Zero-config. Auto-learns. Just works.
97% smaller context · Auto-learns your style · Works with any agent
Compresses AI agent context from 56K tokens to 1.9K tokens. Learns your coding patterns from git history and session logs. Results vary by project size and session history.
You know the problem. You start an AI agent session. It reads your entire project: session logs, git diffs, READMEs, config files. It writes code that doesn't match your style. It wastes tokens and produces generic, bloated code.
taste puts a stop to that. It learns your patterns. It compresses your context. It makes your agents write code like you do.
You ask for a rate limiter. Your agent reads 56K tokens of context, installs a library, writes a generic implementation, and asks about your Redis setup.
With taste:
taste
More examples in examples/.
Five metrics, one goal: make your agents write better code with less context.
| Metric | Without taste | With taste | Improvement |
|---|---|---|---|
| Token usage | 56,000 | 1,950 | 97% reduction |
| Context quality | Generic | Project-specific | Better |
| Pattern learning | Manual | Automatic | Zero-config |
| Agent support | Single | Multiple | 3+ agents |
| Setup time | Hours | Seconds | Instant |
97% smaller context, auto-learns your style, and works with any agent. Every pattern taste learns is marked in the code with confidence scores. Reproduce it yourself: run taste learn
in any project. Method and raw numbers: benchmarks/. Real-world examples: examples/.
That is the byproduct, not the pitch. These are average numbers, and they vary by project. Larger projects with more session history see better compression. Smaller projects with less history see smaller savings. And all of this is iterative: each time you run taste learn
, it learns more patterns, which makes the next compression better. The rule was never "fewest tokens." It is: learn only what the project needs, and never skip validation, error handling, security, or accessibility. The context ends up small because it is necessary, not trimmed, and that is the part that stays useful. Better code quality is a side effect of learning your style, and that is the part that matters.
Before compressing context, taste learns your patterns:
1. Collect session data → git diffs, session logs, taste config
2. Create summary → compact format, essential information only
3. Analyze with agent → calls opencode or claude
4. Extract patterns → naming, architecture, imports, error handling, style
5. Update taste config → append to TASTE.md, update .agent-taste.json
Lazy, not negligent: validation, error handling, security, and accessibility patterns are never skipped.
The most effort taste will ever ask of you:
curl -fsSL https://raw.githubusercontent.com/dvcoolarun/taste-ai/main/install.sh | bash
git clone https://github.com/dvcoolarun/taste-ai.git
cd taste-ai
chmod +x taste
cp taste ~/.local/bin/
taste help
That was it. It would be proud. It won't say it.
taste
taste learn
taste learn --depth 5
taste learn --dry-run
taste init
taste show
| Command | What it does |
|---|---|
taste |
|
Pack session context into .session-doc.md |
|
taste pack [file] |
|
| Pack to specific output file | |
taste init |
|
Create default .agent-taste.json |
|
taste show |
|
| Show current taste config | |
taste learn |
|
| Learn patterns from recent sessions (agent-assisted) | |
taste help |
|
| Show help |
taste learn analyzes your coding sessions and extracts patterns:
┌─────────────────────────────────────────────────────────┐
│ Data Collection │
│ - Last 3-5 session logs │
│ - Last 3-5 prompt logs │
│ - Git diffs (last 3-5 commits) │
│ - Current taste config │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ Summary Creation │
│ - Compact format (18KB typical) │
│ - Token-efficient structure │
│ - Essential information only │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ Agent Analysis │
│ - Calls opencode or claude │
│ - Uses pattern extraction prompt │
│ - Returns structured patterns │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ Pattern Extraction │
│ - NAMING conventions → TASTE.md │
│ - ARCHITECTURE patterns → TASTE.md │
│ - IMPORTS style → TASTE.md │
│ - ERROR_HANDLING patterns → TASTE.md │
│ - STYLE preferences → TASTE.md │
│ - BANNED_PATTERNS → .agent-taste.json │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ Auto-Update │
│ - Positive patterns → TASTE.md │
│ - Banned patterns → .agent-taste.json │
│ - Preserve existing patterns │
│ - Avoid duplicates │
└─────────────────────────────────────────────────────────┘
| Category | What It Learns | Output Location |
|---|---|---|
| NAMING | Function naming conventions (snake_case, camelCase, etc.) | TASTE.md |
| ARCHITECTURE | Project structure patterns, dependency management | TASTE.md |
| IMPORTS | Import style, ordering, lazy vs eager imports | TASTE.md |
| ERROR_HANDLING | Try/catch patterns, error propagation | TASTE.md |
| STYLE | Code formatting, function length, comments | TASTE.md |
| BANNED_PATTERNS | What NOT to do, with reasons | .agent-taste.json |
taste learns both positive patterns (what to do) and negative patterns (what NOT to do). Banned patterns are extracted from user corrections, past mistakes, and feedback.
