| name | draft-setup |
| description | Context initialization. Runs a short conversational interview to derive product and team context, then writes structured files to the workspace so every future session starts grounded. Run once at onboarding, or anytime to refresh context after a significant shift.
|
/draft:setup — Context Initialization
You are loading the shared context workspace with information about what the user is building, who they're building it for, and how they work. This interview takes about 3–5 minutes.
The goal is derived context, not transcribed answers — synthesize what the user says into structured knowledge, don't just copy their words.
Before starting
Profile detection (first-time users only)
Check whether a profile is already configured:
python3 -c "
from pathlib import Path
active_file = Path.home() / '.draft' / 'active-profile'
if active_file.exists() and active_file.read_text().strip():
print('has_profile')
else:
print('no_profile')
"
-
If has_profile: skip this section and go directly to the context check below.
-
If no_profile (first-time user): ask for a profile name before the interview:
Use the AskUserQuestion tool to ask:
"What should we call this workspace? (Default: <cwd-slug> — based on your current directory)"
Derive the CWD slug first, then ask:
python3 -c "
import os, re
cwd = os.path.basename(os.getcwd())
slug = cwd.lower()
slug = re.sub(r'[\s_]+', '-', slug)
slug = re.sub(r'[^a-z0-9-]', '', slug)
slug = slug[:40].strip('-') or 'default'
print(slug)
"
Wait for the user's response. If they accept the default, use the CWD slug. Otherwise use their typed name (apply the same slug transformation).
Then create the profile directory and set active-profile:
python3 - <<'PYEOF'
import shutil
from pathlib import Path
name = "<chosen-slug>"
draft_global = Path.home() / ".draft"
base = draft_global / "workspaces" / name
personal = draft_global / "personal"
old_ws = draft_global / "workspace"
migrating = old_ws.exists() and not old_ws.is_symlink()
for subdir in [
base / "context" / "company" / "log",
base / "context" / "product" / "log",
base / "context" / "team" / "log",
base / "context" / "priorities" / "log",
base / "context" / "decisions",
base / "docs",
base / "config",
]:
subdir.mkdir(parents=True, exist_ok=True)
if migrating:
migrated = []
old_ctx = old_ws / "context"
if old_ctx.exists():
for dim_dir in old_ctx.iterdir():
if not dim_dir.is_dir():
continue
dim = dim_dir.name
new_dim = base / "context" / dim
new_dim.mkdir(parents=True, exist_ok=True)
idx = dim_dir / "index.md"
if idx.exists():
shutil.copy2(idx, new_dim / "index.md")
migrated.append(f"context/{dim}/index.md")
log_dir = dim_dir / "log"
if log_dir.exists():
new_log = new_dim / "log"
new_log.mkdir(parents=True, exist_ok=True)
for f in log_dir.iterdir():
if f.is_file():
shutil.copy2(f, new_log / f.name)
migrated.append(f"context/{dim}/log/ ({len(list(log_dir.iterdir()))} entries)")
tensions = old_ctx / "tensions.md"
if tensions.exists():
shutil.copy2(tensions, base / "context" / "tensions.md")
migrated.append("context/tensions.md")
old_decisions = old_ws / "context" / "decisions"
if old_decisions.exists():
new_decisions = base / "context" / "decisions"
for f in old_decisions.iterdir():
if f.is_file():
shutil.copy2(f, new_decisions / f.name)
migrated.append(f"context/decisions/ ({len(list(old_decisions.iterdir()))} files)")
old_research = old_ws / "context" / "research"
if old_research.exists():
new_research = base / "context" / "research"
new_research.mkdir(parents=True, exist_ok=True)
for f in old_research.iterdir():
if f.is_file():
shutil.copy2(f, new_research / f.name)
migrated.append(f"context/research/ ({len(list(old_research.iterdir()))} files)")
old_docs = old_ws / "docs"
if old_docs.exists():
for item in old_docs.rglob("*"):
if item.is_file() and item.name not in (".DS_Store", ".gitkeep"):
rel = item.relative_to(old_docs)
dest = base / "docs" / rel
dest.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(item, dest)
migrated.append("docs/")
old_config = old_ws / "config"
if old_config.exists():
for f in old_config.iterdir():
if f.is_file():
shutil.copy2(f, base / "config" / f.name)
migrated.append("config/")
old_skills = old_ws / ".claude" / "skills"
if old_skills.exists():
new_skills = base / ".claude" / "skills"
shutil.copytree(old_skills, new_skills, dirs_exist_ok=True)
migrated.append(".claude/skills/")
old_brainstorm = old_ws / "brainstorm-sessions"
if old_brainstorm.exists():
new_brainstorm = base / "brainstorm-sessions"
shutil.copytree(old_brainstorm, new_brainstorm, dirs_exist_ok=True)
migrated.append("brainstorm-sessions/")
print(f"Migrated from ~/.draft/workspace/ to profile '{name}':")
for item in migrated:
print(f" ✓ {item}")
else:
for dim in ["company", "product", "team", "priorities"]:
idx = base / "context" / dim / "index.md"
idx.write_text(f"---\nname: {dim}\ndescription: >\n No information recorded yet.\nlast_updated: \"\"\nsource: \"\"\n---\n")
(base / "context" / "tensions.md").write_text(
"# Tensions\n\nActive contradictions and inconsistencies noticed across context dimensions.\n"
)
(base / "docs" / ".gitkeep").touch()
print(f"Profile '{name}' created (fresh workspace).")
