Build Cycle

Build Cycle

The ritual you run again and again — all the way to PMF and beyond.

What This Is

Not a tool you pick up once. A rhythm.

Engineers document every solved problem so the next occurrence takes minutes, not hours. That's /compound. The Build Cycle is the founder equivalent: after every meaningful ship, you run this ritual. It asks the honest questions, synthesizes what you learned, and produces one record that compounds over time.

After 10 cycles, you have a complete record of your product's evolution — what you tried, what you learned, how you navigated each failure point, and exactly how you got to where you are.

Cadence: Run after each meaningful ship. Or weekly if you ship continuously. Never less than once a week. Never more than once a day — daily cycles are too granular to surface real signal.

Quick Start

Human-First Response

Default: answer in chat first with Verdict, Do this now, and Details saved. Keep chat under 150 words unless asked; do not paste the full artifact. Write/update the requested .md file as the durable record for agents.

Say: "Run build cycle" or "Let's do our cycle" or "What did we learn?"

This takes 20–30 minutes. Do it with intention — not rushed. Total time invested compounds.

What gets produced:

• cycles/YYYY-MM-DD.md — the permanent cycle record
• Updated founder-context.md — your source of truth

How It Opens

Before asking you anything, the skill reads two things:

1. founder-context.md — your product's current state, north star, stage
2. The most recent cycles/ document — what happened last cycle, what you decided

If neither exists, the skill runs a one-time First Cycle Setup (see below).

After reading both, the skill opens with a single orienting statement — not a question:

> "Last cycle you decided to [X]. Before we reflect on how that went, I want to name one thing I noticed from your context: [observation]. Let's start there."

This is borrowed from the best therapy and coaching: name what you see before asking what they feel.

---

The Six Phases

Run in order. Don't skip phases even when you're in a hurry — the order matters.

---

Phase 1 — The Honest Open

One question: "What actually happened since last cycle — not what you hoped would happen, but what actually happened?"

This is not a status update. The agent is listening for:

• The gap between what was planned and what shipped
• What got avoided or deprioritized
• Tone: are you energized, exhausted, confused, or numb?

After you answer, the agent reflects back what it heard — including the things you said but might have glossed over. This is the "document-review" move: surface what's being minimized.

> The uncomfortable question for this phase: "What are you not saying?"

---

Phase 2 — What Users Actually Did

One question: "What did your users do — not what you hoped they'd do, not what they said they'd do, but what they actually did?"

This phase separates signal from noise by distinguishing:

| Signal Type | Weight | |------------|--------| | Behavior (what they did without being asked) | High | | Unprompted feedback (what they said without being asked) | High | | Prompted feedback (what they said when you asked) | Medium | | What you want to believe | Zero |

The agent asks follow-up questions here — one at a time, never more than three:

• "Did any users do something you didn't expect?"
• "Did any users NOT do something you expected?"
• "Has anyone told someone else about this without you asking them to?"

> The uncomfortable question: "Is there user behavior you've been avoiding looking at?"

---

Phase 3 — The Honest Assessment (Parallel)

Phase 3 runs three independent analyses simultaneously — MPP scoring, PMF signal assessment, and cycle pattern reading. These have no dependencies on each other. Run them in parallel.

Spawn these 3 agents simultaneously:

Agent A — MPP Scorer Reads: founder-context.md, last cycle record, any user behavior data from Phases 1–2 Task: Score MPP using the 5 criteria from [mpp-criteria.md](../mpp-evaluator/mpp-criteria.md) Returns: Composite score (1–10) + the single criterion holding the score back + one sentence on why

Agent B — PMF Signal Analyst Reads: founder-context.md, all cycles/ records Task: Assess PMF signal strength Signal levels:

• None: Users use it once and don't return. No unprompted sharing.
• Faint: Some users return. 1–2 unprompted mentions across all cycles.
• Building: A core group retains consistently. Word of mouth starting without incentive.
• Clear: Users would be upset if it disappeared. Usage grows without effort.

Returns: Signal level + the specific evidence for that rating + the one behavior that would move it to the next level

Agent C — Pattern Reader Reads: All cycles/ records (last 3 minimum, all if available) Task: Identify patterns across cycles Patterns to find:

• Momentum: Is MPP score trending up, flat, or declining?
• Effort: What type of work keeps repeating without producing signal?
• Avoidance: What has been mentioned across cycles but never acted on?
• User: Is there a user type showing stronger signal than others?

Returns: Momentum trend + the most important pattern (effort or avoidance, whichever is stronger)

Wait for all 3 agents. The orchestrator presents findings.

