How to build
complex software
with AI

1. 001Start with a rough idea rather than a usable engineering specification.Encountered / reported
2. 002Decide whether the intended outcome is a business, internal tool, client project, personal utility, or experiment.Documented decision / task
3. 003Identify who benefits from the product and who actually pays for it.Documented decision / task
4. 004Write down what a successful first version must accomplish.Documented decision / task
5. 005Separate the founder's vision from the implementation the AI happens to suggest.Documented decision / task
6. 006Explain the same product thesis to ChatGPT, Claude, Codex, and separate agents.Encountered / reported
7. 007Decide what belongs in the first usable slice.Documented decision / task
8. 008Determine what can be deferred without losing the central promise.Documented decision / task
9. 009Change direction when new evidence reveals a wrong first approach.Encountered / reported
10. 010Record rejected directions so the next AI doesn't revive them.Documented decision / task
11. 011Decide how much founder involvement is wanted versus what can be delegated.Documented decision / task
12. 012Identify nonnegotiable customer ownership and spending limits.Documented decision / task
13. 013Identify available time, budget, risk tolerance, and launch urgency.Documented decision / task
14. 014Decide which actions need founder approval versus an automated safe default.Documented decision / task
15. 015Define evidence that something is actually finished.Documented decision / task
16. 016Distinguish a functioning demo from a durable real-world result.Documented decision / task
17. 017Re-explain why generating more code was not necessarily progress.Encountered / reported
18. 018Establish an explicit finish line before opening another task lane.Documented decision / task

1. 019Research the actual pain of building ambitious products with AI.Encountered / reported
2. 020Identify the buyer, everyday user, and alternatives they already know.Documented decision / task
3. 021Study Lovable, Replit, Base44, Cursor, Claude, Codex, and emerging competitors.Encountered / reported
4. 022Separate a competitor's advertised capability from a proved customer outcome.Documented decision / task
5. 023Build a competitor matrix without assuming everyone else lacks planning or testing.Documented decision / task
6. 024Compare your idea with existing tools, including the project tools you already use.My real build: agent managers and Linear.Encountered / reported
7. 025Ask which parts of the existing product idea could be solved by buying or integrating something else.Documented decision / task
8. 026Decide when founders need an initial prototype and when they need to graduate an existing one.Documented decision / task
9. 027Clarify why BYOA is more than an extra connector.Encountered / reported
10. 028Define the smallest compelling differentiating experience.Documented decision / task
11. 029Rework the pitch after it sounded like a generic orchestration product.Encountered / reported
12. 030Research which founder segments experience the pain most acutely.Documented decision / task
13. 031Test whether the pain is actually worth paying to remove.Documented decision / task
14. 032Separate technical success from customer adoption and business value.Documented decision / task
15. 033Decide whether a feature advances the founder's actual goal.Documented decision / task
16. 034Seek candid feedback from potential design partners rather than assuming the idea sells itself.Encountered / reported
17. 035Compare market evidence, founder reports, provider claims, and AI-generated inference.Documented decision / task
18. 036Reevaluate the project thesis as stronger builders acquire yesterday's differentiators.Documented decision / task

1. 037Start with an internal code name so naming doesn’t block the work.My real build: Relay.Encountered / reported
2. 038Explore alternatives that don't sound like generic AI-product jargon.Encountered / reported
3. 039Reject names that feel clichéd, confusing, or too close to existing brands.Encountered / reported
4. 040Explore the connection between praxis, judgment, and purposeful action.Encountered / reported
5. 041Shortlist a distinctive public-facing name.My real build: Praxeri.Encountered / reported
6. 042Check the availability of the exact-match .com domain.Encountered / reported

1. 058Choose the legal entity that will run your beta.Encountered / reported
2. 059Resolve personal funding versus company ownership of product assets.Documented decision / task
3. 060Clarify member and intellectual-property ownership arrangements.Documented decision / task
4. 061Document rights to code, domain, design, and contractor-created work.Documented decision / task
5. 062Decide when a separate new company becomes appropriate.Documented decision / task
6. 063Check entity name, registered details, and public legal contact.Documented decision / task

1. 080Discover that connecting an AI to a project also creates a data-handling product.Encountered / reported
2. 081Inventory every category of information the application collects.Documented decision / task
3. 082Account for prompts, responses, conversation history, files, and repository snippets.Documented decision / task
4. 083Account for the project’s saved decisions, archived assumptions, and uploaded documents.Documented decision / task
5. 084Account for OAuth authorization details and connection identifiers.Documented decision / task
6. 085Account for error logs, diagnostic traces, model usage, and QA screenshots.Documented decision / task

1. 105Learn the difference between a successful login and sufficient project permissions.Encountered / reported
2. 106Diagnose missing GitHub App permissions and account selection.Encountered / reported
3. 107Encounter permission problems only after attempting a repository move.Encountered / reported
4. 108Determine the minimum scopes required for each tool and operation.Documented decision / task
5. 109Separate personal accounts, organizations, workspaces, projects, and repositories.Documented decision / task
6. 110Handle OAuth authorization and callback redirect configuration.Documented decision / task
7. 111Deal with different visible states for connected versus verified services.Encountered / reported
8. 112Store tokens in appropriate secret management, not plain prompts.Documented decision / task
9. 113Keep secrets out of worker context, browser JavaScript, and logs.Documented decision / task
10. 114Avoid placing API keys in ordinary chat or Git commits.Documented decision / task
11. 115Decide whether a provider key is for one project, one environment, or an entire organization.Documented decision / task
12. 116Use separate development, staging, and production credentials where justified.Documented decision / task
13. 117Confirm the account that owns each resource before making changes.Documented decision / task
14. 118Test token expiry, revocation, reconnect, and refresh failures.Documented decision / task
15. 119Verify least-privilege access after every expanded App permission request.Documented decision / task
16. 120Control what third-party tools can see about a private project.Documented decision / task
17. 121Document when consent is required before sharing source or strategy externally.Documented decision / task
18. 122Avoid reusing a founder's personal browser profile in cloud tests.Documented decision / task
19. 123Prove another founder cannot read the current founder's data.Documented decision / task
20. 124Prove an agent cannot approve its own sensitive tool access.Documented decision / task
21. 125Decide who can publish, delete, buy, transfer, or change production settings.Documented decision / task
22. 126Test secret redaction at every known output and evidence boundary.Documented decision / task
23. 127Retest privacy assumptions after changing the execution architecture.Encountered / reported
24. 128Prepare an incident-response plan for leaks, bad permissions, or tenant crossover.Documented decision / task

