write-milestone-brief - SKILL.md Agent Skill

name: write-milestone-brief description: Synthesize the current conversation into a milestone brief (PRD). Writes to `M###-CONTEXT.md` by default, or files a GitHub issue only with explicit user confirmation. Use when asked to "turn this into a PRD", "draft a milestone brief", "capture this context", "write it up", or when enough has been discussed to commit the plan to paper. Does not interview — it synthesizes what is already known.

Take everything established in the current conversation (plus repo reality) and produce a milestone brief that a future agent can execute from with zero additional context. The output is a populated `M###-CONTEXT.md`, matching the template at `src/resources/extensions/gsd/templates/context.md`. Optionally, with explicit confirmation, also a GitHub issue. This skill runs at the end of a discussion phase, after enough grilling/design has happened that the plan has stabilized. It does NOT interview — use the `grill-me` skill for that. This skill collapses what is already on the page into a durable artifact.

Typical invocation points:

User says "capture this" or "write it up" after a planning discussion
End of a discuss phase before moving to plan
User wants to hand the work off to another agent or a human teammate

**SYNTHESIZE, DO NOT RE-INTERVIEW.** Use what is already in this conversation. If a decision is genuinely missing, note it in `Open Questions` — do not relitigate.

REAL OUTCOMES, NOT TASKS. M###-CONTEXT.md describes what the user can do when the milestone ships, what scenarios must pass for "done," and what architectural decisions were made. It does NOT list tasks or implementation steps — those belong in M###-ROADMAP.md and S##-PLAN.md.

NO FILE PATHS OR LINE NUMBERS IN THE BRIEF. Those go stale. Describe modules, interfaces, and behaviors. The roadmap and plans can cite code locations; the brief should outlive refactors.

Step 1: Locate the target

Find the active milestone:

Read .gsd/STATE.md — it names the active milestone.
If no milestone is active, ask the user whether this is a new milestone (create directory + files) or appending to an existing one.
For a new milestone, use gsd_milestone_new or the /gsd new-milestone flow — do not create directories by hand.

Step 2: Read the template

Read src/resources/extensions/gsd/templates/context.md (the full path is shown in the templatesDir system-prompt field). Match its structure exactly — parsers and downstream prompts depend on the headings.

Step 3: Sketch the modules

Before filling the template, sketch the major modules the milestone will build or modify. Actively look for opportunities to extract deep modules — ones that encapsulate a lot of functionality in a simple, testable interface that rarely changes. A deep module is more testable, more AI-navigable, and survives refactors.

If the modules are non-obvious, offer the user a brief check-in: "The milestone touches modules A, B, and C. A can be extracted as a deep module that hides X. Does that match your thinking?" One round. Then proceed.

Step 4: Fill the template

Populate M###-CONTEXT.md using the template. Key sections:

Project Description — one paragraph, plain English, what this milestone is.
Why This Milestone — the problem it solves and why now, from the user's perspective.
User-Visible Outcome — literal user actions in the real environment. "User can complete the import flow end-to-end" not "Adds import API."
Completion Class — contract / integration / operational. Be honest about what level of proof is required.
Final Integrated Acceptance — the real end-to-end scenarios that must pass. Name things that cannot be simulated.
Architectural Decisions — one ### Decision Title block per decision, with rationale and alternatives considered. If there were no architectural decisions, write "None — straightforward execution" and skip the subsections.
Error Handling Strategy — approach for failures, edge cases, and error propagation. Include retry policies, fallback behaviors, and user-facing error messages where relevant.
Risks and Unknowns — only real ones. Do not invent risks.
Existing Codebase / Prior Art — module names, not line numbers. Brief description of how each relates.
Relevant Requirements — requirement IDs this milestone advances and how.
Scope — In Scope / Out of Scope / Non-Goals. Be specific about tempting adjacent work that is out.
Technical Constraints — binding constraints (runtime, platform, dependencies, performance budgets).
Integration Points — external systems/services this milestone touches and how.
Testing Requirements — test types (unit, integration, e2e), coverage expectations, and specific scenarios that must pass.
Acceptance Criteria — per-slice, testable criteria gathered during discussion.
Open Questions — anything material that is genuinely unresolved. Note current thinking so future agents have a starting point.

The template headings above are mandatory — do not omit, rename, or reorder them. The exact required order matches src/resources/extensions/gsd/templates/context.md: Project Description → Why This Milestone → User-Visible Outcome → Completion Class → Final Integrated Acceptance → Architectural Decisions → Error Handling Strategy → Risks and Unknowns → Existing Codebase / Prior Art → Relevant Requirements → Scope → Technical Constraints → Integration Points → Testing Requirements → Acceptance Criteria → Open Questions.

Step 5: Write it

Use the write tool to create or overwrite .gsd/milestones/<MID>/<MID>-CONTEXT.md. Do not ask for approval of the file contents before writing — the user will see the rendered file and can edit directly.

Then append a one-line summary to .gsd/DECISIONS.md for any genuinely hard-to-reverse architectural decision: - YYYY-MM-DD [MID]: <decision> — <one-line rationale>.

Step 6: Offer next steps

After writing, offer the user (do not auto-execute):

Proceed to planning — run /gsd dispatch plan to generate M###-ROADMAP.md from this brief.
File as GitHub issue — use the active GitHub issue-write tool if one is available, using its exact active tool name. Requires explicit "yes" per the outward-action rule. Use the PRD template below for the body.
Iterate on the brief — if something feels wrong, run grill-me to stress-test before moving on.

GitHub issue body template (only if user chooses option 2)

## Problem Statement
<user-perspective problem — one paragraph>

## Solution
<user-perspective solution — one paragraph>

## User Stories
1. As a <actor>, I want <feature>, so that <benefit>.
2. ...
<cover all aspects; be exhaustive>

## Architectural Decisions
<list of decisions with rationale — copy from M###-CONTEXT.md>

## Testing Strategy
<what a good test looks like here: external behavior only, which modules get tested, prior art>

## Out of Scope
<things explicitly not covered>

## Further Notes
<any extra context>

---
See `.gsd/milestones/<MID>/<MID>-CONTEXT.md` for the full brief.

Re-interviewing the user. Use context, not fresh questions. If something is missing, put it in Open Questions.
Listing implementation tasks. The brief describes outcomes and decisions — the roadmap and plans describe tasks.
Citing specific line numbers. They rot. Cite modules or interfaces.
Inventing risks or decisions. If the discussion didn't surface a real architectural choice, the Architectural Decisions section is short — that's fine.
Auto-filing the GitHub issue. Any outward-facing action needs explicit confirmation.

M###-CONTEXT.md exists and follows the template structure.
User-Visible Outcome section names literal user actions, not internal tasks.
Completion Class is honest — no "just unit tests" when the milestone demands live integration.
Every architectural decision has a rationale and named alternatives.
Open Questions captures genuinely unresolved items — not decisions the user already made.
.gsd/DECISIONS.md has a dated one-liner for any hard-to-reverse decision.