name: decision-create
description: Create a Decision Group file in wiki/work/decisions/ with one or more proposed decisions, table-only comparisons, and mermaid flowcharts
category: planning
model: claude-opus-4-8
effort: high
argument-hint: <decision topic, group title, or "group: " for an explicit multi-decision file>
disable-model-invocation: false
user-invocable: true
Always obey .docs/guides/mcp-tools.md. Read it now if not already in context.
Run /primer first if you have not already this session.
Run /research <topic> on the decision topic BEFORE drafting any decision block. Do not skip this step.
Create Decision
Create a new Decision Record file. A decision file is a Decision Group — one or more architecturally significant decisions that share context and are documented together. Each decision tracks its own Status, Date, Deciders, Tags, and supersession independently.
Decisions follow a hybrid of Michael Nygard's 2011 template and MADR 4.0 (2024), extended with this project's per-decision status model. They are append-only history: when a later decision overrides an earlier one, write a new decision (in a new file or as a new D* block in an existing related file) that supersedes the old one — never rewrite the past.
Read wiki/work/decisions/lifecycle.md first — it defines the Decision Group model, the supersession rule, the relationship graph, and the index format this command must respect.
Topic: $ARGUMENTS
When to write a Decision
See wiki/work/decisions/lifecycle.md for the canonical "When to Write" / "When NOT to Write" tables. In short: write a decision when at least one decision in the topic affects structure, NFRs, dependencies, interfaces, or construction techniques — and is hard to reverse. Skip for bug fixes, refactors, naming, or trivially reversible choices.
If the decision does not clear that bar, stop and tell the user — don't pad the log.
Instructions
Step 1: Parse the topic and recall project context
- Extract the core topic from
$ARGUMENTS. If the topic is too vague (e.g. "auth"), use AskUserQuestion to narrow it.
- Decide single-decision vs Decision Group — see Step 1.5.
- Recall Serena memories that may inform any decision in the group:
mcp__serena__list_memories (filter by topic if applicable)mcp__serena__read_memory for any architecture/gotcha/integration memory that touches the topic
- Read project context:
CLAUDE.md, PROJECT_STATUS.md (if present), and wiki/work/decisions/lifecycle.md (the index) plus a sample of recent decisions to learn local conventions.
Step 1.5: Single-decision vs Decision Group triage
Ask the user (or decide from $ARGUMENTS) whether the file will hold one decision or several. Use the lifecycle doc's "When to Group vs Split" table:
| Signal | Group into one file | Split (this command writes one decision now; user invokes /decision-create again later) |
|---|
| Shared context, drivers, deciders | ✅ Group | — |
| Tightly coupled (one constrains the next) | ✅ Group | — |
| Different decision areas with little overlap | — | ✅ Split |
| Different audiences / timelines | — | ✅ Split |
| One is much larger than the others | — | ✅ Split |
Default to single-decision unless $ARGUMENTS starts with group: or the user explicitly lists multiple decisions. When in doubt, ask via AskUserQuestion with the table above visible. If grouping, request a list of the decisions (one short title each, 2-6 entries).
Step 2: Locate the decisions directory and assign a number
-
Find or create the decisions directory at wiki/work/decisions/.
If it does not exist, create it with a lifecycle.md index. Confirm via AskUserQuestion before creating a new directory.
-
Read wiki/work/decisions/lifecycle.md to learn the current chain, decision-area conventions, and relationship graph.
-
Determine the next file number: scan existing files via mcp__serena__list_dir, find the highest 4-digit prefix, increment by 1. First decision is 0001.
-
Derive the file slug — names the group, not a single decision. Lowercase, dash-separated, ≤ 60 chars:
- Single decision "session storage strategy" →
DEC-0007-session-storage-strategy.md
- Group covering session lifetime + storage + invalidation →
DEC-0007-session-management.md
Step 2.5: Detect existing decision area (per-decision supersession check)
Before drafting, determine whether each planned decision enters a decision area that already has an accepted decision somewhere in the log. The supersession rule operates per decision: each D* block must check independently.