1. 129Start building heavily with Claude and evaluate Codex for independent tasks.Encountered / reported
2. 130Compare coding-agent subscriptions with metered provider APIs.Encountered / reported
3. 131Discover that an AI subscription is not the same as API usage.Encountered / reported
4. 132Check which subscription-based integrations are actually authorized for commercial use.Documented decision / task
5. 133Investigate why usage quotas suddenly disappear faster than expected.Encountered / reported
6. 134Measure huge cache-read and cache-write token volumes, not just visible output.Encountered / reported
7. 135Discover a long-lived Claude session had grown beyond 750K context tokens.Encountered / reported
8. 136Discover the cost of multi-agent and helper loops multiplying context processing.Encountered / reported
9. 137Determine which provider and model should handle each task type.Documented decision / task
10. 138Reserve higher-cost inference for work where stronger reasoning helps.Documented decision / task
11. 139Test whether cheap models are adequate for summaries, triage, or other bounded tasks.Documented decision / task
12. 140Compare open-weight models without assuming cheap inference is free infrastructure.Documented decision / task
13. 141Keep API billing, subscription allowances, and cloud-sandbox costs separate.Documented decision / task
14. 142Prevent a fallback from silently switching a user to a different billing method.Documented decision / task
15. 143Estimate costs before authorizing metered tasks.Documented decision / task
16. 144Set monthly, daily, and per-run spend limits.Documented decision / task
17. 145Check a provider's limits and actual reset windows before scheduling work.Documented decision / task
18. 146Keep sufficient allowance for likely repairs and independent checks.Documented decision / task
19. 147Track total cost per accepted task, not cost per agent response.Documented decision / task
20. 148Decide when not to run a model at all because a script can do the job.Documented decision / task
21. 149Move more execution to Codex and realize reduced Claude usage alone doesn't prove savings.Encountered / reported

1. 150Work across local and hosted execution environments—and find out why those behave differently.Encountered / reported
2. 151Determine what runs on the user's laptop versus the cloud.Documented decision / task
3. 152Test local worker connectivity and native provider authentication.Encountered / reported
4. 153Investigate Windows-specific command and tool failure behavior.Encountered / reported
5. 154Determine whether Git, Python, Node, npm, Docker, or WSL is truly required.Documented decision / task
6. 155Avoid requiring beginners to install an entire developer toolchain by default.Documented decision / task
7. 156Develop a desktop helper and discover that signed installers, safe updates, and setup support are separate jobs.Encountered / reported
8. 157Diagnose a working Helper update hidden behind the wrong UI control.Encountered / reported
9. 158Encounter local Helper/desktop paths that distracted from cloud beta delivery.Encountered / reported
10. 159Decide to park the desktop Helper for the initial cloud-only beta.Encountered / reported
11. 160Prove a clean first-use environment without relying on the developer's machine.Documented decision / task
12. 161Distinguish local model subscription access from provider-authorized hosted access.Documented decision / task
13. 162Test file permissions and selected-folder access rather than trusting folder labels.Documented decision / task
14. 163Determine which environment variables change the active billing route.Documented decision / task
15. 164Document supportable OS/browser/platform combinations.Documented decision / task
16. 165Use a clean VM or supported isolated environment for novice-install proof when needed.Documented decision / task
17. 166Diagnose missing runtime dependencies without telling users to bypass OS protections.Documented decision / task
18. 167Make installations and upgrades recoverable and reversible.Documented decision / task
19. 168Build a genuine cloud path that does not require keeping the founder's laptop on.Documented decision / task

1. 169Set up and reconfigure GitHub Apps and authorizations.Encountered / reported
2. 170Connect and troubleshoot Claude authentication and API-key alternatives.Encountered / reported
3. 171Connect and test Notion workspaces and page access.Encountered / reported
4. 172Find that an authorized Notion workspace didn't expose the desired project page.Encountered / reported
5. 173Handle the difference between discovering a resource and actually being able to edit it.Encountered / reported
6. 174Investigate Notion reconnect and add-page permission problems.Encountered / reported