Example banned patterns:
{
"banned_patterns": [
"--single-process_Chromium_flag_on_macOS (reason: causes crashes, documented failure)",
"hardcoding_connection_URLs_or_env_specific_values (reason: caused 'Queue service unavailable' failure)",
"jumping_to_implementation_before_design_alignment (reason: wasted work when pricing model wasn't confirmed)",
"removing_comments_during_code_rewrites (reason: user explicitly called out and expects preservation)",
"using_browser_only_Node_APIs_in_subprocess (reason: ErrorEvent caused ReferenceError)"
]
}
Why banned patterns matter:
Specific- Not generic ("don't use classes") but concrete ("don't use --single_process_Chrome_flag")** Actionable**- Clear reasons that explain WHY it's banned** Learned from mistakes**- "was replaced with page.setContent" shows historical context** Platform-aware**- "crashes on macOS" shows environment-specific knowledge
How banned patterns work:
- Agent extracts BANNED_PATTERNS from session data
- Patterns are written to
.agent-taste.json
as a JSON array - When you run
taste
, banned patterns are included in.session-doc.md
- Agents read the banned patterns and avoid those patterns
Each pattern includes a confidence score (0-1):
0.9-1.0: Very high confidence (seen multiple times)** 0.8-0.9**: High confidence (seen consistently)** 0.7-0.8**: Medium confidence (seen occasionally)** 0.6-0.7**: Low confidence (seen once or twice)<0.6: Not included (insufficient evidence)
Each time you run taste learn
, it:
Reads your latest session logs and git diffsAnalyzes patterns with an AI agentExtracts structured patterns with confidence scoresUpdates your taste config (TASTE.md, .agent-taste.json)Preserves existing patterns and avoids duplicates
The more you use taste, the better it learns your style. Pattern confidence increases as it sees the same patterns across multiple sessions.
taste
opencode .
taste
claude .
taste
Source: `.agent-taste.json`
``` json
{
"flavor": "Functional TypeScript, strict types, zero dependencies",
"banned_patterns": ["classes", "any", "console.log"],
"style": "Implicit returns, max 20 lines per function"
}
Branch: main
Last 5 commits:
abc1234 refactor: extract auth to /core
def5678 feat: add token validation
Changed files (last commit):
src/auth.ts | 12 +++---
src/utils.ts | 5 +++-
### `taste learn` Output
taste: Analyzing last 3 sessions... taste: Collecting session data... taste: Summary created: 18818 bytes taste: Calling opencode for analysis...
