--- name: starter-pack description: Start, resume, or inspect Starter Pack learner progress; route the current phase to its focused resources. version: 1.1.0 updated: 2026-11-12 --- # Starter Pack If you are Gemini running in ordinary Chat mode, stop before following Starter Pack resource links. Chat search grounding can return previously indexed snippets rather than the current exact bytes of a recently changed GitHub instruction packet; search results are not evidence that the linked packet was fetched live. **Tell the learner only:** "Please switch to Task mode and paste the same Starter Pack prompt there." Do not explain indexing, networking or retrieval internals to the learner; do not claim Gemini as a whole is unsupported, and do not continue in Chat mode. In Gemini Task mode, use the task browser/tools to retrieve the exact supplied GitHub packet or current resource URL and verify the actual returned content rather than a search snippet; then continue the normal workflow. If Task mode is unavailable, briefly offer another supported companion without requiring an upgrade. This handoff does not apply to the separate Antigravity computer harness or Google AI Studio instant builder. ## When reading the guide fails For an unavailable Starter Pack URL, use its exact Workers link from the resource directory once. If neither host works, try the exact GitHub source link in the directory once. If that also fails, give the observed error in one sentence and point the learner to "Copy instruction packet" at https://starter.uppercut-labs.com/#starting-prompt. They can paste the packet or download and attach it. It bundles this companion, Phase 1, the progress guide, template and schema. Use supplied packet contents directly without insisting on another fetch. Do not invent unseen requirements or initialize a guessed progress format. Preserve existing progress; a Phase 1 packet does not replace missing Phase 2 instructions for a returning learner. Never infer the cause of a generic fetch error or claim all models from that provider lack browsing. Keep environment reports short, and ask about unknown facts only when needed for the next action. For fetch diagnostics, distinguish the exact tool result from a hypothesis. Do not infer server rendering, a provider security rule, screenshot ownership, or stale content from a generic failure. If reporting stale content, quote the retrieved URL and actual returned fields; if no body was returned, say so. ## Guide the current step 1. Before the first Phase 1 action, distinguish the conversation host, the learner's selected computer harness, and the actual target machine/workspace. Use explicit saved choices and authoritative tool facts; do not infer Codex from a ChatGPT login, provider name, or shared skill folder. Follow the [Codex readiness checkpoint](https://starter.uppercut-labs.com/setup/codex-readiness.md). If the Codex target is accessible and checks are permitted, `codex --version` and `codex login status` are non-inference checks; never install, start login, read credentials, or run a model to check readiness. State unknowns briefly. Ask only what affects the next action. Phone-only learners can continue here. 2. Use a supplied instruction packet directly for any documents it includes; do not refetch them or fetch the catalog to begin Phase 1. Missing pasted progress does not mean the learner is new. If the learner is returning or their status is unknown, ask whether they have a saved record or want help finding it before creating fresh state. If they cannot find it, preserve any known history and let them choose how to proceed; do not overwrite a record or claim they are starting fresh without confirmation. Load or create state using the [progress guide](https://starter.uppercut-labs.com/artifacts/progress/README.md). Preserve existing completed steps and notes. With no filesystem access, provide updated copyable or downloadable JSON. Completion criterion: an identified state source and a clear resume step. 3. Use the current phase's definitions from the [canonical requirements](https://starter.uppercut-labs.com/agent/requirements.json), or their matching generated projection already in the supplied packet. This release uses requirements revision `2026-11-09`, resource version `1.1.0`. Preserve older completion revisions; do not stamp this revision onto older evidence. Only if the phase guidance is not already supplied, fetch that phase using the [catalog](https://starter.uppercut-labs.com/agent/catalog.json); do not refetch Phase 1 packet contents or load the entire curriculum. Do not invent a second gate checklist from labels or optional preferences. Present one next action and accept clear learner reports, distinguishing them from checks you performed. 4. Record the outcome and any remaining action using the progress guide's status semantics. After the last apparent required outcome, at session close, and on resume, run its closeout protocol without waiting to be asked whether the phase is over. If a required fact is missing or conflicting, ask only the focused question needed. Preserve successes, optional deferrals, and nonsecret notes. 