personal.mkdir(parents=True, exist_ok=True)
(personal / "user").mkdir(parents=True, exist_ok=True)
(personal / "wip").mkdir(parents=True, exist_ok=True)
def has_real_content(path):
if not path.exists(): return False
text = path.read_text()
return "No information recorded yet" not in text and len(text.strip()) > 100
old_personal = old_ws / "personal" if migrating else None
old_mem = old_personal / "memory.md" if old_personal else None
old_user = old_personal / "user" / "index.md" if old_personal else None
mem_dest = personal / "memory.md"
user_dest = personal / "user" / "index.md"
if old_mem and has_real_content(old_mem) and not has_real_content(mem_dest):
shutil.copy2(old_mem, mem_dest)
print(" ✓ personal/memory.md migrated to ~/.draft/personal/")
elif not mem_dest.exists():
mem_dest.write_text(
"---\nname: memory\ndescription: Vocabulary, working preferences, and non-obvious patterns.\nlast_updated: \"\"\nsource: \"\"\n---\n\n## Vocabulary\n\n## Preferences\n\n## Goals\n\n## Patterns\n"
)
if old_user and has_real_content(old_user) and not has_real_content(user_dest):
shutil.copy2(old_user, user_dest)
print(" ✓ personal/user/index.md migrated to ~/.draft/personal/")
elif not user_dest.exists():
user_dest.write_text(
"---\nname: user\ndescription: >\n No information recorded yet.\nlast_updated: \"\"\nsource: \"\"\n---\n"
)
(draft_global / "active-profile").write_text(name + "\n")
print(f"\nProfile '{name}' is now active.")
print(f"Path: ~/.draft/workspaces/{name}/")
PYEOF
Update DRAFT_WORKSPACE for this session:
export DRAFT_WORKSPACE="$HOME/.draft/workspaces/<chosen-slug>"
Context check
Check whether context already exists in the current workspace:
python3 -c "
import os
from pathlib import Path
ws = os.environ.get('DRAFT_WORKSPACE', os.path.expanduser('~/.draft/workspaces/default'))
files = list(Path(ws, 'context').glob('*/index.md'))
has_content = any(
'---' in Path(f).read_text() and 'No information recorded yet' not in Path(f).read_text()
for f in files
)
print('has_content' if has_content else 'empty')
" 2>/dev/null || echo "empty"
- If
empty: deliver the Welcome orientation below, then run the full interview from Q1
- If
has_content: run the re-run flow instead (see end of this file)
Welcome (first-time only)
Before asking Q1, deliver this orientation. Keep it warm but brief — don't pad it.
"Welcome to Draft — the collaborative context layer.
A few things to know before we start:
Your workspace lives at ~/.draft/workspaces/<profile-name>/. Everything Draft learns about your product, team, and priorities is stored there as plain markdown files — you can read, edit, or version-control them like any other file. You can have multiple named workspaces for different projects or clients — run /draft:profiles to manage them.
How updates happen: Draft writes to your workspace after this setup interview, and keeps it fresh as you work. Whenever you share a decision, a shift in priorities, or a team change, Draft will update the relevant files automatically. You can also say update [company / product / team / priorities] at any point to trigger an explicit refresh.
Change history: Each context dimension has a log/ subdirectory. Draft writes a timestamped entry whenever something meaningful changes — so you always have a record of how your thinking evolved.
