Skip to content

Workflow Walkthrough EN

Sergey Emelyanov edited this page Nov 6, 2026 · 1 revision

Workflow-Walkthrough-EN — step 5 (options) from the inside, on the CVStart demo

A teardown of one workflow step, based on the CVStart demo run (projects/демо/). All quotes come from real repository files: .claude/skills/options/SKILL.md, .claude/agents/red-team.md, config/quality_gate.json, projects/демо/outputs/03_options.md. Russian version: Workflow-Walkthrough. The overall run — Demo-CVStart-EN.

Why step 5

The options step ("Options and objections") is the central one in the process: this is where a vague request turns into a meaningful fork, and the beauty of a recommendation gets tested by an attack. The process rule: at least three meaningfully different options, including "keep as is" and "shut down/wind down", with strong objections, economics and the conditions under which each option holds — and the human makes the choice, not the agent.

What the agent receives as input

By the start of step 5, projects/демо/ already holds the artifacts of steps 0–4:

Input What it gives this step
00_brief.md the decision, the comparison axis ("growth of paying users and ARR"), the constraints (MIT core, solo, local-mode trust) — a filter applied before generation
sources/S001–S011.md, sources.jsonl verified materials: the product, the market, competitors, synthetic data
01_research_plan.md 6 solution-changing questions; which are answered, which are still open
claims.jsonl (C001–C018 by this point) claims with types and grades — the only permitted building material for option economics
02_situation.md known/gaps/contradictions; in the demo — the contradiction "100% local" [C001] vs the cloud

The agent does not "remember" this from chat history — it reads the files. That is why the step is reproducible and resumable after a break.

The step's timeline in one chain

pfs step options
  → read 02_situation.md + the registries (files, not memory)
  → pfs kb search: the cards M-goals-drivers-bets, M-solution-design, M-business-models,
                    M-unit-economics, M-risks, M-strategy-growth
  → pfs questions options (criteria · boundaries · objection quorum)
  → generate 5–7 candidates → keep 5 meaningfully different (A–E)
  → per-option economics: formulas on [C016][C020], labeled forecast
  → the red-team subagent ← 02 + 03 → 20 hits (15 objections + 5 assumptions)
  → integrating the hits: option text + B's mitigations + hidden assumptions
  → the comparison table along the brief's axis (ARR)
  → the recommendation + revision conditions
  → pfs check (the 03_options.md gates) → 0 errors
  → pfs ask "Which option goes into the strategy?" → STOP: waiting for the human

Then — the leader's answer ("B with mitigations") → pfs answer → pfs step choice done. In the demo the choice was recorded as claim [C019] and the leader's answer [S011] — the grounds of the future D001.

What the process prescribes (the options skill)

The step's instruction (.claude/skills/options/SKILL.md) fixes a strict order:

  1. Generate 5–7 candidates, keep 3–5 meaningfully different (a different bet: segment, value, monetization model, channel, investment scale), not variations of one. Mandatory: Option A: keep as is (a baseline forecast of "doing nothing", not "zero") and an exit option (shut down / wind down / hand off).
  2. Per option — a fixed structure: the essence; for whom and which job it solves; what must be true (assumptions [C###]); the expected effect with a range and a type (forecast); costs and revenue impact; risks; Objections (2–4, "as a smart skeptic would say"); what would refute the option.
  3. Hand the document to the red-team subagent; fold its objections into the options; "if an objection kills an option — say so outright".
  4. A comparison table along 4–6 criteria derived from the brief's goal (no invented weights without a basis).
  5. A recommendation + the conditions under which the recommendation changes. pfs check.
  6. The choice: pfs ask "Which option goes into the strategy?" --options "A,B,C,…", stop, and after the human's answer — pfs answer, pfs step choice done.

