Problem Observatory
Scout rotates through public problem signals, normalises evidence, creates problem dossiers, and emits up to ten unranked seeds. It is a top-of-funnel lane, not a ranking run.
Singapore time · idempotent · no external actionOffdata / operating system
Run the company
by intent.
A practical manual for turning your Hermes profiles into a coordinated, evidence-led company factory: discover opportunities, rank the portfolio, build carefully, and keep every important decision visible.
Control surface / 01LIVE MODEL
One front door.
Clear handoffs.
Chief of Staffintent, approval, exception
↓
Scout → Reviewerproblems, evidence, counterevidence
↓
Foundry → Builder / Operatorranked bets, quality gates, reconciliation
01 / Rhythm
Let the system run on a cadence you can trust.
Daily work creates evidence and problem-backed seeds. Monday and Thursday create portfolio movement. You can interrupt the cadence with a directed sprint whenever a signal deserves a closer look.
Scout intelligence + ideas
Scout produces the intelligence brief and up to three unranked, deduplicated seeds. Reviewer checks the packet. Chief of Staff sends recommendations and exceptions to you.
fallback 07:30 · report relay 07:45Directed research sprint
Ask the Chief of Staff to send Scout after a specific market, customer pain, hack, prior signal, or second-order effect. Run it multiple times per day.
Founder direction → Scout → Reviewer → CoSOpportunity Foundry
The Foundry processes open SEED and OPP records, deduplicates, scores, creates eligible canonical opportunities, and proposes movement. It never decides for you.
06:20 · no automatic founder decisionSimple rule: Scout discovers problems and evidence. Reviewer challenges. Foundry ranks. Chief of Staff recommends. You decide. Builder builds only from an approved intent.
02 / Capture model
Every opportunity has a traceable origin.
Do not treat the Scout brief as the only source of ideas. The Problem Observatory converts recurring, observable problems into a durable evidence chain. The Foundry then compares that chain with existing ideas, opportunities, products, companies, prospects, decisions, and outcomes.
01EVD
→
Evidence item
A dated, source-linked observation with a bounded claim, source family, terms mode, and evidence grade.
inbox/evidence/*.json02PROB
→
Problem dossier
A recurring problem assembled from evidence: affected actor, workflow, workaround, consequence, buyer hypothesis, counterevidence, and disproof test.
inbox/problems/*.json03SEED
→
Idea seed
An unranked, model-neutral mechanism or solution hypothesis linked to one or more PROBs. It is not yet a canonical opportunity.
inbox/idea-seeds/*.json04OPP
Canonical opportunity
A Foundry-admitted, deduplicated opportunity with commercial routes, score, dependencies, synergies, validation plan, and founder decision state.
opportunities/OPP-*.mdDaily and directed lanes create EVD / PROB / SEED
Scheduled Observatory runs and founder-directed Scout sprints may create evidence, problem dossiers, and up to ten unranked seeds per run. They do not rank, create canonical OPP records, contact anyone, spend, build, publish, deploy, or change stages.
Only the Foundry creates or refreshes OPP
On Monday and Thursday, the Foundry reconciles open SEED and OPP records. A new OPP requires linked PROB records, at least two independent source families, E2-or-better evidence, counterevidence, and a falsifiable disproof test. Ten is a maximum delta, not a quota.
Founder interest is not the same as stage
You may promote, downgrade, hold, validate, build, merge-review, or kill. The agent recommendation explains the trade-off; it does not change founder_interest, stage, budget, or execution authority without your exact approval.
Source laneWhat it can revealWhat it cannot prove alone
Hacker NewsAttention, practitioner friction, emerging technical patternsBuyer intent or willingness to pay
Stack ExchangeRepeated technical and operational questions, failed workaroundsMarket size or commercial urgency
Greenhouse / LeverResource commitments and roles companies are hiring to fillThat the role’s problem is underserved or buyable
News / web researchChange, regulation, incidents, category movement, second-order effectsFirst-hand customer pain without corroboration
Reddit / Product HuntResearch-only community language, demand signals, product reactionsCommercial use, raw-content reuse, training, or outreach until terms and credentials are cleared
Quality rule: A “good idea” is not a good opportunity. An opportunity earns promotion only when the problem is corroborated, the buyer and route are plausible, the alternative is understood, and the fastest disproof test is explicit.
