| name | tdd-writer |
| description | Draft comprehensive Technical Design Documents. Analyzes codebase, fetches ticket/design/doc context via MCP. Every technical claim is verified against actual code. Unverified information is explicitly marked. Use when the user wants to write a TDD, create a technical design document, plan a large engineering project, or says 'write a TDD'. |
TDD Writer
Draft Technical Design Documents for staging up large or complex projects. A TDD is a communication tool: a reader who has never seen the work should be able to read it start to finish and come away understanding what you're proposing, why, and how it will be built. It is not a form to fill in, and it is not a checklist with the prose removed.
REQUIRED BACKGROUND: Invoke /flagrare:write-docs before drafting. This skill owns what a TDD must cover and how to verify it; write-docs owns how to make the prose readable. The single most common failure of an AI-drafted TDD is the "medicine sheet", every section flattened into terse bullets, no causality, nothing a human wants to read. The craft layer in write-docs is the antidote, and the section "Writing the document" below applies it specifically to TDDs. Read both; do not skip the handoff.
When to Use
- Projects estimated at 2+ weeks
- Complex features with multiple components
- Architectural changes affecting multiple services
- Features requiring cross-team coordination
- User says "write a TDD", "design doc", "technical design"
No Assumptions Policy
Never assume, imply, or hallucinate ANY technical information.
Every piece of information about architecture, code, systems, services, data models, and APIs must be:
- Verified by reading actual source code
- Confirmed via MCP tools (Jira, Confluence, Notion, Figma)
- Double-checked against the actual codebase
When Information is Unknown
Use these markers:
[UNKNOWN: Brief description of what's missing]
[NEEDS VERIFICATION: What needs to be checked and where]
[TBD: Decision pending - who needs to decide]
Before Writing Any Technical Detail
- Service names - search codebase, verify service exists
- API endpoints - read actual proto files or route definitions
- Database tables - find actual schema/migrations
- Data models - read actual model/entity files
- Business logic - read actual implementation code
- Dependencies - check package.json, build.gradle, requirements.txt
Workflow
Phase 1: Gather Context
Before drafting, collect and verify information from all available sources.
Step 1: Fetch External Context via MCP
Jira/Linear ticket (if provided):
- Fetch full ticket details: description, acceptance criteria, linked issues
- Walk the parent chain (epic, initiative) for broader context
- Check remote links for Confluence, Figma, Notion references
Confluence/Notion docs (if referenced):
- Fetch related docs, DACIs, existing tech notes
- Search for related TDDs in the same area
Figma designs (if linked):
- Get design context, component structure
- Capture screenshots for visual reference
Step 2: Explore Existing UI (if modifying an existing system)
If the feature modifies an existing interface, explore the current state:
- Navigate to the relevant pages
- Document current UI layout, components, user flows
- Note current limitations and UX issues
- Identify patterns to maintain or improve
Step 3: Analyze Codebase (required)
Read actual code before writing technical details:
- Search for affected services, read entry points
- Find existing data models, database migrations
- Check proto files for message definitions
- Understand current architecture: API routes, service communication
- Review related code for patterns
- Find similar features for reference
Step 4: Mark All Unknowns
After gathering, explicitly list what could NOT be verified:
- Missing schema information
- Unclear service boundaries
- Unconfirmed business logic
- Pending decisions
Phase 2: Draft Structure
The template below is a coverage checklist for the author, not a layout for the reader. It lists what a good TDD considers. It does not mean each heading gets three bullets and a code block. As you fill it, follow one rule above all others: write each section as prose a colleague could read aloud. Reach for a list or table only when the content is genuinely parallel and order-independent, a roster of endpoints, a t-shirt-size scale, a test matrix. The moment a "list" has bullets that depend on each other (this happens, then that, because of the other), it's a paragraph wearing a list costume. Write the paragraph.
Drop sections that don't apply rather than filling them with "N/A" noise. Mark ANY unverified information with the markers above.
# TDD: [Initiative Title]
**Author:** [Name]
**Team:** [Team/Squad Name]
**Status:** IN REVIEW | GO | NO GO
**Links:**
- Ticket: [link]
- Designs: [link]
- Docs: [link]
---
## Introduction
### Context
[1-2 paragraphs: What problem are we solving? Business-oriented terms.]
