| name | workspace-kickoff |
| description | Orients an agent in a new or resumed workspace by reading the canonical orientation files (CLAUDE.md, AGENTS.md, .cursor/scratchpad.md, and PLAN.md if present), summarizing what the workspace is for, current status, and the top blockers and next actions — then waiting for direction instead of executing. Use when the user says "start this workspace", "resume this workspace", "kick it off", "begin session", "pick up where we left off", "what is this workspace", or opens a scaffolded workspace for the first time in a session. |
Workspace Kickoff
Purpose
Make the first turn of a session load-bearing. Read the orientation files,
produce a short situation report, and stop. Do not execute work until the
user confirms direction. This pattern relies on workspaces that follow the
scaffolding convention (CLAUDE.md, AGENTS.md, .cursor/scratchpad.md,
optional PLAN.md).
The three-layer model
Scaffolded workspaces separate orientation across three files with
different update cadences. Respect the separation when summarizing —
don't blend them.
| Layer | File / System | Grain | Update cadence | What you pull from it |
|---|
| Governing initiative | PLAN.md | Goal, milestones, success criteria | Rarely — at milestone boundaries | "What is this workspace for; how far along are we at the milestone level" |
| Tracked work | beads (bd ready, bd show <id>) | Individual tasks with dependencies | Per task | "What's actually claimable right now" |
| Working memory | .cursor/scratchpad.md | In-flight item, blockers, decision log | Every session | "What's in progress this moment; what's gating us" |
PLAN.md may also be read by workspace tooling (e.g. an MCP plan
reader); treat it as structured metadata, not a to-do list. Beads are
the granular tasks that roll up to milestones. The scratchpad is a
session-level notebook — NEVER read it and think it's authoritative
for "what's next." Use bd ready for that.
Steps
-
Read the orientation files, in this priority order. Some may be
missing; that's fine, read what exists.
| Priority | File / System | What it tells you |
|---|
| 1 | CLAUDE.md or AGENTS.md | Working conventions, constraints, rules of the road |
| 2 | PLAN.md | Goal, milestones, success criteria (coarse) |
| 3 | bd ready (if .beads/ exists) | Priority-sorted list of claimable tasks (fine-grained) |
| 4 | .cursor/scratchpad.md | What's in flight, current blockers, decision log |
| 5 | README.md | Human-facing summary (fallback only — the above are authoritative) |
| 6 | findings/ or equivalent | Latest artifacts from prior sessions |
-
Produce a summary with this exact structure:
## Workspace: <name>
**What it is:** <one sentence from PLAN.md goal or README>
**Working mode / constraints:** <key rules from CLAUDE.md — e.g. learning-only, no PRs, beads-for-tracking>
## Where we are (per PLAN.md)
<1–3 bullets: what's claimed/in-progress right now>
Waiting for your direction. Shall I proceed with #1, pick a different
one, or do something else?
Fallback: workspace has no scaffolding
If none of CLAUDE.md, AGENTS.md, .cursor/scratchpad.md, or PLAN.md
exists, say so plainly and ask:
This workspace doesn't have kickoff scaffolding yet (no CLAUDE.md,
AGENTS.md, scratchpad, or PLAN.md). Want me to bootstrap them based
on what's here, or would you rather tell me in chat what this
workspace is?
Do not invent a summary from filesystem spelunking alone.
Session etiquette the agent should honor
- Start in planner mode. The summary is the whole first turn. Don't
drift into execution.
- Update the scratchpad before ending. Either the agent or the user
rewrites
Executor — next actions at session end so state carries
forward. If the user says "done" or "stop" without updating, prompt
them: "Want me to update .cursor/scratchpad.md with today's progress
before we stop?"
- Respect workspace-local rules over defaults. If CLAUDE.md says
learning-only, don't open PRs. If it says beads-for-tracking, don't
use TodoWrite.
Examples of correct kickoff output
Example: a scaffolded investigation workspace
User prompt:
Start this workspace.
Agent response (after reading PLAN.md + scratchpad + bd ready):
## Workspace: ci-audit
**What it is:** Reduce wall-clock of the CI pipelines that gate customer releases.
**Working mode / constraints:** Learning-only — no PRs, no branches pushed to origin. Findings stay local.
## Where we are (per PLAN.md)
M1 (orientation) DONE. M2 (diagnose) IN PROGRESS. M3/M4 not started.
## What's in flight (per scratchpad)
- Nothing currently claimed. All M2 beads are ready.
## Blockers
- `ci-audit-r6d` — waiting on a conversation with the platform team about the LLM container deprecation. Gates the whole M2→M3 transition.
## Top 3 next actions (per `bd ready`)
1. `ci-audit-r6d` Message the platform team re: LLM container deprecation
2. `ci-audit-xpc` Inventory the main service's CI pipeline config end-to-end
3. `ci-audit-7ss` Attach real wall-clock timings from the CI provider's API
## Ready
Waiting for your direction. Shall I proceed with #1, pick a different one, or do something else?
Example: a platform/tooling workspace
User prompt:
Pick up where we left off.
Agent response:
## Workspace: codebase-atlas
**What it is:** Progressive knowledge store and agent-ready CLI over the org's code-host workspace.
**Working mode / constraints:** Bash + jq + curl only. Scrapers idempotent. Views curated. `data/` gitignored.
## Where we are (per PLAN.md)
M1 (bootstrap) DONE. M2 (ship to the org remote) BLOCKED on token scopes. M3/M4 not started.
## What's in flight (per scratchpad)
- Nothing in-progress. First commit is local (0c50048), push gated on `atlas-le6`.
## Blockers
- `atlas-le6` — code-host API token lacks write scopes. Blocks `atlas-1dh` (create repo + push). Needs the token regenerated with repo-write scope and a keychain update.
## Top 3 next actions (per `bd ready`)
1. `atlas-le6` Regenerate the API token with write + pipeline scopes
2. `atlas-wf8` Author `skills/atlas-cli/SKILL.md` (ships with repo)
3. `atlas-v4g` `scrapers/sweep-pipeline-runs.sh` — unlocks real CI-audit timings
## Ready
Waiting for your direction.
Anti-patterns
- Don't execute on the first turn. Even "obvious" next steps need
human confirmation. The user might know something the scratchpad
doesn't.
- Don't rewrite the scratchpad, PLAN.md, or beads during summary.
Read them, report them. Updates happen at end-of-session or on
explicit instruction.
- Don't substitute filesystem exploration for reading orientation
files. If CLAUDE.md or PLAN.md exists, those are authoritative.
- Don't invent a numbered to-do list from PLAN.md. PLAN.md tracks
milestones (M1, M2…), not next actions. If
.beads/ exists, use
bd ready for the worklist.
- Don't list more than 3–4 next actions. Longer lists dilute focus.
- Don't cite files verbatim. Summarize. The user wrote them; they
know what they say. Your value is compression and reprioritization.