1. 196Develop important product decisions in ChatGPT conversations.Encountered / reported
2. 197Carry those decisions into Claude prompts for implementation.Encountered / reported
3. 198Pass Claude's build findings back for independent analysis.Encountered / reported
4. 199Paste review findings back into the build conversation.Encountered / reported
5. 200Repeat the loop when a reviewer finds another issue.Encountered / reported
6. 201Keep Notion pages to stop decisions disappearing into old chats.Encountered / reported
7. 202Correct Notion pages when new decisions supersede older ones.Encountered / reported
8. 203Review long handoff packets and rewrite instructions when they omit a constraint.Encountered / reported
9. 204Ask one AI to search earlier conversations before handing work to another.Encountered / reported
10. 205Discover that chat history doesn't equal reliable project memory.Encountered / reported
11. 206Use structured context packets to avoid repeating the entire founder story.Encountered / reported
12. 207Explain that a previously rejected architecture is still rejected.Encountered / reported
13. 208Explain which current codebase and deployment an agent must inspect.Encountered / reported
14. 209Tell a new agent not to trust a previous agent's completion claim.Encountered / reported
15. 210Carry the correct release versions, task identities, review notes, and approvals across handoffs.Encountered / reported
16. 211Mark facts as founder-stated, independently observed, inferred, or unknown.Documented decision / task
17. 212Preserve source/date and confidence for consequential project decisions.Documented decision / task
18. 213Separate historical raw conversations from authoritative current project state.Documented decision / task
19. 214Select only task-relevant information for the next builder.Documented decision / task
20. 215Check whether a memory entry became stale after actual code changed.Documented decision / task
21. 216Design a fresh-session handoff that doesn't require the founder to write a novel.Documented decision / task
22. 217Find that a long checkpoint page forced readers to wade through historical clutter.Encountered / reported
23. 218Audit instruction files whose account of the project conflicts with the actual code.Encountered / reported
24. 219Keep a record of every accepted founder correction so it isn't asked again.Documented decision / task

1. 220Document goals, scope, priorities, rejected options, and current dependencies.Encountered / reported
2. 221Try maintaining durable project decisions in a dedicated project-memory system rather than hoping chat remembers everything.Encountered / reported
3. 222Give those decisions a structured history so changes can be tracked and corrected.Encountered / reported
4. 223Decide what the Project Brain is authoritative about—and what the code is authoritative about.Documented decision / task
5. 224Distinguish active, superseded, disputed, archived, and unknown facts.Documented decision / task
6. 225Preserve evidence for why a requirement changed.Documented decision / task
7. 226Track source project identities across builders and GitHub migration.Documented decision / task
8. 227Check that imported research isn't silently promoted into verified truth.Documented decision / task
9. 228Correct a case where a founder's own statement had been treated as independently observed evidence.Encountered / reported
10. 229Decide when a worker must ask for more context rather than invent an answer.Documented decision / task
11. 230Test context continuity across browser refresh, chat threads, and model handoffs.Documented decision / task
12. 231Verify that all unfinished obligations survive a new session.Documented decision / task
13. 232Keep provider-specific implementation details out of universal founder memory.Documented decision / task
14. 233Create small task-sized context packets instead of sending the full Project Brain.Documented decision / task
15. 234Resolve contradictions before authorizing consequential execution.Documented decision / task
16. 235Protect private project context from cross-workspace disclosure.Documented decision / task
17. 236Retain history without forcing each new model to reread everything.Documented decision / task
18. 237Give founders a readable view of what the system believes about their project.Documented decision / task

1. 238Try building a separate evidence system to learn whether AI-assisted work actually helped.Encountered / reported
2. 239Reconstruct past execution events from code history, AI sessions, and review records.Encountered / reported
3. 240Distinguish actual timestamps from reconstructed or uncertain ones.Encountered / reported
4. 241Connect recommendations to later decisions and accepted outcomes.Documented decision / task
5. 242Track which model, agent, role, and reviewer actually did the work.Documented decision / task
6. 243Record retries, repair rounds, failed gates, and founder interventions.Documented decision / task
7. 244Distinguish useful failures found from duplicated scanner noise.Documented decision / task
8. 245Measure cost per founder-accepted task rather than passed run.Documented decision / task
9. 246Record what was tested and what remained unverified.Documented decision / task
10. 247Learn when an extra specialist improves the outcome versus wastes money.Documented decision / task
11. 248Prevent one customer's proprietary code from becoming another customer's model context.Documented decision / task
12. 249Document which model predictions were contradicted by actual runtime behavior.Documented decision / task
13. 250Use outcome evidence to improve future decisions without letting an AI silently rewrite important policy.Documented decision / task
14. 251Decide when a practice should be versioned, superseded, or retired.Documented decision / task
15. 252Separate observed prevention from made-up estimates of hypothetical savings.Documented decision / task
16. 253Build a truthful operator view of learning, cost, and quality.Documented decision / task
17. 254Accept that a system can collect lots of events and still lack outcome linkage.Documented decision / task

1. 255Decide when one Claude session is enough.Encountered / reported
2. 256Experiment with Claude subagents and helper tasks.Encountered / reported
3. 257Experiment with concurrent work and many-agent review sessions.Encountered / reported
4. 258Find that more agents can generate more coordination and usage than useful progress.Encountered / reported
5. 259Define bounded roles for planning, building, security, UX, and independent review.Documented decision / task
6. 260Choose which roles genuinely need different models or providers.Documented decision / task
7. 261Decide when parallel tasks are safe and when they share a dangerous file.Documented decision / task
8. 262Assign a single implementation owner for a touched code area.Documented decision / task
9. 263Create independent reviewer roles without letting them overwrite the builder's work.Documented decision / task
10. 264Set agent permissions for source, filesystem, shell, network, and credentials.Documented decision / task
11. 265Limit agent runtime, context, tool calls, retries, and spending.Documented decision / task
12. 266Prevent agents from reviving abandoned tasks because they saw stale plans.Documented decision / task
13. 267Prevent two agents from changing the same shared state without coordination.Documented decision / task
14. 268Deal with agent passbacks that required additional prompts and clarification.Encountered / reported
15. 269Retest a repair after a different agent's review contradicted the first one.Encountered / reported
16. 270Record exact changes and evidence from every worker, not just a confident summary.Documented decision / task
17. 271Retire temporary agents and clean their branches when work is complete.Documented decision / task
18. 272Detect when repeated repairs mean the approach is wrong, not just the code.Documented decision / task
19. 273Stop expensive browser and multi-agent loops when they became unproductive.Encountered / reported
20. 274Define an inexpensive single-builder path for ordinary low-risk changes.Documented decision / task
21. 275Keep founder attention focused on approvals, not agent staffing decisions.Documented decision / task