### Problem Statement (Current State)
[Current limitations, pain points, gaps.]
### Current Flow (if modifying existing system)
[Document the existing UI and user flow.]
### Proposed Solution
[2-3 sentences: How do you plan on solving it?]
### Long Term Vision
[How does this bring us closer to the team's long-term goals?]
---
## Decision Record
- **Driver:** [Author]
- **Approver:** [EM or senior IC with domain expertise]
- **Contributors:** [Your team, affected teams]
- **Informed:** [Engineering, PM, relevant stakeholders]
---
## Phase 1
### LOE (T-Shirt Size)
| Size | Time |
|------|------|
| XS | 1-2 days |
| S | 1 week |
| M | 2 weeks |
| L | 4 weeks |
| XL | 4+ weeks |
**This phase:** [SIZE]
### External dependencies & impact
[Prose. Who else is affected and how? Name the teams that depend on this or whose
systems you touch, the vendors (with cost) you're adding, the libraries or internal
services you'll lean on, and the consequence of each. A reader should learn not just
*that* Team X is involved but *what they have to do because of this work*.]
### Concepts
[Prose. Introduce each new idea or model and, crucially, how they relate. This is
where causality lives, "an Order owns many LineItems, and a LineItem can't outlive
its Order", so resist turning it into a glossary of disconnected terms.]
### System architecture, data model & APIs
[The technical heart. Walk the reader through the design as a story: the shape of the
data, what flows where, and the one or two decisions that everything else hangs on.
Lead with the decision and its reason, then show the artifact.
- **Data model:** Describe in prose what's changing and why, then show the actual
schema/migration. A reader should understand the *semantics* before they read the
column list. Reserve tables for the column inventory itself.
- **APIs:** Say what the client experiences differently and why you chose this shape,
then include the real endpoint/proto/GraphQL definition inline (verified against
source). A list is right for the *roster* of endpoints; prose is right for the *why*.
- **Business logic:** Explain non-trivial logic as a narrative of what happens and
when. Number the steps only when they're a true sequence. Call out load, response,
and caching considerations where they bite, not as a trailing bullet dump.]
### Testing
[One or two sentences on the testing strategy, what gives you confidence this is
correct, then the matrix. The table is genuinely parallel data, so a table earns its
place here. Follow `/flagrare:testing-philosophy`: integration-heavy, behavior over
implementation, and, for any user-facing flow, at least one end-to-end test of the
critical happy path (a browser journey, a running-service HTTP test, a CLI subprocess,
or a public-API consumer, depending on the surface). "Mostly integration" is not "skip
e2e."]
| Scenario | Type | Data Considerations |
|----------|------|---------------------|
| [scenario] | unit/integration/e2e/manual | [special data] |
**What won't be tested:** [Prose. State the exclusions and *why* they're acceptable.]
### Localization
[Prose, if applicable. Translation strategy and what triggers it.]
### PII
[Prose. New PII fields? PII persisted outside the primary database? Name the field and
the handling, not just "yes".]
### Observability & alerting
[Prose. What new signal will exist, and, more importantly, what question each
metric/dashboard/alert answers when something goes wrong.]
### Security
[Prose. Auth, permissions, and any vulnerability surface this opens or closes, with
the mitigation in the same breath.]
### Analytics
[Prose. New or modified events and where they land, framed by what decision the data
supports.]
### Rollout strategy
[Prose. Walk through how this reaches production, dry run, stealth mode, feature
flags, gradual percentages, as a sequence with reasons, not a checklist of toggles.
The reader should understand the *risk posture*, not just the mechanics.]
### Post-deployment monitoring
[Prose. How will we know it's working? Name the queries/dashboards and who watches
them, so a reader could verify success themselves.]
### Risks
[Prose. Each known risk paired with its mitigation in the same sentence or paragraph.
A risk listed without a mitigation reads as an unanswered worry.]
### Documentation
[Prose. What docs need to exist or change, and for whom.]
### Open questions
- [ ] [Genuinely open question, a real fork, not a placeholder]
- [ ] [Another, if any]
### Alternative solutions
[Prose. The options you considered and *why you discarded them*. This is one of the
most-read sections in any TDD, reviewers come here first to check you weren't naive,
so give each alternative a fair sentence and an honest reason it lost.]