The orchestrator does NOT soften the assessment. If the score is 3/10, say 3/10. Present all three agent findings, then ask:

> "What's the most honest reason we're at this score?"

This is not rhetorical. It's the most important question in the cycle. Take time here.

> The uncomfortable question: "If this score doesn't improve next cycle, what does that mean?"

> If Pattern Reader returns 3+ consecutive flat or declining cycles: Do not continue to Phase 5. Name it: "I need to flag something. We've had [N] cycles without improving signal. That's not a reason to panic, but it is a reason to diagnose before deciding. I want to run the Failure Navigator before we choose what to do next." See [failure-modes.md](failure-modes.md).

Sequential Fallback (Codex / OpenCode)

If the agent cannot run parallel subagents, run Phase 3 in this order:

1. MPP Scorer
2. PMF Signal Analyst
3. Pattern Reader
4. Orchestrator synthesis

Same output. ~2–3× longer.

---

Phase 4 — Synthesis

With all three agent returns from Phase 3 in hand, the orchestrator (no subagents — this is synthesis work) connects the findings:

• What does the MPP score tell us about the Phase 1 and 2 findings?
• What does the PMF signal tell us about which parts of the product are working?
• What does the pattern read reveal that neither score alone would show?

One paragraph. No lists. This is the co-founder thinking out loud — connecting the dots between what was built, what users did, and what the numbers say.

---

Phase 5 — One Decision

Not a list. One thing.

The agent synthesizes everything from Phases 1–4 and proposes a single decision for the next cycle. The format is always:

> "The most important thing you can do in the next [cycle period] is [specific action], because [reason grounded in the cycle's findings]. You'll know it worked if [specific observable signal]."

Then the agent asks: "Does this feel right — or is there something the data is pointing to that I'm missing?"

This is the co-founder moment. Not just presenting the analysis — checking it against your instinct, which sometimes sees things data doesn't.

The one decision becomes the commitment field in the cycle record.

---

Phase 6 — The Record

The agent writes two things:

1. cycles/YYYY-MM-DD.md — the permanent record

Following the template in [cycle-template.md](cycle-template.md). This document is never edited after the cycle closes — it's a permanent record, like a ship's log. Future cycles can reference it, but not change it.

2. Updated founder-context.md

The source of truth is updated:

• Stage (has it changed?)
• North star metric and current value
• MPP score
• PMF signal strength
• Current focus (the one decision from Phase 5)
• Last cycle date and key learning

---

First Cycle Setup

If no founder-context.md or cycles/ exists, the agent runs a one-time setup:

It asks 6 questions — one at a time:

1. "What are you building, in one sentence?"
2. "Who specifically is your first customer — describe their situation, not their demographics."
3. "What stage are you at? (Pre-launch / Just launched / Have users / Revenue)"
4. "What's the one metric you're tracking right now?"
5. "What's the biggest thing you're unsure about?"
6. "What did you try most recently, and what happened?"

This creates founder-context.md and cycles/[today].md (the first cycle). Setup takes 10 minutes.

---

The Cycle Record

Each cycle produces one document. The document structure is in [cycle-template.md](cycle-template.md).

The cycles/ directory becomes your product's memory:

cycles/
├── 2026-02-01.md    # First ship — the baseline
├── 2026-02-08.md    # First users
├── 2026-02-15.md    # First feedback
├── 2026-02-22.md    # First sign something is working
└── ...              # The full story, in your own words

After 10 cycles, you'll have something most founders don't: a clear record of why you made every significant decision, what you learned from it, and how the product evolved. This is not just useful for you — it's the clearest possible input for a fundraising narrative, a co-founder conversation, or a moment of doubt where you need to remember why you started.

---

What the Build Cycle Is Not

• Not a retrospective format borrowed from engineering sprints. Those are team-coordination tools. This is a founder-learning tool.
• Not cheerleading. If your signal is weak, the cycle says so. Honest over kind.
• Not a feature planning session. Features happen after the cycle. The cycle decides what problem to solve — not how to solve it.
• Not skippable when things are going well. The best cycles happen when things feel good. They find the fragility before it becomes a crisis.

---

Related Skills

• Use mpp-evaluator for a deep-dive MPP scoring session between cycles
• Use failure-navigator when build-cycle detects stagnation (3+ flat cycles)
• Use co-founder (Partner phase) for the broader PMF journey and beyond
• Use pmf-signal-reader (PMF phase) when build-cycle shows "Building" PMF signal and you want to understand it deeply

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gvkhosla
• Source: gvkhosla/founder-skills
• License: MIT
• Homepage: https://fskills.xyz

Install and usage instructions live in the source repository linked above.