1. 276Build and repeatedly refine a project roadmap and milestone structure.Encountered / reported
2. 277Create task lanes and independent review gates.Encountered / reported
3. 278Reconcile duplicate and superseded checklists.Encountered / reported
4. 279Discover that a huge master checklist can itself become a navigation problem.Encountered / reported
5. 280Choose one active source for task ownership and status.Documented decision / task
6. 281Distinguish planned, built, tested, reviewed, integrated, deployed, and accepted.Documented decision / task
7. 282Assign dependencies and explicit blockers before starting parallel work.Documented decision / task
8. 283Limit how many tasks can be actively built at once.Documented decision / task
9. 284Keep a founder-visible short list of the actual next decisions.Documented decision / task
10. 285Check whether a task still advances the accepted finish line.Documented decision / task
11. 286Challenge feature creep when beta is close.Documented decision / task
12. 287Park worthwhile ideas without letting them interrupt a release.Documented decision / task
13. 288Decide when a failed attempt merits repair, rollback, or a new design.Documented decision / task
14. 289Retain historical failures without resurfacing every one as a new task.Documented decision / task
15. 290Ask which existing components already solve a requested problem.Documented decision / task
16. 291Record why new vendors and infrastructure were rejected.Documented decision / task
17. 292Find an internal tool that was built but never connected to the actual experience.Encountered / reported
18. 293Keep acceptance criteria behavior-focused instead of just output-focused.Documented decision / task

1. 294Create source repositories and choose ownership accounts.Encountered / reported
2. 295Create branches for experiments, integrations, and fixes.Encountered / reported
3. 296Create worktrees so parallel agents don't stomp on each other.Encountered / reported
4. 297End up with hundreds of worktrees and branches needing inventory.Encountered / reported
5. 298Discover a worktree folder name didn't match its active branch.Encountered / reported
6. 299Find a default working folder hundreds of commits behind the live app.Encountered / reported
7. 300Find a live copy reportedly hundreds of commits ahead of the remote backup.Encountered / reported
8. 301Prove the exact deployed page and backend are tied to known source commits.Documented decision / task
9. 302Create off-machine recovery copies before dangerous cleanup.Documented decision / task
10. 303Inventory uncommitted, unmerged, and unpublished changes.Documented decision / task
11. 304Identify which branches contain useful work versus superseded experiments.Documented decision / task
12. 305Prevent an agent from building a feature on an old copy of the source code.Documented decision / task
13. 306Handle file conflicts between simultaneous agents.Documented decision / task
14. 307Choose safe merge/rebase policies for approved work.Documented decision / task
15. 308Check whether a fix already exists on another branch.Documented decision / task
16. 309Reconcile work copied across branches without pretending it merged cleanly.Documented decision / task
17. 310Verify the diff instead of trusting the agent's commit description.Documented decision / task
18. 311Test protected-baseline rules before promotion.Documented decision / task
19. 312Require actual source ownership transfer, not an empty replacement repository.Documented decision / task
20. 313Avoid deleting customer source during trial-repository cleanup.Documented decision / task
21. 314Preserve commit history and relevant project evidence across a transfer.Documented decision / task
22. 315Prove repo identity and branch state after an account migration.Documented decision / task
23. 316Record which exact SHA each reviewer actually examined.Documented decision / task
24. 317Clean stale worktrees only after approving the recovery plan.Documented decision / task

1. 318Try to create the interface using verbal prompts alone.Encountered / reported
2. 319Discover the design is functional but not good enough.Encountered / reported
3. 320Revise layouts, navigation, information hierarchy, and copy multiple times.Encountered / reported
4. 321Identify what the current product experience actually needs to communicate.Documented decision / task
5. 322Review other apps' UX patterns rather than relying only on model imagination.Documented decision / task
6. 323Experiment with Mobbin as an upstream reference tool.Encountered / reported
7. 324Find that generic design searches produced weak results.Encountered / reported
8. 325Learn to search for specific screens, interactions, states, and users.Documented decision / task
9. 326Avoid copying a single competitor design even if it looks great.Documented decision / task
10. 327Compare independent reference patterns before extracting a visual direction.Documented decision / task
11. 328Check usage and retention rights for third-party visual materials.Documented decision / task
12. 329Use the external source as inspiration, not a screenshot library to redistribute.Documented decision / task
13. 330Generate original concepts and founder-visible preview options where warranted.Documented decision / task
14. 331Obtain approval of a meaningful design direction before spending hours coding.Documented decision / task
15. 332Record the accepted design direction so the next agent doesn't redesign it.Documented decision / task
16. 333Review what actually rendered in the browser, not just what the code suggests.Documented decision / task
17. 334Work through competing design-mode ideas and tool choices.Encountered / reported
18. 335Reconcile screens that duplicated status or sent the user to the wrong place.Encountered / reported
19. 336Test desktop and phone behavior independently.Documented decision / task
20. 337Do not bury the essential action under impressive technical controls.Documented decision / task
21. 338Check accessibility, copy, empty states, and responsive hierarchy.Documented decision / task
22. 339Compare the final rendered experience against the accepted reference.Documented decision / task