Here's the layout:
~/.draft/
├── workspaces/
│ └── <profile-name>/ ← your active workspace
│ ├── context/
│ │ ├── company/ ← what you're building, business model, stage
│ │ ├── product/ ← product state, target user, key bets, roadmap
│ │ ├── priorities/ ← active sprint, top priorities, blockers
│ │ ├── team/ ← structure, who does what, capacity
│ │ └── decisions/ ← key decisions with status (active/superseded/parked)
│ ├── config/ ← collaboration config (created when you set up team sharing)
│ └── docs/ ← written artifacts (analyses, PRDs, strategies)
└── personal/ ← your global layer (shared across all profiles, never shared with team)
├── user/ ← your role and working style
├── memory.md ← vocabulary, preferences, recurring patterns
└── wip/ ← drafts not ready to share
How Draft works: Draft uses three specialized agents:
- @draft-researcher — finds and retrieves information (workspace files, integrations, web)
- @draft-executor — takes concrete actions (writes PRDs, creates Linear issues, sends Slack messages)
- @draft-learner — updates persistent memory (keeps your workspace files accurate over time)
Most requests follow the same pattern: researcher gathers context → executor acts → learner saves what changed.
Let's load your shared context workspace. This takes about 3–5 minutes."
Then ask Q0.5 before moving to the interview.
Q0.5 — Existing source (ask before Q1)
Use the AskUserQuestion tool to ask:
"Do you have existing notes, a docs folder, or a GitHub repo with product context? If so, I can import from there to bootstrap setup — saves a lot of typing.
(Paste a local path like ~/notes or a repo like owner/repo, or press enter to skip.)"
If they provide a source: delegate to @draft-researcher before continuing.
Give @draft-researcher this prompt:
"Read all markdown files under <source> and return a structured summary. For each file, identify which Draft dimension it most likely belongs to (product, company, team, priorities, decisions, research — or 'unmapped'). Return a JSON-like summary with one block per dimension:
DIMENSION: product
CONFIDENCE: high | medium | low
SUMMARY: <2-5 sentences synthesizing key facts — product name, what it does, target user, key bets>
DIMENSION: team
CONFIDENCE: high | medium | low
SUMMARY: <team size, structure, who does what>
...
For GitHub repos (owner/repo format): clone first using gh repo clone <repo> $(mktemp -d /tmp/draft-import-XXXXXX) -- --depth 1 --quiet, read the files, then delete the temp dir with rm -rf.
Do NOT return raw file contents — summaries only. Mark confidence high if the file clearly addresses that dimension, medium if inferred, low if speculative."
Store @draft-researcher's output as import_context — you will use it throughout the interview.
If they skip (press enter or say no): set import_context to empty and proceed to Q1 normally.
The interview
Ask questions one at a time. Wait for the full answer before asking the next. Adapt each question to what you've already heard — skip or merge questions if the user already answered them.
If import_context exists: before asking each question, check whether the import already provides a confident answer for that dimension. If confidence is high, skip the question and briefly acknowledge what you already know: "Got your product context from the import — just a couple of gaps to fill." If confidence is medium, ask a shorter targeted follow-up rather than the full question. Only ask the full question when confidence is low or the dimension is absent from the import.
If the user says "skip" at any point, stop the interview immediately and say: "No problem — run /draft:setup anytime you're ready to load your shared context workspace."
Q1 — What you're building
Use the AskUserQuestion tool to ask:
"What are you building, and who's it for? Give me the one-liner."
Listen for: product name, problem being solved, target customer, stage (pre-launch / early users / scaling / growth).
If the answer is vague on the customer, probe once using the AskUserQuestion tool:
"Who specifically — a job title, a type of company, a situation they're in?"
Q2 — Distribution and traction
Use the AskUserQuestion tool to ask, adapted to stage:
- Pre-launch: "Who are you building with right now — any design partners or early testers?"
- Post-launch: "How do users find you today, and how many do you have?"
Listen for: acquisition channel, number of users or customers, whether they're paying, how the user thinks about growth right now.
Skip this question if they already answered it in Q1.
Q3 — Team
Use the AskUserQuestion tool to ask:
"What does your team look like?"
Listen for: team size, who does what (eng, design, PM, founders), whether they're solo. One follow-up max if the answer is thin. Keep this short.
Q4 — Current focus
Use the AskUserQuestion tool to ask:
"What's the most important thing you're working on right now?"
Listen for: specific milestone or sprint goal, the bet they're making, blockers, what success looks like this week or month.
If vague, probe once using the AskUserQuestion tool:
"What would make this week a win?"
Q5 — How you work
Use the AskUserQuestion tool to ask:
"What tools do you use day-to-day, and what does a typical week look like?"
Listen for: project management (Linear, Jira, Notion), communication (Slack, email), planning cadence, how decisions get made.