For each planned decision in the group:
-
Read every existing decision file in wiki/work/decisions/ (use mcp__serena__list_dir, then Read each; exclude lifecycle.md)
-
For each existing decision block (D1, D2, …) with status accepted, compare:
| Tier | Signal | Detection |
|---|
| 1 | Tag overlap | ≥ 1 shared tag with the planned decision |
| 2 | Decision-title noun overlap | Substantial overlap |
| 3 | File-slug stem overlap | Only relevant when both are single-decision files |
-
Build a candidate-supersession table per planned decision:
| Planned Decision | Existing Decision | Overlap Tier | Likely Same Area? |
|---|
| (group D1) Session storage backend | DEC-0003#D2 | Tag | Yes |
| (group D2) Session lifetime | (none) | — | No |
-
If any row is Yes/Maybe, surface to the user in Step 4 and offer to set Supersedes on that block. Otherwise the block enters a fresh area.
This check is mandatory — it prevents accidentally creating two parallel accepted decisions in the same area, which breaks the chain.
Step 3: Research each decision
Invoke /research <topic> for each decision in the group — call the research skill directly, do not substitute with ad-hoc web searches or Serena lookups alone. Do not proceed to Step 4 until research for every planned decision is complete. Each invocation must produce concrete inputs:
| Decision Section (per decision) | Research must produce |
|---|
| Context | Code state relevant to this decision; constraints from prior decisions; business/compliance forces |
| Decision Drivers | Concrete forces specific to this decision: performance budgets, team skill, security, ecosystem fit, deadline |
| Considered Options | ≥ 2 viable options. Use Context7 for library/framework docs, Brave Search (1/sec, sequential) for ecosystem patterns |
| Pros & Cons | Symmetric assessment per option |
| Consequences | Real follow-on: new dependencies, migration cost, what becomes harder/easier |
| Validation | Measurable signal + threshold + timeframe (E-C-A-D-R "R") |
Decisions in the same group may share evidence sources (Shared Context handles that), but each block needs its own drivers, options, and consequences.
If any block lacks evidence for an option or driver, drop the option/driver from that block rather than fabricating. Note the gap in the Step 9 report.
Step 4: Clarify with the user (mandatory)
Use AskUserQuestion to resolve, for each decision in the group, at minimum:
| # | Question | Notes |
|---|
| 1 | Status to assign at draft time | Strongly prefer proposed (default). Each decision can be a different status if needed |
| 2 | Recommended option | Present research-backed recommendation with the comparison table; user confirms or overrides |
| 3 | Deciders | Who owns this decision (may differ per block); defaults to current git user |
| 4 | Tags (mandatory, non-empty) | 1-3 short tags identifying the decision area for this block. Required for future supersession detection |
| 5 | Supersedes | If Step 2.5 surfaced candidates for this block, ask which (if any) it supersedes. DEC-NNNN#DM, none, or "decide at finalize time" |
Refuse to write the file if any decision block has empty Tags.
For Decision Groups, ask all decisions' questions in a single AskUserQuestion batch when feasible to minimize round-trips, but keep questions per-decision (not collapsed into "the group's deciders").
Step 5: Generate the decision file
Write the decision to wiki/work/decisions/DEC-NNNN-<group-slug>.md using this template. The two formatting rules — tables only, mermaid for flows — are mandatory in every decision block.
# DEC-NNNN: <Decision-Group Title>
> Decision Group covering <area-1>, <area-2>, <area-3>.