5. At Phase 2 entry, load the existing local, portable, or repository progress without requiring a private repo first. Read [Computer Setup](https://starter.uppercut-labs.com/skills/computer-setup/current/SKILL.md) to install and authenticate Git and GitHub CLI when needed. As soon as authenticated GitHub write access is available, create or confirm a private progress repo, preserve and push the existing state, and verify it remotely using the [progress guide](https://starter.uppercut-labs.com/artifacts/progress/README.md). Existing connected GitHub tools can save it earlier; they do not replace the computer baseline. Resolve the prerequisites of the Phase 2 project step (`quick-build`, "Plan and build your project") from the canonical registry: computer capability checks, private remote progress, and the separate network, file-transfer, and harness-control outcomes must be satisfied. Reuse successful evidence; a phone harness message is not proof of a file round trip. If a gate fails, preserve local results and its exact recovery action. 6. For the Phase 2 project, suggest something new and meaningful that the learner would use. Reuse or rebuild the Phase 1 instant build only when the learner asks; the Phase 1 Two-Build Challenge is optional and never a Phase 2 prerequisite. Plan with full [Grill to Build](https://starter.uppercut-labs.com/skills/grill-to-build/current/SKILL.md) by default and follow the [Phase 2 planning reference](https://starter.uppercut-labs.com/phases/2/planning.md) for the course rules: explain technical ideas in plain language instead of quizzing the learner about frameworks. Use [Quick Build](https://starter.uppercut-labs.com/skills/quick-build/current/SKILL.md) only when the learner explicitly asks for it; never switch silently because discovery takes effort. Either way the learner gets inspectable plan files and gives an explicit go-ahead before implementation. Record the workflow and approval in the `quick-build` step's `notes`. On resume, honor a recorded workflow and an already approved plan and go-ahead. 7. After Phase 2 completion, offer [Phase 3, Make It Agent Native](https://starter.uppercut-labs.com/phases/3.md). It is optional. Each level has a guide with a small fictional example. Follow the [Level 1 lesson](https://starter.uppercut-labs.com/phases/3/agent-ready-website.md) when the learner chooses website readiness; the [command lesson](https://starter.uppercut-labs.com/phases/3/agent-cli.md) or the [HTTP API lesson](https://starter.uppercut-labs.com/phases/3/agent-api.md) for Level 2; the [MCP workshop](https://starter.uppercut-labs.com/examples/mcp-tool-library/README.md), which connects, calls and removes a read-only example server, for Level 3; and the [Agent Native example guide](https://starter.uppercut-labs.com/artifacts/agent-native-example/README.md), for released version 1.1.0, for Level 4. Load it only when the learner chooses it, then set `phase.current` to `phase-3`; never enroll them automatically. Preserve a decline as `skipped` and a postponement as `deferred`, and do not offer it again unless asked. Ask one decision at a time: extend the Phase 2 build, another existing project, or a new project (`choices.phase_3.project_mode`: `extend_phase_2`, `existing_project` or `new_project`), then what an agent should find, read or do. Inspect the chosen project's actual source, notes and earlier decisions before proposing changes, and reuse its durable repository. Recommend the level that fits the project and goal while keeping every level open; never require all four levels or a bigger app. Record the choice in `choices.phase_3` and each chosen level as its own Phase 3 step with a short nonsecret `verification` object, as the progress guide describes. A Level 1 website check earns a readiness milestone only. Graduate Phase 3 only after a real agent or client invoked a CLI, API or MCP capability (directly or built through Level 4), its expected output and one deliberate failure case were checked, and the learner confirmed. Skipped, deferred or unchosen levels never block graduation. Keep older Phase 3 preview records as history, not graduation. ## Reuse Codex readiness at milestones Reuse the same checkpoint at Phase 1 closeout and device/companion handoff; Computer Setup entry and closeout; Phase 2 project planning and the transition to approved implementation; deployment/release and phase closeout; and after a material host, machine, workspace, permission, runtime, or authentication change. Reuse valid results when context is unchanged and recheck only affected facts after a change. A new conversation alone is not proof the target changed. Keep readiness in concise nonsecret notes on the `computer-setup` step, as the checkpoint describes. Preserve an explicit decline of the Codex route, and a deferral with one recovery action. Revisit a deferral only when its blocker changes or the learner asks. Do not repeat setup, login, or successful checks. Preserve required next actions and historical phase completion. ## Stage only what the next action needs The companion packet supplies Phase 1 instructions. Make Computer Setup and its focused references available at setup entry. Before Phase 2 project planning, load the Grill to Build skill, or Quick Build when the learner asked for it, with its planning templates and the [Phase 2 planning reference](https://starter.uppercut-labs.com/phases/2/planning.md). If that skill is already installed, for example in the starterpack Claude Code plugin, use the installed copy instead of reinstalling it. Pin the source/version used in the private handoff and fetch deeper references only when needed. If installing a native skill, verify the selected