Which method cards the agent reads

Before working, the agent pulls in the methodology — via the step's card list and knowledge-base search:

pfs step options          # the step heading, gates, question block
pfs kb search "options and bets"    # as needed — sections of cards, not the whole base

Cards prescribed by the skill: M-goals-drivers-bets (options as bets), M-solution-design (solution hypotheses and risks), M-business-models (monetization models), M-unit-economics (option economics), M-risks (risks and mitigations), M-strategy-growth (growth types). Related cards (pfs kb search): cannibalization (flows between options), traction-model (growth models), model-validation (checking the numbers). The economy rule: do not read the whole base — 1–3 cards found by search.

The step's question block: criteria before the choice

pfs questions options gives the leader 3 questions. The point of the block is to fix the comparison frame before a recommendation exists, so the recommendation cannot be fitted to it:

Block question Why The tool's recommended answer
options-criteria: which criteria to compare on? criteria are fixed BEFORE the choice "contribution to the metric"
options-borders: what is out of bounds? conscious rejections are a mandatory part of strategy "reduced privacy" (in the demo this meant: cloud is opt-in only)
options-quorum: how many red-team objections must options survive? attack depth drives committee trust "2–3 per option"

In the demo the comparison axis was already given by the leader at intake ("growth of paying users and ARR") — so the comparison table in 03_options.md is built around it, without invented weights.

Generation: 5–7 candidates → 5 meaningfully different

The agent sketches candidates along the bet axes (who we sell to · what we sell · how we monetize · where we invest) and collapses duplicates. The demo kept 5, and they differ by the bet itself, not by wording:

  • A. Keep as is — a free local OSS with no commercial version (the baseline forecast).
  • B. Hosted freemium SaaS — cloud, global; the bet: a subscription for hosted convenience + multilingual coverage.
  • C. B2B white-label — licenses to career centers and universities; the bet: selling to organizations.
  • D. Open-core dual-license — enterprise self-host + ATS integrations; the bet: enterprise licenses.
  • E. Shut down the commercial branch — the project stays a non-commercial OSS (the exit option).

Each option's economics — formulas on assumptions [C016][C020] with an explicit type label (forecast/model forks); market numbers — tagged [C005][C006][C007][C011].

Calling the red team: what the subagent receives and returns

The subagent is defined in .claude/agents/red-team.md: "a smart skeptic from the board of directors who dislikes this project". Input: 02_situation.md and 03_options.md (+ pfs claim list when needed). It must return, per option:

  1. 2–3 of the strongest objections — concretely: which assumption is false, which data is missing, how competitors/regulators/users would react;
  2. one observation that would refute the option fastest and cheapest;
  3. which meaningfully different option is missing from the set;
  4. separately on the recommendation: where a hypothesis is passed off as a fact, where numbers lack sources, where "keep as is" is painted too dark. "Do not soften. Do not propose your own strategy."

In the demo the red team returned 20 hits: 15 objections + 5 hidden assumptions. Examples of hits against the recommended B: "PII/GDPR and scraping liability shift to the developer" (jurisdiction and litigation risk), "'$0 forever' is a public promise; HN/PH will read a subscription launch as bait-and-switch", "freemium conversion 2–5% [C020]; the product is episodic — churn is naturally high". Against A: "MIT = someone else's monetization — a fork will wrap the code into a closed SaaS (the WordPress/WP Engine precedent)".

Integrating the objections: how hits land in the document

Objections are not "considered: yes" — they are built in through three mechanisms:

  1. In the option's text — a "Red team objections" block per option, worded as a skeptic would.
  2. As mitigations inside the option — for B: a BYOK free tier (zero inference cost on free — removes the "money is against us" hit), an opt-in cloud on the local core (removes the "trust collapse" hit), honest pricing [C014], stop thresholds from the future 06.
  3. As a hidden assumption and a validity condition — each option names the assumption it stands on and the condition under which it should be dropped from the agenda.

Some hits become counter-metrics of the validation plan (step 8): "trust collapse" → a GitHub-stars drop > 30% as a stop signal. Full teardown — Validation-Plans-EN.

