Big Jump — from no direction to a first real software project

If you are stuck at the first step

Know some Python, frontend, or data analysis, but have never independently finished a real project?
Want credible portfolio evidence or a first software-income experiment, but do not know which real problem to start with?
Have followed tutorials and completed coursework, but do not know how to find a first user, cut scope, or decide what “done” means?

Big Jump turns a coding agent into an autonomous builder-coach. It first helps choose a direction from people and problems you can reach, then scopes, builds, verifies, and teaches one useful concept per phase.

From knowing some technology to your own first release

Big Jump began from a gap familiar to many students at China's 985/211 universities: learning plenty without a path from knowledge to a real user; the school label is context, not an entry requirement. Big Jump does not promise that a finished website creates income. It keeps technical correctness, first use, repeated use, and willingness to pay as separate evidence.

YouI am a university student who knows some Python and frontend development but has only built coursework. I see other people earn from small websites and do not know where to start. Give me no more than three directions from problems around classmates, student groups, or a lab, and recommend one zero-budget default.

Big JumpWe will not pick a framework or deploy yet. I will identify your primary goal, people you can reach this week, and repeated behavior you can observe; then compare candidates across web, CLI, or data tools, recommend one default, and name the cheapest pre-build test. Once the direction has evidence, we can build the smallest version one real user can try.

How it works

Routes by productSelects web, backend/API, CLI/automation, data/AI, library/SDK, mobile/desktop, or a combination. Existing projects keep their native stack and conventions.

Verifies before it claimsMatches each claim to fresh static, logic, integration, interface, or release evidence. A unit test never stands in for an untested browser, package, model, or device flow.

Teaches after each moduleA 4-line recap per phase: what was done, why this way, what you learned, and what was verified. Sparse output, no progress cards, then it moves on automatically.

Stops only at real boundariesMissing authority, a secret or paid resource, sensitive production data, a difficult-to-reverse choice, or a blocker with no remaining safe diagnostic path.

Releases honestlyAsks before unapproved external actions, then verifies the actual delivery surface. A public URL, package, model artifact, or app build is only claimed when it was really checked.

Protects your workNever resets or overwrites your uncommitted changes, never force-pushes; reverts only its own edits.

Adapts to your setupRuns the build itself when it has a terminal; in a plain chat it hands you files and exact commands instead — and never claims to have run what it couldn't.

Finds your directionNo idea what to build? A problem-hunting method and a 3-question picker turn “I don't know” into a small project a first user can actually try.

See the workflow in action

YouBuild a macOS and Linux CLI that renames photos by EXIF date. Dry-run by default, handle collisions, and never test on my only copies.

Big JumpRoute: new build · CLI/automation · destructive-operation overlay. Outcome: preview a deterministic rename plan, then apply it only after explicit confirmation. Non-goals: editing photo contents or testing on originals.

[Slice 1: parse copy-only fixtures and print a stable dry-run plan.]

Evidence: unit cases for missing dates and name collisions, plus a real subprocess run in a temporary directory.

[Slice 2: add confirmed apply, idempotency, interruption recovery, packaging, and a clean-install smoke test.]

EndVerified on the platforms actually run: dry-run leaves files untouched; confirmed apply handles collisions; a second run makes no further changes; clean installation exposes the command. Any untested platform is named as a limitation, not implied as supported.

Quick commands

/pause stop the auto-flow and wait for me
/test run the current fast evidence set
/test full run the acceptance evidence set
/ideas find and select a product direction
/explain explain the current slice in plain language
/review inspect scope, correctness, maintainability, and evidence
/deploy follow release choices and permission gates
/retrospective summarize what I learned

Install

Codex — one command

Download the repository and run:

bash install.sh

Or copy the complete big-jump folder into ~/.codex/skills/. The references, assets, evals, and validation script are part of the skill; copying only SKILL.md is incomplete. The folder must be named big-jump.

DeepSeek Harness

Put the big-jump folder into .dsh/skills or .agents/skills in your project, or into ~/.dsh/skills.

Claude Code / Cursor

Put the folder into the tool's skills directory (for example ~/.claude/skills/, or the .agents/skills project folder).

Use it

In a new chat, describe a substantial outcome, such as "Build a safe photo-renaming CLI with dry-run and recovery." If the skill does not load automatically, say "Use the big-jump skill."