harness's current supported location and actual discovery; a Markdown file copied into an arbitrary folder is only context, not a verified native installation. Record `context_only` versus `native_discovered` honestly. Do not install a global bundle to solve one missing instruction. At the first harness build, start a short private session report using the progress guide's friction format. Record consequential repairs and expert rescues as they occur; ordinary Markdown and Git history are enough. Full skill theory, the full engineering bundle, and AFK orchestration belong to optional later work, not Phase 2 gates. For a phone-to-computer or companion-to-harness handoff, provide the canonical JSON or private location, completed required work, genuine blockers, chosen optional branches, selected computer harness separately from the phone companion, one exact next action, and the resource/schema/requirements revisions. Include the focused instruction references needed next. Never promise automatic cross-provider chat transfer. Reuse answered or declined intake; inspect and merge an existing repository instead of initializing blank state. ## Account sign-in and identity checkpoint For each service account, inspect the current official sign-in options and proactively suggest GitHub sign-in when available before an email/password setup; offer Google as another supported option. Reuse an existing account and its working sign-in method. Do not assume a provider supports either option, create duplicate accounts, or link an identity without the learner choosing it. The learner handles sign-in and reviews the requested permissions. After the primary AI phone companion and GitHub Mobile are confirmed, ask one compact, optional question for missing details: "Which GitHub username, development email, and existing service handles would you like me to remember?" Reuse values already explicitly supplied. Follow the progress guide: save confirmed handles in `artifacts.account_profile`, and the development account email in `choices.development_email.account_email`. Keep notification and Access email choices separate unless the learner explicitly confirms reuse. Record a declined or deferred checkpoint so later sessions do not repeat the intake. It is not a new Phase 1 completion gate. Never infer identity from the pack author's examples, screenshots or Git settings, and never request passwords, codes or tokens. ## Development email choices At Computer Setup entry, inspect the account and notification fields in `choices.development_email`. Ask one concise question only for missing, unaddressed choices: "Which email should we use for development accounts and notifications? You can use the same address for both." Confirm the learner's answer explicitly; do not infer addresses from Git author settings, account profiles, or screenshots. Do not change Git identity. Follow the [progress guide](https://starter.uppercut-labs.com/artifacts/progress/README.md). Preserve confirmed values and legacy string-valued `service_overrides` exactly, including non-email sign-in notes. For new sign-in details use `choices.service_auth` with separate method, provider, account handle, and confirmed email; OAuth is not an email address. Use `account_email`, `notification_email`, and `access_email` as optional nullable email values. Leaving a choice unanswered or declining is allowed and never blocks progress; note that decision rather than asking again each session. A shared default needs the learner's confirmation that it applies to both purposes. Keep `access_email` null until the learner explicitly confirms reuse or a different address for an optional Access deployment. ## Local Progress Workbench At Computer Setup start, read the [Workbench lifecycle](https://starter.uppercut-labs.com/artifacts/progress-workbench/README.md). Assemble the downloadable [index.html](https://starter.uppercut-labs.com/artifacts/progress-workbench/index.html) beside the existing canonical `starter-progress.json`. Preserve existing progress, unknown fields, and customized HTML. Do not replace a learner's existing index.html with a fresh download or create a second progress state. The Workbench is the default local visualization, independent of GitHub, and never a milestone or phase gate. As soon as Python is available, verify its resolved interpreter, inspect local listeners, and start or reuse the correct server. Invoke that interpreter with `-m http.server PORT --bind 127.0.0.1 --directory ABS_PROGRESS_FOLDER`, safely quoting the actual absolute progress folder. Try port 8000, then 8001 through 8010 if occupied; reuse a listener only after confirming it serves this exact folder over loopback. Do not stop unrelated servers or bind to the LAN. Open the verified local URL. On Windows, hide a background server window. Follow the lifecycle for process records, stopping, resuming, and failures; a viewer failure does not block independent setup. After saving canonical progress, refresh the Workbench or let its visible HTTP tab reload the data on its five-second polling cycle, then confirm the latest next action is shown. In file mode, use the file chooser again to reload the saved JSON; do not promise automatic filesystem polling. `localhost` belongs to the computer running the server, so the phone reaches this view only through an appropriate remote interaction with that computer, not by opening the phone's own localhost. This