Start here

If you want proper results from AI,
you need to help it stay on track.

Nudge is a structured approach to working with AI — from casual users who want better outputs, to developers shipping production code, to teams building machine-executable agent pipelines.

73%
fewer tokens on agent actions
N=5
calls to break even
5
AI roles. One per stage.
The problem

The longer you talk to AI
without structure,
the worse it gets.

The problem isn't the AI. It's the lack of clear roles, scope, and constraints. Without them, the AI accumulates dead context and starts reasoning against its own noise.

WITHOUT STRUCTURE
You dump context and hope for the best
AI makes uninstructed architectural decisions
Dead history degrades every completion
Expensive rewrites and repeated explanations
WITH NUDGE
Each message has a clear role and scope
AI executes exactly what you specified
Context stays sharp — managed deliberately
Token efficiency means better reasoning
Your path

Where do you want
to go from here?

All three paths converge at the same place. Pick what fits you today — you can jump to any node from the sidebar at any time.

Path A · Step 1

Every task needs
five different roles.

Stop asking one AI to be everything at once. The moment you separate these five modes, output quality jumps — for writing, for code, for anything.

01
Challenger
High autonomy
Breaks the plan. Finds weak points. Challenges every assumption first.
02
Planner
Medium
Sequences work. Flags hidden dependencies before execution starts.
03
Specifier
Medium
Translates plans into precise specs with acceptance criteria.
04
Executor
Low
Executes against specs. No decisions. Stops if ambiguous.
05
Reviewer
Zero
Validates against spec only. Does not improve. Three verdicts.
Path A · Step 2

The one habit that
transforms AI output.

Fill in six fields. Copy. Paste before any AI request. Works on Claude, ChatGPT, Gemini — anything. No account required.

Fill → Copy → Paste before your next AI requestFree · Works anywhere
// Fill the fields above — your structured brief appears here
Path A · Step 3

How work actually gets done.

You don't build everything at once. You pass the work through these 5 stages, starting a new chat when needed.

1. Discovery Brief
Lock in the scope. Ask questions to figure out what you are NOT building.
2. Blueprint
The Master Plan. Map out the overall strategy.
3. Task List
The To-Do List. Break the Master Plan down into an ordered sequence.
4. Shard
The Task Brief. A highly specific brief for ONE single item from your To-Do list.
5. Deliverable & Review
The AI executes the Shard. You review it against the constraints.
Path A · Step 4

Don't lose the plot.

When you move to a new chat to clear the AI's memory, you must pass the baton. Follow this strict 6-step flow.

1
Check Pipeline
2
Set Profile
3
Work (1 task)
4
Log Output
5
Handover Summary
6
Index Update
The Handover RuleBefore ending a chat, ask the AI to write a short, bulleted summary of all decisions made. Paste this into the next chat.
Path A · Step 5

28 cards.
Start with the essentials.

Each card is one habit. Start with Essential. Add more when you feel the problem they solve. Tap any card to flip it and copy the prompt.

Path A · Step 6

When the AI fails.

AIs hallucinate. They get stuck in loops. Here is the non-technical protocol to recover without wasting tokens or ruining your work.

Stop
Classify the error
Is the brief unclear? Or is the AI carrying dead context? Don't just rephrase.
Two Attempts
The strict limit
If it fails twice to follow the brief, the brief is broken. Stop and rewrite it.
Undo
The Kill Switch
Never let the AI overwrite your only copy. Revert to the last good prompt immediately.
Path B · Step 1

Five stages.
Five roles. One pipeline.

Strict separation of thinking and doing. Human in the loop at every gate. Select a stage to see the full copy-able prompt.

Path B · Step 2

Classify before
you touch anything.

Debugging is investigation, not development. A dev agent never switches roles mid-session — if it hits a bug during feature work, it stops feature work. Full stop.

Environment
Fix the setup
Code is correct, environment is wrong. Wrong version, missing variable, port conflict.
Never touch code for an env bug
Logic
Fix the code
Environment is fine, implementation is wrong. Read the error fully. Minimal fix only.
Smallest fix that passes the test
Spec
Fix the spec
Code does what was asked, but what was asked was wrong. Stop. Rewrite the shard.
Stop and rewrite — do not patch
Critical ruleA dev agent never switches roles mid-session. Hit an env bug during feature work → stop, classify, report to orchestrator. Do not become a debugger.
Path B · Step 3

Give the AI a map,
not a maze.