LEARNED PATTERNS (last 3 sessions):
NAMING:
- functions_describe_action_verbs (confidence: 0.9)
- classes_use_PascalCase_Prefixed (confidence: 0.9)
- variables_underscore_separated_snake_case (confidence: 0.8)
ARCHITECTURE:
- Python_FastAPI_fronts_with_Node_subprocess_backend_via_stdin_stdout_bridge
- async_job_queue_with_redis_backend_and_RQ_worker
- dual_storage_PDF_disk_and_Redis_cache
IMPORTS:
- lazy_import_inside_endpoint_to_avoid_side_effects
- from_stdlib_then_third_party_then_local_grouped
- explicit_imports_not_star_imports_used
ERROR_HANDLING:
- log_then_raise_precise_HTTPException_with_detail
- check_rate_limit_before_database_operation
- refund_credits_by_saving_values_before_session_closes
STYLE:
- short_direct_corrections_fix_agent_behavior_precisely
- comment_preservation_expected_across_rewrites
BANNED_PATTERNS:
- --single-process_Chromium_flag_on_macOS (reason: causes crashes, documented failure)
- hardcoding_connection_URLs_or_env_specific_values (reason: caused 'Queue service unavailable' failure)
- jumping_to_implementation_before_design_alignment (reason: wasted work when pricing model wasn't confirmed)
- removing_comments_during_code_rewrites (reason: user explicitly called out and expects preservation)
- using_browser_only_Node_APIs_in_subprocess (reason: ErrorEvent caused ReferenceError)
Updated: TASTE.md, .agent-taste.json
## Auto-Capture
taste learn automatically captures your current session if no recent session files exist:
``` bash
taste learn
What it captures:
- Git history (last 3 commits + diffs)
- Current taste config (
.agent-taste.json
orTASTE.md
) - Last 3-5 session log summaries
What it does NOT capture:
- Terminal histories
- Agent session logs
- Full file contents
Before taste:
Raw context:
- Session logs: ~40,000 words
- Git diffs: ~10,000 words
- README.md: ~1,000 words
- Session notes: ~500 words
Total: ~51,000 words (~66,000 tokens)
After taste:
Compressed context:
- Taste config: ~120 words
- Git summary: ~200 words
- Session notes: ~200 words
- Agent config: ~500 words
Total: ~1,500 words (~1,950 tokens)
Savings:
Words: 51,000 → 1,500 (97% reduction)** Tokens**: 66,000 → 1,950 (97% reduction)
Create ~/.config/taste/base.json
for global settings:
{
"flavor": "Standard idiomatic development",
"banned_patterns": [],
"style": "Prefer clarity over brevity"
}
Create .agent-taste.json
in your project root:
{
"flavor": "Functional TypeScript, strict types, zero dependencies",
"banned_patterns": [
"classes",
"any",
"console.log"
],
"style": "Implicit returns, max 20 lines per function"
}
After running taste learn, banned patterns are automatically populated:
{
"flavor": "Standard idiomatic development",
"banned_patterns": [
"--single-process_Chromium_flag_on_macOS (reason: causes crashes, documented failure)",
"hardcoding_connection_URLs_or_env_specific_values (reason: caused 'Queue service unavailable' failure)",
"jumping_to_implementation_before_design_alignment (reason: wasted work when pricing model wasn't confirmed)",
"removing_comments_during_code_rewrites (reason: user explicitly called out and expects preservation)",
"using_browser_only_Node_APIs_in_subprocess (reason: ErrorEvent caused ReferenceError)"
],
"style": "Prefer clarity over brevity",
"learned": {}
}
Does it need a config file?
No. An optional .agent-taste.json
or TASTE.md
can be created, but nothing is required. taste works with zero configuration.
What if I really need that 120-line cache class? You don't. Insist anyway and taste will learn your pattern. Slowly. Correctly. While looking at you.
Does it scale? The context you never waste scales infinitely. Zero tokens wasted, zero generic code, 100% style matching since forever.
Why "taste"? You know exactly why.
- bash 4.0+
- git
- opencode or claude (for
taste learn
)
- Multi-agent support (claude, codex, commandcode)
- Session auto-capture (daemon mode)
- Global taste config (
~/.config/taste/
) - JSON output for agents
- Integration with more agent harnesses
Contributions are welcome! Please feel free to submit a Pull Request.
- Fork the repository
- Create your feature branch (
git checkout -b feature/amazing-feature
) - Commit your changes (
git commit -m 'Add amazing feature'
) - Push to the branch (
git push origin feature/amazing-feature
) - Open a Pull Request
MIT. The shortest license that works.