- **File created**: YYYY-MM-DD
- **Last updated**: YYYY-MM-DD
- **Tags (group)**: <umbrella tags shared by all decisions, optional>
## Shared Context
<Context that applies to every decision in this group. Avoid restating in each block.>
<Link to relevant code areas (repo-relative paths), prior decisions ([[DEC-NNNN-slug|DEC-NNNN#DM: title]]), or external constraints.>
---
## D1. <Short Title — Verb Phrase>
- **Status**: <proposed | accepted | deprecated | superseded by DEC-XX#DY>
- **Date**: YYYY-MM-DD
- **Deciders**: <names or roles>
- **Consulted**: <names or roles, optional>
- **Informed**: <names or roles, optional>
- **Supersedes**: <DEC-MMMM#DK | none>
- **Tags**: <area-1>, <area-2>
### Context (decision-specific)
<Anything that applies to D1 but not to siblings. Reference Shared Context when possible.>
### Decision Drivers
| # | Driver | Why it matters |
|---|--------|----------------|
| 1 | <e.g. p99 latency budget < 200ms> | <consequence if violated> |
| 2 | <...> | <...> |
### Considered Options
| Option | One-line summary |
|--------|------------------|
| **A. <name>** | <≤ 15 words> |
| **B. <name>** | <≤ 15 words> |
| **C. <name>** | <≤ 15 words> |
### Option Comparison
| Criterion | Option A | Option B | Option C |
|-----------|----------|----------|----------|
| Meets driver 1 | ✅ | ⚠️ | ❌ |
| Meets driver 2 | ✅ | ✅ | ❌ |
| Implementation cost | Low | Med | High |
| Operational cost | Low | Low | High |
| Reversibility | Easy | Hard | Hard |
### Trade-off Detail per Option
#### Option A: <name>
| Aspect | Assessment |
|--------|------------|
| Pros | <one cell, comma-separated or short prose> |
| Cons | <one cell> |
| Risks | <known unknowns> |
| Exit cost | <how hard to walk away later> |
#### Option B: <name>
| Aspect | Assessment |
|--------|------------|
| Pros | … |
| Cons | … |
| Risks | … |
| Exit cost | … |
#### Option C: <name>
| Aspect | Assessment |
|--------|------------|
| Pros | … |
| Cons | … |
| Risks | … |
| Exit cost | … |
### Decision Outcome
**Chosen option**: **<Option X — name>**, because <one-sentence justification anchored to the highest-priority driver>.
### Decision Flow
<Include a mermaid block when the decision involves a flow, sequence, or state transition. Skip only for purely static choices.>
```mermaid
flowchart LR
...
Consequences
| Type | Consequence |
|---|
| ✅ Positive | <e.g. eliminates current bottleneck> |
| ⚠️ Negative | <e.g. adds new operational dependency> |
| 🔁 Follow-up | <e.g. write task: instrument metric> |
Validation
| Signal | Threshold | When measured |
|---|
| <e.g. p99 latency> | < 200ms | 30 days post-deploy |
Links
- Related decisions: [[DEC-NNNN-slug|DEC-NNNN#DM: title]], [[DEC-NNNN-slug|DEC-NNNN#DM: title]]
- Supersedes: [[DEC-MMMM-slug|DEC-MMMM#DK: title]] (if applicable)
- Source task(s):
wiki/work/tasks/TASK-NNN-slug.md
D2. <Short Title of Decision 2>
- Status:
- Date: ...
- Deciders: ...
- Consulted: ...
- Informed: ...
- Supersedes: ...
- Tags: ...
Context (decision-specific)
...
D3. ...
**Single-decision files** simply have one `D1` block and no `D2`/`D3`. The `Shared Context` may be merged into D1's `### Context` or omitted in that case.
### Step 6: Maintain the Decision Index
After writing the decision file:
1. **Append a row to `wiki/work/decisions/index.md`**: `- [DEC-NNNN — Title](DEC-NNNN-slug.md) — one-line summary · D1 proposed`
2. **Add one row to the Index per decision in the new file** in `wiki/work/decisions/lifecycle.md` (a 3-decision file adds 3 rows). Columns: `File`, `Decision`, `Title`, `Decision Area`, `Status`, `Date`, `Deciders`, `Supersedes`, `Superseded By`. See lifecycle doc's index column spec.