AIs waste massive context running ls or reading entire files for one function. These six scripts eliminate that permanently. Copy once, use everywhere.

pre_session.ps1
Snapshots active phases, file existence, stubs before every session.
eliminates: ls, find, dir
update_structure.py
Maps every function to its exact line number in structure_map.txt.
eliminates: navigation reads
find_symbol.ps1
Returns a function + 2 lines context. Replaces 3000-token full-file reads.
eliminates: full-file reads
run_task.ps1
Prints OK on success. Writes 20 error lines to .last_error.txt on failure.
eliminates: verbose test output
get_skeleton
Tree-sitter powered. Outputs signatures without reading file content.
eliminates: structure reads
read_anchor
Fetches exact logic blocks via @ANCHOR comments. Zero-read dependency maps.
eliminates: context searches
Path C · Step 1

nudgeDSL —
the encoding layer.

A token-dense, human-readable Domain Specific Language for encoding executable intent from LLMs to any backend. One string. One parser. Go, Python, Dart — same AST.

Today — your .handover.md (prose, ~180 tokens)
slice: auth-engine-v1 status: complete files_modified: - auth.go / auth_test.go decisions_locked: - JWT not sessions - Refresh tokens in Redis next_slice_needs: - contracts.yaml
nudgeDSL (~38 tokens, executable)
MARK("auth-v1", "done") >> MOD("auth.go") // MOD("auth_test.go") >> NOTE("JWT not sessions") >> NOTE("Redis for refresh") >> NEXT("api-layer")
Path C · Step 2

Four operators.
Any intent.

The grammar is fixed. The vocabulary is yours. Four operators encode every relationship between actions — sequential, fallback, parallel, and repeat.

>>
Chain
Sequential. Right runs only if left succeeds.
MARK("done") >> NOTIFY("ops")
|
Fallback
Try left. Use right if left fails.
FETCH("primary") | FETCH("replica")
//
Parallel
Concurrent execution. Both run simultaneously.
WRITE("db") // CACHE("redis")
**N
Amplify
Repeat N times sequentially. N must be ≥ 1.
PING("health") ** 3
Precedence: **//>>|  ·  Use parentheses to override
Path C · Step 3

The grammar is fixed.
The vocabulary is yours.

An atom registry is a JSON file that defines your domain's vocabulary. Import it into the playground or use with the CLI. Three registries ship out of the box.

nudgedsl-core
Core
14 atoms · Generic
MARKCREATEMODNOTEFETCH+9
nudge-framework
Framework
27 atoms · Pipeline
SHARDACCEPTVERIFYRESULT+23
game-ai
Game AI
14 atoms · Tactical
SELMOVATKHEALCONF+9
Path C · Step 4

Type nudgeDSL.
See it parse.

The parser runs entirely in your browser. Validates syntax, returns the full JSON AST. No server, no account, no key needed for validation.

input nudgeDSL
ast / errors
// Type nudgeDSL above or pick an example
Almost there

Nudge in your browser.coming soon

The Brief Builder and role profiles as a sidebar on any AI chat. Works on Claude, ChatGPT, Gemini. One click switches the AI's role — no copy-paste, no tab switching.

AI Profiles
Switch roles instantly
Challenger, Planner, Specifier, Executor, Reviewer. One click — the AI changes mode immediately.
Click profile → AI switches mode
Brief Builder
Always one click away
Structured context in your sidebar. Fill in six fields. Copy. Perfect brief every time.
Fill → copy → paste before request
nudgeDSL widget
DSL in the browser
Validate DSL, load registries, generate agent prompts without leaving your AI chat. In discussion for v2.
In discussion for v2
Watch on GitHub
The upgrade layer

Everything you just learned
made machine-executable.

nudgeDSL sits on top of the entire framework. Your handovers, shards, agent outputs — all expressible as dense, human-readable, executable strings. The protocol is always free.

Free
No account · No key · Browser
Validate DSL syntax
Generate agent prompts
Import custom registry
Export GBNF grammar
Browse all registries
Prose → DSL translation
Token savings benchmark
Bring your key
Your Anthropic key · Browser only
Everything in Free
Prose → DSL translation
Token savings benchmark
Registry inference from code
Key stays in your browser
CLI pipeline integration
Batch processing
Local CLI
Self-install · Pi/MCP · Offline
Everything above
nudge analyze ./output.json
Batch entire projects
Editor integration
Pi / Jetson MCP wiring
Offline execution
nudgeDSL v0.1.0 · pre-release · Structure beats hope.