Q5.5 — Team collaboration
Use the AskUserQuestion tool to ask:
"Do you want to use Draft together with your team? This lets you share your context layer and keep it fresh across multiple people."
If yes: Say "Let's set that up." → invoke /draft:setup-collab inline (call it as a skill before proceeding to Q6).
If no: Continue to Q6 normally. No config files created.
Q6 — Hardest problem
Use the AskUserQuestion tool to ask:
"What's the hardest problem you're dealing with right now?"
This is the most important question. Let them answer fully — don't rush or redirect. Listen for the actual underlying tension.
After the interview
Once all answers are in (or the user says they're done), do the following in order:
1. Synthesize
Do not pass raw answers to @draft-learner. Before writing, derive a merged synthesis from both sources:
If import_context exists: treat researcher summaries as the baseline. Interview answers take precedence where they conflict or add detail — the user's own words are more authoritative than inferred file content.
Derive the following for each dimension, combining import_context + interview answers:
- company: name, what they do, business model (B2B/B2C/marketplace/etc.), stage, funding if mentioned
- product: product name, problem it solves, target user (specific, not vague), how users find it today, key bets or hypotheses
- user: their role, team context, working style, communication preferences, tools they use
- team: size, structure, who does what, capacity constraints if mentioned
- priorities: current milestone or sprint goal, the single most important thing, open questions or blockers
- memory/vocabulary: domain-specific terms they used, working preferences, goals they named
If import_context provided a high-confidence summary for a dimension and no interview answer contradicts it, use the summary directly — no need to re-derive from scratch.
2. Confirm dimensions
Before writing files, tell the user what will be created and give them a chance to customize:
"I'll create context areas for: company, product, team, priorities.
Want to rename any or add custom ones? (e.g. research, decisions, legal)
Press enter to continue with defaults."
Wait for their response. If they want changes, adjust the dimension list. If they press enter or say nothing, use the defaults.
Pass the confirmed dimension list to @draft-learner in the next step.
3. Call @draft-learner
Pass the synthesized content as a single, structured message. Instruct @draft-learner to write:
Context files (under $DRAFT_WORKSPACE/context/) — write only the dimensions confirmed in step 2:
$DRAFT_WORKSPACE/context/company/index.md (if confirmed)$DRAFT_WORKSPACE/context/product/index.md (if confirmed)$DRAFT_WORKSPACE/context/team/index.md (if confirmed)$DRAFT_WORKSPACE/context/priorities/index.md (if confirmed)- Any additional custom dimensions the user requested (e.g.
$DRAFT_WORKSPACE/context/research/index.md)
Personal files (under ~/.draft/personal/ — global layer, NOT inside $DRAFT_WORKSPACE):
~/.draft/personal/user/index.md~/.draft/personal/memory.md
Each index file must have complete frontmatter:
name: dimension namedescription: 2–10 sentences, specific and factuallast_updated: today's date (YYYY-MM-DD)source: /setup interview
For priorities/index.md, also include the full body content.
Do NOT write log entries during /setup — this is the initial state.
4. Confirm
After @draft-learner writes the files, confirm to the user:
- Which files were written (paths)
- One-line summary of what's in each
- Invite corrections: "Does that capture it right? Say 'update [company/product/team/priorities/user]' to fix anything."
5. Ask the first real question
End with one sharp, specific question grounded in what you just learned. Base it on the hardest problem they named in Q6, or the most interesting tension you noticed.
Re-run flow
If context files already have content, show a one-line description per dimension (read from frontmatter), then:
Check for uninitialized dimensions. Scan $DRAFT_WORKSPACE/context/ for subdirectories that exist but have no index.md. If any are found, mention them:
"I also see these folders with no context index: research/, decisions/ (etc.)
Want me to scaffold them so Draft can load them? (y/n)"
If yes: for each confirmed folder, call @draft-learner to create context/<name>/index.md with the standard frontmatter template and create context/<name>/log/.
Then use the AskUserQuestion tool to ask:
"Your shared context workspace already has context. Want to do a full refresh, or just tell me what's changed?"
- Full refresh: run the interview from Q1, overwrite all context files
- What's changed: use the AskUserQuestion tool to ask "Walk me through what's different since we last synced." Then call @draft-learner to update only the relevant dimensions and write log entries for any changes to
product/ or priorities/.
Rules
- One question at a time — never ask two questions in the same message
- No preamble before questions ("Great answer!" / "Thanks for sharing" = never)
- Adapt — skip questions already answered, merge where natural
- Synthesize before writing — derived context beats transcribed answers
- The description field in frontmatter is loaded every session — write it like a briefing note, not a data dump
