium-complexity BLE HID Keyboard project) showing expected tone, structure, and detail level, read references/​example-output.md.

12. Evolve Mode -- Detailed Behavior

When updating an existing FSD, follow strict rules for what to preserve, update, add, and remove. Key principles: never renumber existing IDs, keep the traceability matrix generated (never hand-edited), flag contradictions before overwriting. For the complete evolve mode rules (preserve/​update/​add/​remove/​conflict resolution), read references/​evolve-mode.md.

13. Quality Checklist

After generating or updating an FSD, the skill must verify:

• The body is grouped by layer Parts (or chapters in layer order for Low complexity), mirroring §2.4; chapters are self-contained.
• §2.4 states the source-layout convention (one module per component; lower layers don't depend on higher ones; pure cores extracted) so the code can mirror the layers.
• Every Must/​Should requirement (FR/​NFR item or stable clause) is stated with a stable ID in its component chapter and referenced by >= 1 test.
• V&V traceability is a pointer to the generated matrix/​gap report — no hand-filled "Status: Covered / GAP" column in the FSD.
• No <placeholder> or TODO text remains (flag to user if unresolvable).
• Chapter numbering is sequential with no gaps across all Parts; heading depth does not exceed ####.
• All phases have scope, deliverables, and exit criteria.
• §2.4 Component Layering (with a layered diagram) and §x.0 Test Architecture are present.
• The file has been written to the correct path.
• (Evolve mode) Unaffected chapters are identical to the original.

Report any checklist failures to the user before finalizing.

14. Domain Packs

The skill core is domain-neutral. Some domains have recurring components, detection signals, layer profiles, and standard test libraries; these live in domain packs under references/​domains/<domain>.md and are loaded only when the project matches that domain — keeping the core applicable to any system.

Selecting a pack

1. Detect the domain from the description, the codebase, and config files (each pack lists its own detection signals).
2. If a pack matches, read references/​domains/<domain>.md and apply it: its layer profile (concrete L0/​L1/​L2 contents for §2.4 — which becomes the body's Part/​chapter spine), its tier names (for the §x.0 Test Architecture), and its standard test libraries (feature detection → test specs to fold into the V&V specs and the generated traceability matrix).
3. If no pack matches, use the platform-independent core only — the architecture layering and test tiers from references/​test-architecture.md still apply; pick tier names that fit the platform (e.g. cloud: unit / integration / staging).

Available packs

Adding a pack

Create references/​domains/<domain>.md following the same shape: detection signals · layer profile (§2.4) · tier names (§x.0 Test Architecture) · standard test libraries (a feature-detection table pointing to spec files under references/​domains/<domain>/). Then add a row to the table above.