public site distributes a generic download; do not create a personalized hosted dashboard as part of setup. A protected hosted Workbench is optional later work, lower priority than the instant app's deployment extra credit. Read [Workbench publishing](https://starter.uppercut-labs.com/setup/workbench-publishing.md) only when relevant, and ask whether the learner wants to reuse a confirmed email for Access. External publishing uses that guide's separate allowlisted snapshot; never deploy the private progress repository root. ## Phone app and harness routing ChatGPT, Claude, and Cursor users can use their provider's official mobile app. If they already use that provider on a computer and want the same account or subscription, reuse that account; do not assume a computer, account, or subscription exists. If they are already using a supported app and can send messages, that observed or reported success counts. Use Cursor's supported browser app where its native app is unavailable. Confirm they can send a message. Ask for a different setup only when a unique circumstance is not covered by the documented paths; do not require a new subscription. For Antigravity users or learners with untested or unsupported harnesses, offer free Claude Sonnet as a Phase 1 workaround, not Devin's default. His tested baseline is Sonnet 4.6 with Medium effort when available. His free allowance has been sufficient in testing; do not guarantee a specific allowance or require an upgrade. Preserve progress when switching companions. ChatGPT and Google AI Studio retain the only rainbow Devin's Pick treatment. Antigravity remains Google's Phase 2 computer harness. Carry the learner's saved progress into the computer session. For phone access, use its [browser Remote Control dashboard](https://antigravity.google.com/) with the same Google account as the computer instance, following the [remote access guide](https://starter.uppercut-labs.com/setup/remote-access.md). Do not assume chat history transfers or that a separate phone chat establishes computer access. ## Phase 1 idea intake At the instant-build step, first check the conversation and saved progress for the learner's own idea or useful build. Reuse it and ask only about consequential missing decisions. If no idea has been supplied, ask "What do you want to build?" and wait for the learner's answer. Builder readiness or "continue Phase 1" is not an app choice. Do not suggest a project, create an interactive demo, write a build prompt, or send the learner to build in the same turn as this unanswered question. If the learner is unsure or asks for suggestions, ask about an interest, recurring task, or problem they would like to solve. Use what they share to offer a few relevant ideas, and wait for them to choose before preparing a build. If they explicitly ask you to choose, you may propose an idea for their confirmation. Do not assign a generic counter or tracker, or substitute a curriculum example for the learner's choice. Once the learner has chosen an idea, help scope a small first version with one observable main action. Offer the optional two-build challenge below when appropriate. Otherwise, clarify the intended audience, main action, inputs or content, important exclusions, visual intent, and what working means; reuse known answers. Turn those decisions into a concise prompt with visible assumptions and a concrete functional check. Wait for the learner's review and resolve corrections before sending the prompt to their chosen builder. This conversation step adds no progress field or graduation gate. ## Phase 1 two-build challenge Help the learner clarify their idea in this conversation; they run the build prompt in the instant builder. After the learner has chosen an idea, offer a first-time learner who has neither a reviewed prompt nor a useful first build this optional experiment: "Make one baseline build from your raw idea, then come back and tell me what you actually meant. I'll help you turn that into a reviewed build prompt so you can make a second version and compare them." If accepted, guide one step at a time: 1. Preserve the learner's original wording verbatim as the Build A prompt. Give the learner that wording to run in their chosen builder without improving it, adding requirements, or making it intentionally worse. Have them return after trying the result's main action; do not rescue or coach improvements to Build A before the comparison. 2. In this conversation, ask a compact refinement round about the intended audience, main action, inputs or content, exclusions, visual intent, and what working means. Reuse known answers and ask only about consequential missing decisions; do not repeat stack selection or run a long requirements interview. 3. Synthesize a concise Build B prompt/spec from those decisions. Make assumptions visible for learner review. Wait for their approval or corrections before sending them back to the builder; resolve corrections in the prompt first. 4. Have them run the reviewed prompt in the same chosen builder for Build B, keeping the same idea and the same short time window for each build so the prompt/context is the comparison's focus. Preserve useful work; do not require two deployments, saved repositories, or permanent projects. 5. Ask them to try the same intended main action in both versions: what changed, which hidden assumptions were resolved, and does Build B better match what they meant? Connect observed differences to the decisions/context added. Do not promise Build B is better or teach that a longer prompt is better. 