---
## Phase 2
[If applicable]
## Future Work
[Deferred items]
Verification (a drafting artifact, NOT a section of the shipped TDD). While drafting, keep a private checklist of what you verified and what you couldn't, sources checked (ticket, docs, code files, proto, schemas) and information gaps ([UNKNOWN], [NEEDS VERIFICATION], [TBD]). This keeps you honest and tells you what's still unproven. It is scaffolding for the author, so strip it from the document you present. A reader answering what/why/how does not need your sources list; genuinely-open decisions live in Open Questions, not a gaps dump. (Earlier versions of this skill shipped a "Verification Summary" section in the doc, that was wrong; it is a checklist you run, not content you publish.)
Phase 3: Writing the document
This is where TDDs live or die. You've gathered verified facts; now you have to turn
them into a document a busy senior engineer will actually read. The craft layer is in
/flagrare:write-docs, invoke it, but here is how its principles land in a TDD.
Lead each section with the reader's situation, not the section's topic. Compare
"This section describes the data model changes." (metadata, it has done zero work)
with "Orders currently can't be partially refunded because the schema models a refund
as a single boolean. We're replacing that with a refund ledger." The second sentence
tells the reader where they are and where you're taking them. The opening line of every
section should state a conclusion, name a problem, or ask a question.
Prose carries causality; lists flatten it. A TDD is mostly because and so that
and before, exactly the relationships a bullet list erases. When you catch yourself
writing a list whose items depend on each other, you've found a paragraph. Keep lists
for the genuinely parallel: a roster of endpoints, a t-shirt scale, a test matrix, a
checklist of sources verified. (This is the single biggest lever against the "medicine
sheet" feel the operator complained about.)
Put context at the point of need. The first time a reader meets client_secret or
OrderProjection or a service name, define it inline in the same sentence, not in a
glossary, not three sections up. The reader who knows skips it; the reader who doesn't
is rescued exactly where they got lost.
Voice and tone. Use "We will…", conversational but precise. Be direct about scope
and limitations; a TDD that hides its gaps loses reviewer trust faster than one that
names them. Keep the voice the same throughout; let the tone tighten in the dense
technical sections and warm slightly in the context and vision.
Diagrams. A good diagram replaces three paragraphs of prose, so prefer Mermaid for
any non-trivial flow or architecture:
flowchart TD
A[User Action] --> B{Decision Point}
B -->|Option 1| C[Service A]
B -->|Option 2| D[Service B]
C --> E[Database]
D --> E
Color-code consistently, yellow/maize for existing architecture, blue for new, so a
reader can see the boundary of the change at a glance. When you show alternatives in a
diagram, mark the chosen path. Reference exact, verified file paths and show directory
layout for new components, but do it in service of the narrative, not as a standalone
inventory.
Phase 4: Review Before Presenting
Do not end your turn the moment the draft is assembled. A complete-looking TDD reads as "done," but the craft pass (Phase 3 / /flagrare:write-docs) and this review still have to happen before you present it. Drafting and stopping is a stall, continue through review in the same turn. (Same pattern as docs/research/2026-06-11-claude-code-goal-anti-stall.md.)
Run the write-docs self-check, read the document aloud; sentences that choke on the
tongue usually nominalize a verb or rely on a bullet list to carry a relationship the
prose should have carried. For a finished draft heading to reviewers, /flagrare:editorial-pass
is the full version of this review: a whole-document read (throughline, pacing, consistency)
plus the noise hunt the Restraint rules below only summarize. Then confirm:
- Could a new team member read this start to finish and understand what's proposed, why,
and how? (If they'd have to reassemble the argument from bullets, rewrite those bullets
as prose.)
- Is every technical claim backed by code you actually read?
- Are all unknowns marked explicitly with the markers from the No Assumptions Policy?
- Did you run the verification checklist (every claim traced to a source), and then strip that checklist from the presented doc?
- Are diagrams clear, accurate, and consistently color-coded?
- Does every section's opening sentence do work, rather than describe what the section is
about?
Once the review passes, present the TDD and close with a tool, not prose: issue an AskUserQuestion (e.g. Looks good / Revise a section / Mark ready for review) so the turn ends on a clear next step rather than trailing off after a long document.