03 / Prompt library
Copy, paste, and replace the brackets.
Use the Chief of Staff as the normal front door. The templates are deliberately explicit about evidence, authority, and stop conditions so the system can operate without guessing.
Daily company intelligence
Use when you want the normal intelligence loop and its small, brief-linked seed set.
Run today's Scout intelligence workflow under the active operating rules. Produce the complete brief and up to three evidence-backed, deduplicated, unranked idea seeds linked to the brief. Also report any Problem Observatory EVD, PROB, or SEED records created today; do not treat them as ranked opportunities. Preserve source traceability, counterevidence, inference labels, evidence grades, and falsifiable disproof tests. Do not rank the portfolio, create canonical opportunities, contact anyone, spend, build, publish, deploy, or change stages. Reviewer must verify the packet, then Chief of Staff must send me recommendations with justification. I remain the decision owner.Run a Problem Observatory sprint
Use daily or on demand when you want problem-first research beyond the intelligence brief.
Run a Problem Observatory sprint on [problem space / actor / workflow]. Use the active source rotation and at least three independent source families where available: Hacker News, Stack Exchange, Greenhouse, Lever, news/web, and approved research-only adapters. Capture dated EVD records with URL, source family, terms mode, bounded claim, evidence grade, and retrieval time. Cluster recurring evidence into PROB dossiers with affected actor, current workaround, consequence, buyer hypothesis, counterevidence, and a falsifiable disproof test. Create up to ten deduplicated, unranked, model-neutral SEED records. Do not create canonical OPP records, rank, promote, merge, contact, spend, build, publish, deploy, or change stages. Reviewer must verify the packet and Chief of Staff must return recommendations with justification. I remain the decision owner.Directed Scout Research Sprint
Use for a market, customer pain, hack, prior Scout signal, or second-order effect. This is the first-class multi-run lane.
Scout Research Sprint: investigate [direction]. Context or signal: [link, prior report, problem ID, customer pain, or hypothesis]. Use the relevant Problem Observatory source lanes and explain: (1) what changed or is structurally true, (2) who has the pain and why now, (3) what workarounds and alternatives exist, (4) what could be sold without becoming generic consulting, (5) what would disprove it quickly. Create or update EVD and PROB records as warranted, then return up to ten deduplicated, unranked SEED records with sources, evidence grades, counterevidence, duplicate classification, commercialisation options, and confidence. Do not create canonical OPP records, rank, promote, merge, contact, spend, build, publish, deploy, or change stages. Chief of Staff should give me recommendations; I make the decision.Turn a signal into a real opportunity packet
Use after a seed looks interesting and you want a tangible business model, not a loose narrative.
Build or refresh the opportunity packet for [OPP ID, SEED ID, or PROB ID]. Trace the full EVD → PROB → SEED → OPP lineage and deduplicate against the repository, GBrain, existing products, companies, and prospects. Include: target customer and reachable buyer; painful current workflow; why-now evidence; alternatives and incumbent workaround; one-sentence offer; SaaS, service, marketplace, channel, licensing, or other viable business-model options; distribution routes; first 20 reachable prospects; gross-margin and delivery assumptions; regulatory and trust risks; defensibility; synergies; smallest disproof test; 14-day validation plan; success and kill thresholds. Label every claim observed, source-backed, inferred, disputed, or unknown. If admission criteria are not met, keep it as SEED/PROB and say exactly what is missing. Do not advance the stage or contact prospects. End with a recommendation, confidence, and exact founder approvals required.Process the SEED and OPP repository
Use on demand for a read-only preview, or let the Monday/Thursday Foundry run use the same rules.
Run the next Opportunity Foundry intake for [all open records / selected SEED IDs / selected OPP IDs]. Reconcile EVD → PROB → SEED → OPP lineage; deduplicate against existing records and classify duplicate, variant, complement, conflict, or novel. For each proposed NEW OPP, require linked PROBs, at least two independent source families, E2-or-better evidence, counterevidence, a falsifiable disproof test, buyer and route hypotheses, and a bounded validation plan. Create or update no more than ten OPP deltas in this run. Produce scores, confidence, synergies, missing evidence, and recommended promote/hold/validate/merge-review/kill actions. Do not change founder interest, stage, budget, contact anyone, spend, build, publish, deploy, or execute a recommendation without my exact approval.Inspect an opportunity’s provenance
Use when you want to understand exactly why an OPP exists and what would invalidate it.
