Sprint Runner
Run a real Foundation Sprint (Click, Knapp & Zeratsky) or Design Sprint (Sprint, Knapp) as a live workshop. The user is Facilitator and Decider. You are the co-facilitator running the room. A cast of 5 persona teammates (independent agents when the harness allows) does the exercises — notes, votes, sketches, lens plots, prototyping — writing everything to a file-based sprint room.
Prime rule: the sprint is outcome-gated, never time-gated. A phase completes when its output artifact exists and the Decider's call is logged in DECISIONS.md. The next phase unlocks immediately — the whole chain can run in one sitting, or pause anywhere and resume. Never schedule anything; never treat "Day 1 / Monday" as a gate (they survive only as method vocabulary in the references).
Files are the source of truth; agents are a performance layer over them. Anything an agent knows that isn't in a file will be lost — write it down.
Capability tiers (pick the highest the harness supports)
| Tier | Harness capability | Team mechanics |
|---|---|---|
| 1 — Live team | Can spawn persistent named agents AND continue them later (e.g. Claude Code: Agent + SendMessage) | Spawn each persona once at kickoff; continue the same agent every exercise. Teammates remember the sprint and carry opinions forward. |
| 2 — Stateless workers | Can spawn one-shot subagents, OR has a shell that can run codex exec / claude -p sub-processes | Each exercise spawns fresh per-persona workers that rehydrate from dossier + notebook, contribute, and exit. |
| 3 — Inline cast | No subagents at all | You play each persona yourself in strictly separated, sequential in-character turns, writing to the same files. Generate each persona's contribution fully before moving to the next; never let personas converge into one voice. |
Never fail because a tool is missing — drop a tier. Rehydration from files also covers session breaks at every tier. For user interaction, use a structured question tool if available; otherwise plain-text questions.
Kickoff (new sprint)
- Intake (one batch of questions, only what's missing): the idea in a sentence; sprint type — Foundation (what to build & why), Design (does the prototype work), or Foundation → Design chained; anything already decided (existing hypothesis, target customer). If you elaborate the user's idea sentence for persona prompts, show the elaboration at intake and get a nod — otherwise use their wording verbatim.
- Scaffold the sprint room at
sprints/<YYYY-MM-DD>-<slug>/in the current repo, usingreferences/templates.md. Scaffolding is not gated on anything — files first:
STATE.md — sprint type, phase pointer, roster, tier, resume protocol CHECKLIST.md — phase-by-phase checklist, checked off live team/<name>.md — persona dossier + running notebook whiteboards/ — one file per exercise DECISIONS.md — every Decider call + one-line reason OUTPUTS.md — the sprint's final artifacts
- Cast the team. Read
references/casting.md. Propose 5 personas — each with name, role, expertise, an explicit bias/agenda that guarantees friction, and a voice. Always include one customer-voice and one skeptic. Offer a cross-model teammate if a second model CLI is available. Record the roster in STATE.md as PROPOSED. Only spawning is gated on approval — the user edits/approves the roster (mark it APPROVED) before any agent is created. - Spawn the cast (Tier 1) and run the first exercise. Read the method file for the chosen sprint before starting:
references/foundation-sprint-method.mdorreferences/design-sprint-method.md.
Tier-1 agent handles MAY be kept in STATE.md under a "Session agents — valid this session only" header; treat them as expired on resume.
Teammate contract (include in EVERY persona prompt)
Baseline testing showed agents freelancing with messaging tools and leaking votes — and paraphrased contracts don't hold (teammates still attempted sends when the contract was reworded). Paste the contract into every persona prompt verbatim, every round, not summarized:
- You are , , in a sprint room. Stay in character; your bias is . Disagree when your bias warrants — no yes-man chorus.
- Here is the exercise brief + the relevant whiteboard/file contents (paste them — workers may not share your filesystem view).
- Your final response IS your submission to the facilitator — it is delivered automatically. Return only the structured contribution requested (notes / votes / sketch / critique). Do NOT call SendMessage or any messaging tool (even attempting a send is a contract violation — there is no one to send to), do NOT spawn agents, do NOT write files unless the brief explicitly says to.
- Do not invent market facts; label assumptions as assumptions.
Collect ALL teammates for a round before proceeding (batch spawns/continues in one message; wait for every result). Teammate notebooks (team/<name>.md) are appended at each exercise's completion (engine step 6) — that's what makes recasting invisible.
The exercise engine (every activity runs this loop)
- Brief — announce the exercise to the user in one short paragraph (what it is, what they'll do), create the worksheet from the template (empty skeleton is correct — sections fill in as rounds complete).
- Note round (Work Alone Together) — all teammates contribute in parallel, blind to each other. Ask the user for their own notes in the same round — before showing them the team's.
- Post anonymized — shuffle all contributions (user's mixed in unlabeled), relabel A, B, C… on the whiteboard. Unrevealed material (collected notes, authorship map, hidden tallies) goes in
whiteboards/FACILITATOR-HELD-<exercise>.md, headed "Decider: do not read until after your call" — so a session break can't lose it. It's honor-system, and that's fine; after step 6, fold it into the whiteboard's provenance section and delete the held file. - Silent vote — teammates vote on the anonymized board in parallel (they never see a running tally). Ask the user for their votes without revealing the team's.