1. 340Design multiple founder onboarding routes for existing projects and new ideas.Encountered / reported
2. 341Watch a novice test expose missing founder facts.Encountered / reported
3. 342See the user mention four interviews and still get told to conduct five.Encountered / reported
4. 343Discover a screen reporting setup success while its main action remained unusable.Encountered / reported
5. 344Find that a Chat Send button could be disabled after onboarding.Encountered / reported
6. 345Find buttons and links that worked in mocks but not the real user journey.Encountered / reported
7. 346Ask the founder's intended outcome before connecting ten tools.Documented decision / task
8. 347Handle users who don't know GitHub, CLI, PATH, worktrees, or environment variables.Documented decision / task
9. 348Explain billing choices without forcing an unfamiliar API setup.Documented decision / task
10. 349Preserve all previously entered customer and product facts.Documented decision / task
11. 350Avoid asking the founder to provide information already recorded.Documented decision / task
12. 351Make the first useful recommendation understandable without technical jargon.Documented decision / task
13. 352Show a clear recovery action if an external account is unavailable.Documented decision / task
14. 353Place only meaningful approvals in the Needs You queue.Documented decision / task
15. 354Provide contextual help tied to the real error, not generic documentation links.Documented decision / task
16. 355Check that completion takes the founder somewhere useful.Documented decision / task
17. 356Avoid repeatedly showing connected services as unverified in contradictory screens.Documented decision / task
18. 357Prove the experience on the actual deployed site with a fresh account.Documented decision / task
19. 358Ensure the product is usable on phone-sized screens where supported.Documented decision / task

1. 359Build a web frontend and a reliable backend that actually agree on the state of a project.Encountered / reported
2. 360Keep project and workspace identities consistent across the different services involved.Encountered / reported
3. 361Define source of truth for projects, members, resources, and tasks.Documented decision / task
4. 362Choose when to use hosted databases versus local files.Documented decision / task
5. 363Separate the founder chat interface from the coding worker's actual execution.Documented decision / task
6. 364Turn chat intent into a bounded, approved work package.Documented decision / task
7. 365Preserve source, task, and result identities through the whole workflow.Documented decision / task
8. 366Treat provider requests and real actions differently from simulated demos.Documented decision / task
9. 367Fix server errors surfaced only during realistic browser navigation.Encountered / reported
10. 368Find a missing backend field caused a Home 500 while a cloud task existed.Encountered / reported
11. 369Separate routine automated changes from actions requiring real owner authorization.Documented decision / task
12. 370Make retried operations idempotent or safely recoverable.Documented decision / task
13. 371Plan concurrency behavior when several workspaces run simultaneously.Documented decision / task
14. 372Avoid accidental cross-project state writes.Documented decision / task
15. 373Keep credentials separate from visible task packets.Documented decision / task
16. 374Expose useful progress without leaking raw internal paths or secrets.Documented decision / task
17. 375Return build outcomes to the same founder conversation that initiated them.Documented decision / task
18. 376Keep architecture modules interoperable without creating duplicate task systems.Documented decision / task
19. 377Distinguish implementation success from feature acceptance and deployment.Documented decision / task

1. 378Evaluate Vercel Sandbox for isolated cloud work.Encountered / reported
2. 379Wire and test a hosted runner instead of relying on a founder laptop.Encountered / reported
3. 380Determine when a sandbox is needed and when a simple API call suffices.Documented decision / task
4. 381Reserve compute cost before starting a metered execution.Documented decision / task
5. 382Limit execution time, wall clock, process count, and egress.Documented decision / task
6. 383Keep each task in an isolated workspace and source boundary.Documented decision / task
7. 384Limit what a build task can access, especially sensitive accounts and credentials.Documented decision / task
8. 385Check whether the necessary CLI/model provider can run in that environment.Documented decision / task
9. 386Discover some promising desktop/QA tooling did not work in hosted cloud execution.Encountered / reported
10. 387Try Midscene and conclude it wasn't usable in the present cloud setup.Encountered / reported
11. 388Continue with plain Playwright instead of pretending Midscene was integrated.Encountered / reported
12. 389Use a single authorized coding writer where shared files are involved.Documented decision / task
13. 390Let parallel research or review happen without multiple workers overwriting the same source files.Documented decision / task
14. 391Handle canceled builds and restore readable results after interruption.Documented decision / task
15. 392Prevent two jobs from spending the final budget allocation simultaneously.Documented decision / task
16. 393Stop idle sandboxes and reclaim temporary resources.Documented decision / task
17. 394Record the exact model payer and infrastructure payer per run.Documented decision / task
18. 395Make a failed sandbox report its true failure instead of fake success.Documented decision / task
19. 396Prove that cloud workers do not depend on hidden local machine state.Documented decision / task

1. 397Prototype private cloud previews before broad public deployment.Encountered / reported
2. 398Learn that a Vercel account requirement can break the intended first-run experience.Encountered / reported
3. 399Design a preview route that works without the founder owning a Vercel team account.Documented decision / task
4. 400Bind the preview to the exact built source candidate.Documented decision / task
5. 401Create short-lived, user-bound preview access.Documented decision / task
6. 402Verify that private previews aren't exposed to the wrong viewer.Documented decision / task
7. 403Prevent unauthenticated access to a private preview.Documented decision / task
8. 404Review subdomain versus direct sandbox-host preview architectures.Documented decision / task
9. 405Encounter Safari/WebKit iframe and separate-tab behavior differences.Encountered / reported
10. 406Build an Open private preview fallback for blocked embedded frames.Encountered / reported
11. 407Expire and revoke preview access reliably.Documented decision / task
12. 408Check service-worker behavior doesn't preserve stale access.Documented decision / task
13. 409Separate a preview from staging and live publication.Documented decision / task
14. 410Show an honest error when a preview cannot be opened.Documented decision / task
15. 411Handle preview stop, reopen, and idle-expiry behavior.Documented decision / task
16. 412Track preview compute and hosting costs independently of AI token usage.Documented decision / task
17. 413Test desktop and phone access where the claimed product supports it.Documented decision / task
18. 414Prove the owner can access the preview through the real product UI.Documented decision / task
19. 415Never claim a static screenshot proves a working app preview.Documented decision / task