What pfs check verifies at this step

The gates for 03_options.md from config/quality_gate.json:

Check Rule In the demo
min_options: 3 at least three options 5 (A–E)
status_quo_keywords "keep as is" present Option A
exit_keywords an exit option present Option E ("shut down the commercial branch")
per_option_sections: ["Objections"] every option has objections 15 red-team objections
numbers_need_tags: true numbers tagged [C###]/[S###] all registry numbers
option_heading_regex machine-readable headings (## Option X) A–E

A failed gate = the step does not close: pfs step options done will not pass until the document complies.

If a gate failed: what a verification failure looks like

For contrast — what the agent sees when a rule is broken. Hypothetical examples (in the demo the gates passed on the first presentation of the final version):

Violation pfs check reaction
3 options in the document, all variations of "add features" error: no status_quo_keywords ("keep as is") and no exit_keywords (shut down/wind down)
Option C has no "Objections" block error: per_option_sections unmet for one option
A number "$2B" without a tag numbers_need_tags warning/error: a number without [C###]/[S###]
Economics built on a previously retracted claim error: a reference to an inactive claim
Trying to record "conversion 4% (per our data)" as a fact without a source pfs claim add refuses at the registry stage — it never reaches the document

The step cannot be "pushed through" with open errors — that is the point of the gates: the next step starts only from a valid document.

The step's output: 03_options.md — a real fragment

The comparison table from the demo (quoted verbatim from projects/демо/outputs/03_options.md):

## Comparison along the "paying subscribers/ARR" axis (the leader's goal)

| Option | ARR in 12 months (model) | Risk | Fit with constraints (MIT core, solo, trust) | Red team |
|---|---|---|---|---|
| A | $0 | low | ✓ | "an expensive hobby" |
| B | $29/mo × 100–300 = $35–105K ARR [C016][S008] | medium | ✓ with BYOK-free and opt-in cloud | the best growth profile |
| C | $24–60K [C020] | high (sales) | ✓ | no sales competence |
| D | $10–50K [C020] | high (relicensing) | ✗ (changing MIT) | risky and slow |
| E | $0 | — | ✓ | abandoning the asset |

Recommendation: **B — hosted freemium SaaS, global**, with a BYOK free tier (removes the inference
risk), language waves from the 19 ready locales, honest pricing without auto-renewals as positioning
[C014], and stop thresholds from 06. The choice belongs to the leader (step 6).

Note: the recommendation already carries conditions (BYOK-free, opt-in, honest pricing, thresholds) — that is, the conditions under which it changes are stated together with it, as the skill requires.

Hand-off to step 6: the pause and the human's choice

The document is ready — but the choice is not made automatically:

pfs ask "Which option goes into the strategy?" --options "A,B,C,D,E"
# …the agent stops…
pfs answer "B — hosted freemium SaaS with mitigations"
pfs step choice done --next "strategies: product and communications"

In the demo the leader chose B with mitigations — the answer was registered [C019][S011] and became a ground of decision D001. Steps 7–11 follow; their outputs — 04_product_strategy.md, 05_comms_strategy.md, 06_validation_plan.md, decisions.jsonl — are covered in Demo-CVStart-EN.

What to take away from this walkthrough

  • The frame is set by the process; the model fills it. The number of options, the structure, the objections, the gates — mandatory; the content — the LLM's work over verified claims.
  • Verifiability is not "on one's word of honor". Every number is a tag; every recommendation comes with revision conditions; every attack carries a source and a number for a counter-metric.
  • The human decides at the highest-cost point. The agent prepares the fork and the defense of every path, but the choice — step 6 — always belongs to the human.

See also: Demo-CVStart-EN · Skills-EN · Subagents-EN · Quality-Gates-EN · Validation-Plans-EN · Decisions-and-Monitoring-EN

Стратег Фабрики продуктов

Устройство

Компоненты

Справочник

Процесс

Справка и прозрачность

Clone this wiki locally