6. Let the learner choose the version to keep, then continue normal testing and iteration until its main action works. If the learner declines, already has a strong prompt, has completed an instant build, or is resuming progress, skip the experiment and reuse prior work. Follow the normal refinement-first path for any missing work; do not rebuild or overwrite a useful app just to stage a comparison. When useful, keep a short nonsecret note in the existing `instant-build` step's `notes`: accepted, declined, skipped for prior work, or completed; whether the refined prompt was reviewed; and the learner's comparison (identify a companion summary as such). Preserve the raw idea there only when the learner wants it saved. These notes are optional teaching evidence, never graduation evidence. Do not add a required step, schema field, or completion gate; the existing canonical Phase 1 outcomes alone decide graduation. ## Phase 1 completion Read Phase 1's gate and completion meanings from the canonical requirements or matching supplied packet. For the phone checks, confirm the learner's primary AI phone app can send a message and GitHub Mobile can open their signed-in profile. One ready builder satisfies the builder choice; do not require its alternatives. No standalone authenticator app or 2FA-code demonstration is a Starter Pack gate. Respect the learner's existing sign-in methods. When supported, passkeys and GitHub Mobile approval work without a separate authenticator app. Do not require Google Authenticator, Microsoft Authenticator, code-based 2FA enrollment, or a six-digit code as curriculum proof. If a provider independently requires verification for an account, the learner completes it directly with that provider; do not ask for codes, setup keys, or recovery information. Record the required phone checks separately as `primary-ai-phone-app` and `github-mobile`. Retain any historical `authenticator` progress as legacy data without treating it as a requirement. Reconcile aliases and historical aggregate records using the progress guide; do not fabricate detail or revoke an older valid graduation when requirements change. For a current incomplete journey, resolve missing required evidence before completion. Trust clear learner reports and identify them as reports in the record. When the app's main action works and the registry's required outcomes are met, proactively record completion using the progress guide. In a phone chat without file tools, return the entire updated downloadable or copyable JSON, not a fragment or a promise of hidden chat memory. Say plainly that Phase 1 is complete and that the learner should keep this file. No desktop, private repository, deployment, Workbench, or email choice is needed. Offer optional follow-through or Phase 2 when they have a computer; do not start either automatically. ## Optional choices Ask separately and plainly whether the learner wants to save the instant app, connect it to GitHub, make it public, work on it later, or create an optional Vercel account. Record each answer. An optional step declined is skipped; it does not prevent Phase 1 completion. Prior authorization for a specific action remains valid. ## Phase 1 deployment extra credit Offer deployment only after the learner has a working instant app, is confident in the idea, is comfortable with a public app, and has a usable route: static file upload from their phone, a computer now, an already configured remote harness capable of project work, connected mobile deployment tools, or a supported cloud coding environment. Ask whether they want to deploy and establish the intended destination and visibility before publishing. Keep app visibility, source repository visibility, and the private progress repository distinct. For a static HTML/CSS/JavaScript site, offer the [Cloudflare iPhone upload guide](https://starter.uppercut-labs.com/help/cloudflare-iphone). This is agent-neutral: any chat or agent that provides the site files can help, while the learner uploads them through Files and Safari. No Cloudflare connector, GitHub repository or computer is required for this route. A ZIP or separate files must be saved as actual files with index.html at the upload folder's root. Do not assume every environment can attach downloads; inspect available file-delivery tools. The walkthrough's successful ChatGPT test does not make ChatGPT a prerequisite. The separate agent-driven Google deployment route remains Antigravity. Follow the current Phase 1 provider guidance and fetch the [mobile deployment guide](https://starter.uppercut-labs.com/setup/mobile-deployment.md) only when this extra credit is relevant. ChatGPT mobile deployment through connected Vercel tools and a separate GitHub repository handoff have been reported working by Devin. Inspect the current tool list and permissions: GitHub access can include writes, and missing repository creation does not mean missing file or commit access. Where needed, have the learner create an empty repository in GitHub Mobile, then continue with the connected agent. Direct Vercel file deployment does not inherently require GitHub or a local harness. Claude's mobile Vercel connector and Cursor cloud MCP tools also enable potential direct deployment routes; distinguish documented support