1. 416Write and run large Python test suites.Encountered / reported
2. 417Write and run JavaScript/page suites.Encountered / reported
3. 418Test core context contracts and data transitions.Encountered / reported
4. 419Run focused browser navigation and interaction crawls.Encountered / reported
5. 420Test a real button click rather than just its render function.Documented decision / task
6. 421Discover many passing tests missed an obvious founder-facing failure.Encountered / reported
7. 422Separate mock-provider pass evidence from live-provider acceptance.Documented decision / task
8. 423Check that requirements are tested against the accepted founder intent.Documented decision / task
9. 424Create negative tests that deliberately make a gate fail.Documented decision / task
10. 425Find that a design-word budget was effectively hard-coded to PASS.Encountered / reported
11. 426Ensure the test actually goes red for a bad screen.Documented decision / task
12. 427Run relevant security, accessibility, and error-state checks.Documented decision / task
13. 428Test that previously working behavior survives new changes.Documented decision / task
14. 429Compare actual build output to the approved design and work package.Documented decision / task
15. 430Track which test version, source SHA, and environment were checked.Documented decision / task
16. 431Avoid full expensive QA for disposable exploration when focused checks suffice.Documented decision / task
17. 432Escalate test depth for shared infrastructure and release candidates.Documented decision / task
18. 433Check wrong-account, wrong-tenant, and revoked-access denial paths.Documented decision / task
19. 434Test how the product behaves if its AI provider is unavailable.Documented decision / task
20. 435Treat skipped tests as skipped, not passed.Documented decision / task
21. 436Prevent a builder from declaring itself independently reviewed.Documented decision / task
22. 437Use real-world user journeys instead of only line coverage.Documented decision / task
23. 438Repeat reviews after fixes and discover new defects in a different path.Encountered / reported
24. 439Stop endless repair when the same area fails repeatedly.Documented decision / task

1. 440Run independent security checks rather than relying entirely on builder self-review.Encountered / reported
2. 441Evaluate Aikido after substantial earlier hardening work.Encountered / reported
3. 442Find meaningful additional security issues in older or overlooked code.Encountered / reported
4. 443Distinguish source findings from reachable production vulnerabilities.Documented decision / task
5. 444Review dependency vulnerabilities against actual use and exposure.Documented decision / task
6. 445Scan for secrets and accidental log disclosure.Documented decision / task
7. 446Review static-analysis rules and their commercial licenses.Documented decision / task
8. 447Review cross-tenant access and account authorization paths.Documented decision / task
9. 448Verify production permissions, privileged routes, and release approval checks.Documented decision / task
10. 449Check webhook signatures and replay protections where supported.Documented decision / task
11. 450Recheck private-preview access, expiry and isolation before considering it safe.Documented decision / task
12. 451Test permission boundaries for browser workers and connected tools.Documented decision / task
13. 452Separate the risk a scanner asserts from the impact you’ve confirmed yourself.Documented decision / task
14. 453Document confirmed defects, mitigations, false positives, and unknowns.Documented decision / task
15. 454Re-run the relevant scanner after a bounded fix.Documented decision / task
16. 455Avoid treating an Aikido green result as legal or security certification.Documented decision / task
17. 456Decide when full DAST, pentesting, or human review is warranted.Documented decision / task
18. 457Protect production databases and real users from unsafe active scans.Documented decision / task
19. 458Record the exact source version scanned by each security tool.Documented decision / task
20. 459Check security findings against user-visible privacy and trust promises.Documented decision / task
21. 460Keep the baseline affordable rather than forcing every customer to buy premium pentesting.Documented decision / task

1. 461Create release candidates with named source SHAs.Encountered / reported
2. 462Prepare separate page, API runtime, and infrastructure releases.Encountered / reported
3. 463Keep staging and review environments distinct from production when needed.Encountered / reported
4. 464Encounter Railway deploy/start failures requiring investigation.Encountered / reported
5. 465Investigate a Vercel domain misconfiguration.Encountered / reported
6. 466Confirm correct DNS, environment bindings, and callback URLs.Documented decision / task
7. 467Check every required environment variable before release.Documented decision / task
8. 468Run Linux, Windows, frontend, and provider-specific gates as relevant.Documented decision / task
9. 469Repair a release script that failed to stop on a bad final check.Encountered / reported
10. 470Make a failed gate actually block the release unless explicitly governed.Documented decision / task
11. 471Bind the release packet to exact repository and deployed runtime identities.Documented decision / task
12. 472Keep one known-good rollback target and explicit recovery instructions.Documented decision / task
13. 473Stop new releases while another material build or migration is active.Documented decision / task
14. 474Retain a recoverable source and data backup before destructive changes.Documented decision / task
15. 475Check actual deployment identity rather than trusting pipeline labels.Documented decision / task
16. 476Run a real post-deployment smoke test on the published environment.Documented decision / task
17. 477Verify browser auth, API Chat, source connectivity, and preview after the release.Documented decision / task
18. 478Test rollback instructions rather than assuming they work.Documented decision / task
19. 479Discover a page can appear normal while backend readiness is still incomplete.Encountered / reported
20. 480Treat a green test suite and a green live release as different claims.Documented decision / task
21. 481Write clear founder release notes about what actually changed.Documented decision / task
22. 482Wait for explicit approval when the operation can affect customers.Documented decision / task