Key Principles
What to Focus On
- Context first - clearly communicate the problem and intended impact
- Data models - changes are hard once in production
- API structure - specs that are hard to change after shipping
- Data loading - network activity, especially for mobile users
- Novel architecture - justify any new patterns/technologies
- Rollout and feedback - how to validate success
Data Model Best Practices
- Self-documenting: convey semantics without needing app code
- Use descriptive enums over numeric codes
- Explicit foreign keys for relationships
API Best Practices
- Optimize ergonomics for clients
- Start from user problem, work backward to data
- Minimize client-side data transformation
Anti-patterns
Craft (the "medicine sheet"):
- Don't flatten sections into terse bullets where prose should carry the causality. If a "list" has items that depend on each other, it's a paragraph.
- Don't open a section with metadata ("This section describes…"). Open with the reader's situation, a conclusion, or a question.
- Don't defer every definition to a glossary. Define terms inline at first use.
- Don't ship a document that can't be read start to finish. A reader should never have to reassemble your argument from disconnected bullets.
Verification:
- Don't invent service names that might not exist
- Don't assume API payload structures
- Don't guess database column names or types
- Don't fabricate code examples without reading actual code
- Don't assume how systems communicate without verification
- Don't skip the verification checklist while drafting (but don't ship it as a section either, see Restraint)
- Don't present the TDD without marking gaps, real open decisions go in Open Questions
Restraint (the TDD answers three questions: what are we solving, why, and how, for a reader who has never seen the work). Everything that doesn't serve those is noise that raises the reader's risk and mental load):
- Don't document what you resolved. A decision that's settled is not a question, so it doesn't belong in Open Questions, Risks, or Key Decisions. State the resolved choice once where it's relevant (a flow, the rollout) and delete the deliberation. Open Questions holds only genuinely-open forks that still need an owner's call.
- Don't narrate provenance. How you learned a fact, "X told Y in Slack," "a Notion doc pointed here," "confirmed in the 2026-06-16 thread," "verified at lines 49-52", is research trail, not design. State the fact; cite a person only when their sign-off is the pending action.
- Don't leave self-referential scaffolding in the doc: no "[NEEDS VERIFICATION]" left in once verified, no "this corrects an earlier note," no "as we found above." The reader doesn't care about the draft's history.
- Don't keep a section alive once its content collapses to nothing. If every item under a heading got resolved, cut the heading. Prefer cutting to writing "N/A".
- Don't carry forward the exploration's volume. A TDD is the distilled conclusion, not the log of getting there. When a later pass resolves things, the doc should get shorter, not accrete "resolved:" annotations. If unsure whether a sentence helps the reader answer what/why/how, cut it.
- Two registers, chosen by what the reader does with the passage. Key Decisions and Open Questions get full contextualization, the reader must judge, so give the ingredients. Background sections (context, intros, transitions) get assertion only: no illustrations of concepts the audience already holds, no importance-selling (a gated, staffed TDD does not need to argue that its work matters), no step-by-step derivations where the conclusion suffices, the mechanism is stated once where it is load-bearing. Test each background sentence: deleted, does the reader decide wrongly, or just feel less escorted? Only the first earns the words.
- Clarity without long-windedness, the two failure modes are opposite and both cost the reader. Compression is the sneakier one: a sentence that packs four ideas into one breath, or a paragraph that is six clauses doing five jobs, is correct but forces the reader to unpack it. The densest, most important paragraph (often the core risk) is the one most likely to be over-compressed, so give it the most room: break it into beats, one idea per sentence, ordered as cause then effect. Define every piece of internal jargon at first use, inline, in the same sentence ("a background job pulls the POS menu and reconciles our copy against it" before you ever lean on the bare word "reconcile"); assume a new hire who knows the stack generically but not this corner of it. Cut phrases that sound technical but carry no information for the reader ("holds by construction", "by design", "leverage", "it should be noted that"); replace them with the concrete thing they gesture at, or delete them. The test is the read-aloud: any clause you can't say in one breath, or any term a new hire would have to stop and look up, is a rewrite. Note this is the same craft
/flagrare:write-docs teaches; restraint here means applying it, not just trimming length.