Show the provenance and current decision packet for [ID]. Start at the EVD records and trace every linked PROB, SEED, OPP, source family, evidence grade, counterevidence, inference, deduplication decision, score component, synergy, owner, receipt, and last validation result. Separate observed facts, source-backed claims, inferences, recommendations, and unknowns. Show the fastest falsification test, current gate, stale evidence, and the one founder decision that would change the next state. Do not mutate any record.Portfolio Council review
Use on Monday or Thursday, or ask for a readout of the latest Foundry run.
Show me the latest Opportunity Foundry portfolio ranking. Compare [OPP IDs or “all active opportunities”] on verified pain, reachable buyers, time to first revenue, contribution margin, founder fit, defensibility, delivery complexity, regulatory risk, synergies, and evidence quality. Separate score from judgement. Recommend: promote, hold, validate, build, merge-review, or kill. Explain alternatives and trade-offs. Do not change founder_interest, stage, budget, or status until I explicitly approve the exact change.Record a bounded founder decision
Use when the Chief of Staff presents a recommendation and you want to approve only a precise slice.
Decision for [decision ID]: [APPROVE / REJECT / MODIFY]. Exact scope: [what may happen]. Exact target: [IDs, recipients, repository, environment]. Budget: [$ amount or $0]. Expiry: [YYYY-MM-DD]. Stop conditions: [conditions]. Everything outside this scope remains blocked. Record the approval receipt before dispatch and show me the resulting task graph and idempotency key.Start a product build
Use only after the opportunity and build intent are approved.
Build [product / feature] from approved intent [intent ID]. First show: scope, non-goals, user story, acceptance criteria, authority tier, budget, workspace, dependencies, rollback plan, and approval gates. Use Arena → Pi → DeepSeek V4 Flash. For every UI change, apply the Impeccable workflow and record the result. Run real tests, lint, typecheck, build, security checks, and preview deployment. Do not merge, publish, change production, or contact customers. Hand the complete diff and release receipt to Reviewer.Request an independent release review
Use when Builder says a product change is ready.
Review [repository / task / commit] independently. Check the approved intent, acceptance criteria, diff scope, Impeccable UI evidence where applicable, tests, lint, typecheck, build, security findings, dependency changes, preview behavior, migration safety, rollback steps, and observability. Return PASS, CONDITIONAL PASS, or FAIL with evidence. Do not merge or deploy. State the smallest corrective action and the exact founder approval still required.Prepare an external discovery batch
Agents draft and research; external sending remains R2 and founder-approved.
Prepare an external customer-development batch for [opportunity / ICP]. Maximum recipients: [number, no more than 10]. Channel: [email / LinkedIn]. For each recipient, show why they fit, the evidence to learn, the exact short message, consent or opt-out handling, and the expected next step. Do not send, add to CRM, scrape, bulk automate, or spend. Give me one bounded approval request containing exact recipients, exact copy, batch ID, expiry, and idempotency key.Handle an exception or failed run
Use whenever a cron, receipt, provider, deployment, or reconciliation looks wrong.
Open an incident review for [task / run / provider]. Freeze downstream side effects. Show the last known good receipt, current state, failed step, suspected cause, duplicate risk, affected artifacts, and whether any external action occurred. Reconcile the ledger and Kanban state. Propose the smallest safe recovery, rollback, or retry with exact idempotency key. Do not retry blindly, delete evidence, send messages, spend, deploy, or change credentials. Escalate anything outside R1.Run the weekly operating review
Use when you want the company view without opening every task.
Run a weekly operating review for Offdata. Summarise: outcomes shipped, evidence gained, customer signals, portfolio movement, revenue hypotheses, KPI movement, failed assumptions, blocked work, stale receipts, cron health, provider/API use, cost, GBrain health, open approvals, recovery risks, and the three highest-leverage next actions. Separate facts from recommendations. Do not create new work or change priorities without showing me the decision packet first.Perform a safe system update
Use before updating Hermes, Arena, integrations, or shared tooling.