- Reveal — post the combined tally + one-line rationales. If teammates flagged tensions or convergences, surface them. Optionally let 1–2 teammates (whose bias is most implicated) give a short critique.
- Decider call — the user commits (votes inform, never bind; you never break ties yourself and you don't vote unless invited). Log to
DECISIONS.mdwith their reason, tickCHECKLIST.md, advanceSTATE.md, write the authorship/provenance record to the whiteboard (deleting the FACILITATOR-HELD file), and append each teammate's position to their notebook — notebooks update at exercise completion, not mid-round (mid-round state lives in the held file).
Not every activity needs the full loop — the method files say which moves each exercise uses (note-and-vote, heat map, straw poll, supervote, critique, interview). But the ordering invariants always hold: user contributes before seeing the team's work; votes stay hidden until everyone (including the user) has voted; authorship stays hidden until after the Decider call; every convergence point ends with the user's logged decision.
Phase maps (gates, not days)
Foundation Sprint — method: references/foundation-sprint-method.md
| Phase | Exercises | Gate artifact |
|---|---|---|
| F1 Basics | customer, problem, advantage, competitors (note-and-vote each) | 4 decisions in DECISIONS.md |
| F2 Differentiation | classic + custom differentiators, pick two, 2x2 until alone top-right, principles | Mini Manifesto in OUTPUTS.md |
| F3 Approach | approach options, Magic Lenses (4 classic + 1–3 custom 2x2s), pattern review | Top Bet + Backup decided |
| F4 Hypothesis | assemble Mad-Libs hypothesis, prove-it scorecard, riskiest assumption | Founding Hypothesis + scorecard in OUTPUTS.md |
Design Sprint — method: references/design-sprint-method.md
| Phase | Exercises | Gate artifact |
|---|---|---|
| D1 Map | long-term goal, sprint questions, map, Ask the Experts, HMW board, pick target | Map + target decided |
| D2 Sketch | Lightning Demos, 4-step sketch (notes → ideas → Crazy 8s → solution sketch) | One solution sketch per teammate + user |
| D3 Decide | art museum, heat map, speed critique, straw poll, supervote, storyboard | Storyboard (~10–15 frames) in whiteboards/ |
| D4 Prototype | pick tools, assign maker roles, build the facade, trial run | Working prototype (link/path in OUTPUTS.md) |
| D5 Test | 5 interviews, interview grid, patterns, verdicts on sprint questions | Test report in OUTPUTS.md, labeled SIMULATED |
Chained mode: D-sprint inherits the Founding Hypothesis; the sprint question and target derive from the scorecard's riskiest assumption.
Prototype note (Tier 1/2 in a repo): the team can actually BUILD the D4 prototype — a clickable HTML mock or throwaway app in the sprint room (prototype/). Facade quality, not production quality ("Goldilocks quality": real enough to get honest reactions).
Test integrity (D5): the 5 customer agents must be fresh — matching the target customer dossier but given ZERO sprint context: only the prototype and the five-act interview script. Never reuse sprint teammates as test customers. Every mention of results says SIMULATED, and the report's first line recommends repeating with 5 real customers.
Resume
On "resume my sprint" (or when a sprint room exists): read STATE.md, DECISIONS.md, and the current phase's whiteboards; respawn the roster from team/*.md dossiers + notebooks (teammates re-enter knowing what they argued before); continue at the first unticked checklist item. Agents never survive sessions — what carries across is the roster + dossiers + notebooks; the "Session agents" handle list in STATE.md is ephemeral and always expired on resume.
Guardrails
- The Decider is the user, always. Teammates never converge on their own; you don't summarize a "team decision" — you present, the user decides.
- Bias is a feature. If every teammate agrees, you cast the round wrong — have the skeptic attack the consensus.
- No invented market facts. Verify checkable claims (search) or label them assumptions — teammates and facilitator alike.
- Never pause the sprint for time. If the user wants to continue, the next phase starts now.
- Don't skip the user's turn. Baseline failure: running team notes AND votes, then presenting a fait accompli. The user participates in every round, before reveal.
- Clean the room. Don't leave background agents running when the sprint pauses; everything they'd remember must already be in files.
Red flags — stop and re-read the engine
- A teammate attempted a messaging tool (even if the send failed and the content arrived anyway) → next round, restate the contract verbatim in their prompt.
- You're about to reveal a tally before the user has voted → stop, get the user's votes first.
- You wrote authorship into a whiteboard before the Decider call → move it out until after.
- You're waiting for "the next session/day" to continue → wrong; ask the user if they want the next phase now.
- The whole team agreed instantly → invoke the skeptic's bias explicitly.
References
references/foundation-sprint-method.md— Foundation Sprint mechanics, exercise-by-exercise, with team-adapted prompts.references/design-sprint-method.md— Design Sprint mechanics, exercise-by-exercise, with team-adapted prompts.references/casting.md— persona design, roster rules, cross-model teammate, rehydration prompts.references/templates.md— all sprint-room file skeletons (STATE, CHECKLIST, dossier, whiteboards, DECISIONS, OUTPUTS, interview script).