1. 483Set up application and infrastructure health signals.Documented decision / task
2. 484Review provider service failures and incomplete build status.Encountered / reported
3. 485Track unexpected 500 errors and browser console failures.Documented decision / task
4. 486Keep support diagnostics linked to the exact user task.Documented decision / task
5. 487Set sensible retry and backoff policies for temporary outages.Documented decision / task
6. 488Distinguish rate limits from revoked authorization and provider downtime.Documented decision / task
7. 489Monitor compute quotas, storage allowance, and inactive cloud resources.Documented decision / task
8. 490Discover hosted database compute allowances can be exhausted.Encountered / reported
9. 491Plan restore windows and recovery expectations for hosted data.Documented decision / task
10. 492Confirm secret rotation and credential expiry don't silently break customers.Documented decision / task
11. 493Watch for orphaned workers and previews that keep costing money.Documented decision / task
12. 494Document the owner and next action for a live incident.Documented decision / task
13. 495Decide which failures require immediate founder notification.Documented decision / task
14. 496Give the founder a useful error and recovery choice instead of a stack trace.Documented decision / task
15. 497Support repair without accidentally overwriting accepted source.Documented decision / task
16. 498Keep operational history so future agent sessions can diagnose repeated failures.Documented decision / task
17. 499Check license and provider-term changes for material workflow impact.Documented decision / task
18. 500Don't mistake an idle dashboard for a healthy working product.Documented decision / task

1. 501Compare monthly subscriptions with API-equivalent token estimates.Encountered / reported
2. 502Reconstruct spending when the visible Claude bill didn't explain compute consumption.Encountered / reported
3. 503Compare actual Claude and Codex use rather than one tool in isolation.Encountered / reported
4. 504Separate direct AI costs from your own coordination and hosting costs.Documented decision / task
5. 505Estimate fixed infrastructure cost per active project.Documented decision / task
6. 506Track costs of failed builds, review loops, and unused sandboxes.Documented decision / task
7. 507Set quota and per-task budget ceilings before paid work begins.Documented decision / task
8. 508Decide when a third-party service saves more than a custom implementation.Documented decision / task
9. 509Evaluate whether a paid specialist adds unique, consequential findings.Documented decision / task
10. 510Avoid claiming hypothetical counterfactual savings as historical fact.Documented decision / task
11. 511Measure elapsed time and founder interventions per accepted milestone.Documented decision / task
12. 512Evaluate lower-cost open models on held-out real tasks.Documented decision / task
13. 513Recheck model price, rate limits, and terms before changing provider routing.Documented decision / task
14. 514Protect a customer's ability to use their already-paid supported account route.Documented decision / task
15. 515Explain what pays for Chat, Build, Review, and preview separately.Documented decision / task
16. 516Test the economic assumptions at 10, 100, and 1,000 active projects.Documented decision / task
17. 517Decide when an additional service or human review is optional rather than bundled.Documented decision / task

1. 518Decide that founders should bring and own their preferred AI providers.Encountered / reported
2. 519Design a free first build that means holding a customer’s source code for a while.Encountered / reported
3. 520Disclose temporary source custody before claiming that code is in the founder's account.Documented decision / task
4. 521Verify GitHub App installation on the custody organization.Documented decision / task
5. 522Encounter a transfer-permission failure in the actual source-claim path.Encountered / reported
6. 523Check transfer authority before the founder attempts to claim code.Documented decision / task
7. 524Verify destination account, organization, permissions, and repository visibility.Documented decision / task
8. 525Confirm that full source history transferred to the new owner.Documented decision / task
9. 526Make sure you don’t keep an unauthorized copy of a customer’s source code.Documented decision / task
10. 527Delete trial copies of customer code according to the retention rules you disclosed.Documented decision / task
11. 528Avoid deleting the customer's claimed repository during cleanup.Documented decision / task
12. 529Preserve the user's ability to continue directly in Claude, Codex, or GitHub.Documented decision / task
13. 530Make customer data and accepted project decisions exportable where supported.Documented decision / task
14. 531Define what keeps running in the customer’s own accounts if they stop paying you.Documented decision / task
15. 532Show who owns and pays for every new resource.Documented decision / task
16. 533Prefer secure guided setup in the founder's account when it's simpler and cheaper.Documented decision / task
17. 534Reserve consequential spending, transfer, delete, and publish decisions for the owner.Documented decision / task
18. 535Confirm that leaving your product doesn’t mean rebuilding the whole project.Documented decision / task

1. 536Compare subscription pricing hypotheses for single and multiple projects.Encountered / reported
2. 537Separate your own fee from the third-party AI and hosting costs customers pay directly.Documented decision / task
3. 538Decide what a free or sponsored trial must actually prove.Documented decision / task
4. 539Show where the paywall sits before founders invest major effort.Documented decision / task
5. 540Avoid surprise token depletion, hidden credits, or bait-and-switch pricing.Documented decision / task
6. 541Plan support costs and unit economics for a small founder-led company.Documented decision / task
7. 542Research bootstrapping, angels, accelerators, and VC options.Encountered / reported
8. 543Explore Nashville-area founder networks and design-partner introductions.Encountered / reported
9. 544Discuss an investor one-pager and the evidence it would need.Encountered / reported
10. 545Decide which funding would genuinely accelerate development.Documented decision / task
11. 546Evaluate whether partner and model access matters more than headline valuation.Documented decision / task
12. 547Preserve product adoption and retention evidence before claiming traction.Documented decision / task
13. 548Separate a paid pilot, an invitation list, and active retained paying customers.Documented decision / task
14. 549Decide whether and when human engineering/security review becomes a premium offering.Documented decision / task
15. 550Keep company-readiness obligations proportionate to the founder's real goals.Documented decision / task
16. 551Distinguish a profitable small business from a venture-scale startup when advising founders.Documented decision / task