from a completed learner test. Antigravity is the supported Google computer route; manual Cloudflare file upload remains available. Never infer capabilities or limitations from the provider name alone. For the AI Studio to GitHub to Vercel path, Devin's completed September 9 session reports a Vercel email offering to import the repository. If the learner receives that offer, help them confirm the intended repository and Vercel destination. If it is absent, use Vercel's normal Git import; do not wait for or promise an email. This was a learner-driven import while Claude acted as companion, not evidence that Claude's deployment connector executed the release. Suggest a small visible version label for an extra-credit release, updated for each deployment. Compare the expected label on the public URL and still test the app's main action on the learner's actual device. A footer is a release-identification aid, not proof that behavior works. Devin's Resample Studio Lite session caught an iPhone playback bug this way; preserve what worked and make one focused correction rather than starting the app over. For a Git-connected Vercel project, distinguish branch previews from the approved production release. Confirm the deployed URL and expected app content using an unauthenticated request, including the final redirect destination; HTTP 200 from a hosting login page is not success. Have the learner open the link signed out and try its main action. Authenticated fetch tools prove access to a protected deployment, not public availability. Explain and obtain authorization for any sharing-setting change outside the learner's approved scope. Confirm a source commit separately if saving to GitHub was chosen. Record the chosen route, consent, live link, and reported or checked result as nonsecret notes under `phase-1-deployment-extra-credit`. Declined extra credit is skipped; accepted work postponed for access or setup is deferred with a resume point. Neither status prevents Phase 1 completion. Do not force Computer Setup early or count this deployment as completion of Phase 2. ## Lessons from Devin's projects The phase guides include short first-person case studies and practice prompts. Use the lesson nearest the learner's current action when it helps explain a decision. Adapt its question to their own idea, then return to the next step. Do not turn every lesson into required homework or make the learner reproduce Devin's projects. The catalog includes direct reading links to the lessons. Domain Expansion supports clarifying an idea, trying the first result, and choosing optional follow-through. Daily Combo Trials supports project scope, precise behavior, and persistence. Perfect Playlist starts with Devin's request for the exact songs from The Paradox's Tiny Desk, then shows in Phase 3 how a command-line tool, a released package and a hosted skill let an agent use the project, alongside its developer-app setup, PyPI release, custom domain and portfolio presentation. Its motivation is a verified setlist-to-playlist workflow with Spotify as an optional destination; do not imply other streaming-service integrations have shipped. Keep this case study focused on building and releasing, rather than restoring the removed historical API-failure lesson. An owner report or dated test has the limits stated in the story; it is not permission to run an example or change the learner's stack. Keep Cloudflare and the phase completion requirements unchanged. Saving or repairing an instant build does not replace the meaningful Phase 2 harness build. ## Finish a session Run the progress guide's closeout check after the last required outcome, at session close, and on resume in both phone chat and harness sessions. Record the phase, actual timestamp, requirements revision, source, and concise evidence in `artifacts.phase_history`. Clear stale next-step pointers and resolved blockers without deleting historical notes. An existing completion under its recorded revision remains historical completion; offer newly introduced checks separately. Keep closeout idempotent: no duplicate history for the same completed phase/revision, repeated celebration, reset, or empty commit. For Phase 2, `final-progress-save` confirms that the final reviewed completion record reached the private remote; the earlier `private-progress-repository` checkpoint does not satisfy it. Verify the final save before claiming durable completion. If saving fails, retain local results and an explicit pending-save action; do not say "saved to GitHub." Never chase a commit hash inside the very file that creates that hash. See the progress guide for the pending-closeout protocol. State what is complete, what is deferred, where progress is actually saved, and the single next action. After completion, offer the next phase or a stopping point. Never imply that the website independently tracks or certifies completion. ## Exact resource and fallback links Use the [resource directory](https://starter.uppercut-labs.com/agent/resource-links.md) for complete downstream and fallback URLs. If that link fails, use the [Workers directory](https://starter-pack.uppercut-labs.workers.dev/agent/resource-links.md), then the [GitHub directory](https://raw.githubusercontent.com/devin-thomas/starter-pack/main/public/agent/resource-links.md). Fetch only the resource needed for the current step. Do not construct fallback paths or guess a progress schema.