Prepare a safe update for [component]. First inventory versions, uncommitted changes, running jobs, credentials, scheduled work, migrations, and rollback points. Create a reversible checkpoint without overwriting local changes. Update in an isolated lane, run tests and smoke checks, verify Scout → Reviewer → CoS, inspect schedules and tool allow-lists, and compare configuration before/after. Do not delete, reset, migrate production, or resume normal schedules until I see the update receipt and approve the recovery plan.No templates match that search. Try “approval”, “Scout”, “build”, or “recovery”.
04 / Authority
Make approvals narrow enough to be safe.
An approval is action-specific. It must name the exact target, scope, budget, expiry, and idempotency key. A casual “yes” is not a production approval.
Approval builder
Draft the exact approval request
Nothing is sent or activated by this page. Copy the generated text into the Chief of Staff bot for review.
Generated approval
05 / Coordination
Know what each profile owns.
Keep your founder inbox small. Backend workers communicate through intent-linked tasks, receipts, and handoffs.
RoleOwnsYou contact it for
Chief of StaffIntent intake, routing, synthesis, exceptions, approvalsEverything strategic; your single front door
ScoutObservatory, signals, research, EVD, PROB, SEEDUsually via Chief of Staff; direct sprint output when needed
Portfolio Architect / RevenueOPP comparison, ICP, offers, GTM, customer-development draftsPortfolio recommendations and commercial hypotheses
Builder / ArenaProduct implementation through Pi, tests, previewApproved build intents only
ReviewerIndependent evidence, quality, security, release gateReview receipts and conditional passes
Operator / ControllerDeployment, provider actions, reconciliation, economicsBackend only; Chief of Staff returns exceptions
Scout→Reviewer→Foundry→Chief of Staff→Founder→Builder / Revenue / Operator
06 / Portfolio
Use the idea repository as a memory, not a graveyard.
Every seed is captured once, labelled, deduplicated, and linked to evidence. The Foundry compares projections, impact, cost, customer evidence, synergies, and founder interest before proposing movement.
01
Capture
Daily Scout and directed sprints create unranked seeds with provenance and disproof tests.
02
Challenge
Reviewer checks novelty, counterevidence, inference labels, and whether a seed is a duplicate, variant, complement, or conflict.
03
Rank
Monday/Thursday Foundry scores attention, revenue path, margin, feasibility, defensibility, risk, and synergy.
04
Decide
You promote, downgrade, hold, validate, build, merge-review, or kill. The score never replaces customer evidence or founder judgement.
Founder preference: In the Obsidian opportunity Base, change only
founder_interest, founder_rationale, and next_action. Use the Chief of Staff for stage, budget, merge, and execution decisions.07 / Product engine
Build a solution, then prove the release.
A product task does not become a release because Builder says “done.” Arena and Reviewer must leave an evidence trail.
01
Intent
Approved outcome, user, non-goals, authority, budget, acceptance criteria, rollback.
02
Design
For UI work, use Impeccable: inspect, critique, harden, and polish before release review.
03
Build
Arena orchestrates Pi and DeepSeek V4 Flash in an isolated worktree with deterministic gates.
04
Review
Reviewer checks code, receipts, tests, security, preview, rollback, and scope.
05
Operate
Operator deploys only within authority; Controller reconciles events, cost, revenue, and exceptions.
06
Founder gate
Merge, production deployment, external communication, pricing, refunds, and legal commitments remain R2.
08 / Resilience
Updates should leave the system more recoverable.
Before every update, checkpoint user changes and verify that the update did not overwrite profiles, cron jobs, skills, credentials, policies, or local repositories.
Before
Run git status, Hermes backup, configuration inventory, enabled-job inventory, and migration review. Record the last known good commit.
During
Update in an isolated lane. Preserve local edits. Keep credentials private. Do not reset or delete to make a tool “clean.”
After
Run tests, schedule/tool-surface checks, GBrain/embedding health, Scout → Reviewer → CoS smoke test, Arena gates, and a reconciliation pass.
If wrong
Freeze downstream effects, preserve receipts, reclaim stale workers, restore the last checkpoint, and only then resume scheduling.
09 / Founder checklist
What to do in five minutes.
Use this as the everyday operating loop. The system is designed to keep routine work quiet and bring only decisions, exceptions, and evidence to you.
Remember: the agents can find patterns, build evidence, draft communications, compare routes, and prepare systems. They do not manufacture demand or make your strategic decision. Your job is to choose the next bounded bet.