1. 552Try a positioning that sounds impressive to you, and find that it sounds generic to everyone else.My real build: “orchestrator.”Encountered / reported
2. 553Try positioning it as a founder operating system and find that people needed examples.Encountered / reported
3. 554Build a one-pager for a potential design-partner introduction.Encountered / reported
4. 555Explain what a founder gets that direct Claude use does not provide by default.Encountered / reported
5. 556Receive the question, 'Why wouldn't I just use Claude myself?'Encountered / reported
6. 557Develop a short product explanation for a first-time AI builder.Documented decision / task
7. 558Develop a deeper explanation for someone already paying for Claude and Codex.Documented decision / task
8. 559Show concrete responsibilities normally hidden behind a quick demo.Documented decision / task
9. 560Create an evidence-backed founder story rather than fictional customer testimonials.Documented decision / task
10. 561Decide how much technical complexity belongs behind an optional expandable layer.Documented decision / task
11. 562Keep brand tone witty without belittling inexperienced founders.Documented decision / task
12. 563Test whether the homepage actually communicates execution rather than advice only.Documented decision / task
13. 564Run synthetic messaging checks and recognize they aren't real buyer evidence.Encountered / reported
14. 565Verify every public feature claim against the live route actually supported.Documented decision / task
15. 566Design an onboarding moment that demonstrates useful judgment before asking for payment.Documented decision / task
16. 567Experiment with memorable campaign language instead of a wall of jargon.Documented decision / task
17. 568Measure real comprehension, activation, and willingness to pay.Documented decision / task

1. 569Seek design partners willing to test an unfinished product.Encountered / reported
2. 570Explore Vanderbilt and other founder-community introductions.Encountered / reported
3. 571Choose a small supervised beta cohort the infrastructure can safely support.Documented decision / task
4. 572Ask whether the user has a prototype, an idea, or an internal operational problem.Documented decision / task
5. 573Get permission before accessing customer repos and external accounts.Documented decision / task
6. 574Explain known limitations before asking for trust.Documented decision / task
7. 575Observe the first session without coaching the customer through every screen.Documented decision / task
8. 576Record the first moment the founder gets genuine useful value.Documented decision / task
9. 577Separate customer-observed friction from the builder's self-reported success.Documented decision / task
10. 578Watch which instructions the customer needs to repeat.Documented decision / task
11. 579Find whether initial recommendations match their actual goal and evidence.Documented decision / task
12. 580Test whether a founder returns after seven and fourteen days.Documented decision / task
13. 581Verify what users would pay for, and what they’d keep doing without your product.Documented decision / task
14. 582Ask what was unnecessary rather than only what features they want added.Documented decision / task
15. 583Track one credible real result before multiplying onboarding campaigns.Documented decision / task

1. 584Send Claude a task and review its confident completion report.Encountered / reported
2. 585Open the actual app and find the interaction still wrong.Encountered / reported
3. 586Write a correction that explains what the founder originally wanted.Encountered / reported
4. 587Send the correction back to the builder.Encountered / reported
5. 588Ask an independent model to review the new change.Encountered / reported
6. 589Find a different error in the review.Encountered / reported
7. 590Send that finding back to the builder.Encountered / reported
8. 591Wait for the next fix.Encountered / reported
9. 592Run the test suite again.Encountered / reported
10. 593Find a test that was checking the wrong thing.Encountered / reported
11. 594Rewrite the test so the obvious failure actually fails.Encountered / reported
12. 595Re-run the suite after the new test.Encountered / reported
13. 596Check that another screen didn't lose the original state.Encountered / reported
14. 597Discover the reviewer inspected an old branch.Encountered / reported
15. 598Explain which branch was actually deployed.Encountered / reported
16. 599Move the review to the current source SHA.Encountered / reported
17. 600Check whether another agent already fixed the same issue somewhere else.Encountered / reported
18. 601Reconcile overlapping worktrees and conflicting file edits.Encountered / reported
19. 602Retell the accepted decision to a new Claude session.Encountered / reported
20. 603Copy the latest handoff back into Notion.Encountered / reported
21. 604Correct an outdated Notion checkpoint that is still being read.Encountered / reported

1. 605Review an API-billing surprise or spent quota before continuing.Encountered / reported
2. 606Reduce agent parallelism because the coordination became more expensive than the work.Encountered / reported
3. 607Change the approach after one more repair fails the exact same way.Encountered / reported
4. 608Checkpoint what really worked so the next model doesn't repeat the whole cycle.Documented decision / task
5. 609Stop calling a task done because one agent wrote the word 'DONE'.Documented decision / task

A deeper pass surfaced additional real episodes and more specific decision checks. Many expand a general item above. Research labels distinguish reported experience from documented work.

1. 610Register a real praxeri.com business email workspace; the domain alone is not an inbox.Encountered / reported
2. 611Configure the Google Workspace Admin Console, account aliases, and team email identities.Encountered / reported
3. 612Set up two-step verification for the business email accounts and administrators.Encountered / reported
4. 613Set recovery phone and backup email so losing one login does not strand the company.Encountered / reported
5. 614Track the Google Workspace free-trial end date and the transition to actual billing.Encountered / reported
6. 615Connect Resend separately from Workspace and distinguish platform email from customers’ app email.Encountered / reported

Documented lessons from Eruvi, Stocks, Collectibles, and Core Intelligence. Not a universal requirement to operate four products.

1. 666Run a separate career-intelligence product alongside the AI-development platform.Encountered / reported
2. 667Decide whether that product is a one-time advice tool or an ongoing relationship with the user.Documented decision / task
3. 668Distinguish what a person said, what they demonstrated, what external evidence showed, and what an AI inferred.Documented decision / task
4. 669Keep several plausible career routes alive rather than forcing one impressive-sounding answer.Documented decision / task
5. 670Check career-market data source, release vintage, permissions, and staleness before making a recommendation.Documented decision / task
6. 671Separate a job's advertised pay from whether the person can realistically access that opportunity.Documented decision / task