3. **If any decision was created with status `accepted` AND supersedes an existing decision** (rare — Step 4 #1 was set to `accepted`), apply the lifecycle doc's two-block cross-reference rule per affected pair. The atomic update covers:
| Site | Edit |
|------|------|
| Superseded decision block | `Status: accepted` → `Status: superseded by DEC-N#DX`; insert callout `> **Superseded by [[DEC-NNNN-slug\|DEC-N#DX: title]]** on YYYY-MM-DD` directly under that decision's H2 |
| Superseded decision's siblings | **Untouched** — sibling decisions retain their own status |
| New decision block | `Supersedes` metadata set; `### Links` references the superseded `[[DEC-MMMM-slug\|DEC-MMMM#DK: title]]` |
| Index | Two rows updated (the superseder's row gets `Supersedes`, the superseded's row gets `Status` and `Superseded By`) |
| Relationship Graph | New node added; supersession edge drawn; superseded node's class flipped |
**Default to `proposed`.** Supersession at create-time is rare; `/decision-finalize` enforces it atomically when the user is ready.
4. **Do not delete or mutate any existing index row** beyond what supersession requires. The index is append-mostly.
### Step 7: Append to wiki/log.md
Append an entry to `wiki/log.md`:
[YYYY-MM-DD] decision-create | DEC-NNNN
Use `Edit` to append — never `echo >>`.
### Step 8: Cross-link from related artifacts
For any decision in the group that was prompted by an active task or feature work, cross-link both directions:
- In the source task file (`wiki/work/tasks/TASK-NNN-slug.md` if it exists), append: `**Decision**: [[DEC-NNNN-slug|DEC-NNNN#DM: title]]`
- In the decision's `### Links` section, list the source task path
Use `Read` then `Edit` for these updates — **never** `echo >>` or `sed`.
### Step 9: Update memories
If a decision in the group establishes a non-obvious pattern, integration constraint, or known gotcha:
- `mcp__serena__write_memory` with a topic-hierarchical name (e.g. `architecture/data-layer/cache-strategy`)
- The memory body should reference the specific decision (`DEC-NNNN#DM`), not just the file
Skip for self-documenting decisions.
### Step 10: Report completion
Print a tabular summary:
| Field | Value |
|-------|-------|
| File path | `wiki/work/decisions/DEC-NNNN-slug.md` |
| Decisions in this file | `D1: <title>`, `D2: <title>`, … |
| Status assigned per decision | per-decision list |
| Supersession queued for finalize? | yes (per decision) / no |
| Index updated | yes (N rows added) |
| Graph updated | yes / no (only if status=accepted at draft) |
| Log entry appended | yes |
| Suggested next steps | `/decision-finalize <file>#D1`, `/decision-finalize <file>#D2`, … |
If any decision was dropped due to lack of evidence, note it in a separate "Gaps" section.
---
## Output Formatting Rules (mandatory — these override any default style)
1. **All comparisons are tables.** No bullet-list pros/cons, no bullet-list option summaries. If you find yourself writing `- Good, because ...` / `- Bad, because ...`, stop and convert to a table row.
2. **Use mermaid for any flow, sequence, or before/after architecture.** Static choices may skip. When in doubt, include — diagrams age well.
3. **One decision per `D*` block.** A single `D*` block must NOT cover two decisions. Split into `D1` and `D2`.
4. **Stable decision IDs.** Once a decision is `D2`, it stays `D2` forever — even if `D1` is later marked `deprecated`. Never renumber siblings.
5. **Present tense, full sentences.** "We will use Redis as the session cache." Not "Redis chosen" or "Going with Redis I think".
6. **Immutable per decision once accepted.** Never edit an accepted decision block's content — create a new decision (in this file as `Dnext` or in a new file) that supersedes it. Status changes (accepted → superseded) and the supersession callout are the only allowed in-place edits.
---
## CRITICAL Rules
1. **Maximum 3 sub-processes at a time** if delegating research steps
2. **Always terminate processes when done** (dev servers, type checkers)
3. **Never invent options or consequences** — every option in a block's tables must be backed by Step 3 research; every consequence must be a real implication
4. **Tables not bullets** for every comparison — hard rule, not a preference
5. **Mermaid for flows** — include a diagram unless the decision is purely static
6. **Per-decision metadata is mandatory** — `Status`, `Date`, `Deciders`, `Tags` cannot be empty on any block; `Tags` non-empty is required for supersession detection
7. **Decisions in a group are independent units** — never assume sibling status, sibling deciders, or sibling supersession on behalf of a block
