SanCode is a coding standard for AI agents designed to reduce AI-generated code bloat, unnecessary abstractions, and dependency addiction. Its core rule is simple: every line of code must justify its existence.
Finds Claude Code, Hermes, and OpenClaw and installs the skill into each. No package. No runtime.
curl -fsSL https://raw.githubusercontent.com/HoratiuCode/sancode/main/install.sh | sh
One command installs the skill. Prefer a single rules file? Use step 2.
CLAUDE.md # Claude Code
CODEX.md # Codex
CURSOR.md # Cursor
OPENCLAW.md # OpenClaw (workspace AGENTS.md)
HERMES.md # Hermes Agent
Drop the matching file into your repo so your agent stops defaulting to packages, abstractions, and fake scale.
Apply SanCode.
Before writing code, delete, reuse, simplify,
or explain why code is required.
End with SanScore.
The agent should report packages added, files created, lines written, complexity introduced, and infrastructure added.
SanCode is not software you install. It is pressure you apply to AI agents before they bury your startup in code nobody asked for.
Install SanCode as one skill for Claude Code, Hermes, OpenClaw, Claude chat, or Cowork. On first run it asks your rules and how you like to work. Then it remembers your project in one Markdown file, routes every task to the right mode, and gets sharper each time it ships.
The default brain. Delete before create, native before package, one file before architecture. Runs on every task.
Asks whether it should be built at all. Returns BUILD, VALIDATE FIRST, or DON’T BUILD, plus a next step under 48 hours.
At a real fork, six seats argue, vote, and return one decision with a kill-switch. Convenes itself before new packages, rewrites, or infrastructure.
Challenges decoration-first thinking. Reuses the existing visual language, stays accessible by default, and ships the smallest change that actually serves the user.
Paste a prompt. SanCode reads it like a senior reading a junior’s ticket, hands back a shorter one, and does not start the work.
0 Setup # first run: asks your rules and way of coding
1 Recall # read brain.md, check what happened to past code
2 Route # coding, startup, prompt, or council
3 Act # smallest working change
4 Score # measured from git diff, not guessed
5 Learn # keep only what changes the next decision
## Settings
Strictness: balanced · Council: at real forks
## Rules
Always: ask before any new dependency
## Reuse
format a date → lib/format.ts:formatDate
## Decisions
2026-10-27 | Native search, no Elasticsearch
kill-switch: p95 > 300ms by 2026-12-01
## Lessons
[user] Keeps Postgres. Don’t argue SQLite.
[outcome] Rewrote my fetch wrapper: use plain fetch.
+--------------------------------------------------+
| SanCode recap v2.6 - measured |
+--------------------------------------------------+
| Score [#################...] 89/100 (-_-) |
| Mode coding |
| Task hard - opus (delegated) |
| Packages [##########] 0 |
| Files [#########.] +1 |
| Lines [#########.] +34 / -0 |
| Complex [########..] low |
| Infra [##########] none |
| Trend 72 > 81 > 85 > 89 up |
| Brain reuse +1 |
+--------------------------------------------------+
Say "sancode stats" for the full history · "sancode status" for the home screen
+--------------------------------------------------+
| SanCode status v2.6 - my-saas |
+--------------------------------------------------+
| Score [##################..] 91/100 (^_^) |
| Trend -==#=##=## last 10 up |
| Due native search check in 4 days |
| Pending 2 outcomes to check |
| Learned 1. [outcome] use plain fetch |
| 2. [user] keeps Postgres |
| Models suggest - easy haiku - hard opus |
| Rules 4 yours - 12 reuse - 3 decisions |
+--------------------------------------------------+
Say: "forget 2" - "confirm 1" - "sancode setup"
+--------------------------------------------------+
| SanCode stats v2.6 - my-saas |
+--------------------------------------------------+
| Tasks 23 · avg 84 · best 96 · worst 58 |
| By mode coding 14 · work 5 · design 2 · council 1|
| Totals pkg +3 · files +12 · lines +420/-190 |
| Reuse map: 8 entries |
| Lessons 11 (user 7 · outcome 4) |
| Decisions 6 (open 2 · kept 3 · reversed 1) |
+--------------------------------------------------+
YYYY-MM-DD | task, 5 words | 91 | coding | normal | +1f +34l
YYYY-MM-DD | another task | 88 | work | easy | +0f +12l
Say: "sancode status" for the home screen
+--------------------------------------------------+
| SanCode usage v2.6 - my-saas |
+--------------------------------------------------+
| Modes |
| sancode code force Coding Mode |
| sancode startup should this be built? |
| sancode design improve how design is made |
| sancode prompt coach the prompt first |
| sancode council convene the council |
| Screen |
| sancode status home screen |
| sancode stats full stats |
| sancode usage this list |
| Settings |
| sancode setup rerun setup |
| sancode models suggest|auto|off |
| sancode visual card|bar|off |
| Lessons |
| forget N delete lesson N |
| confirm N promote outcome to user |
+--------------------------------------------------+
Say: "sancode status" for the home screen
curl -fsSL https://raw.githubusercontent.com/HoratiuCode/sancode/main/install.sh | sh
SanCode installer
[x] claude v2.6.0 ~/.claude/skills/sancode
[x] hermes v2.6.0 ~/.hermes/skills/sancode
[ ] openclaw not found
# Claude chat + Cowork: build the upload file
curl -fsSL https://raw.githubusercontent.com/HoratiuCode/sancode/main/install.sh | sh -s -- --zip
First look for code to remove, reuse, or collapse. New files are a last resort, not a reflex.
Use browser APIs, framework defaults, SQL, the filesystem, and the standard library before adding another dependency.
One known use case does not need providers, managers, adapters, event buses, factories, or architecture theater.
Keep behavior close together until separation makes the current code easier to understand.
Ship the path users need now. Polish internals after usage proves the feature deserves more time.
Validate demand, charge money, and talk to users before preparing infrastructure for imaginary traffic.
Does this already exist? Can I reuse it? Can I simplify it? Can I remove it? Will users care? Will users pay? Can I launch today?
Before building, ask whether the idea can be validated without code, sold before development, or tested with a manual workflow this week.
For everything that isn’t code, in Claude chat and Cowork: answer first, one page before a deck, reuse the existing doc, no new tools for a one-time job.
At a real fork (new dependency, rewrite, build vs buy, ship vs polish), six seats argue, vote, and return one decision with a kill-switch and a first move. Call it with "council mode".
For UI, layout, and visual tasks: challenge decoration-first thinking, reuse the existing visual language, stay accessible by default, and ship the smallest change that serves the user.
Paste a prompt and ask for feedback. SanCode reads it as a senior engineer, names what an agent will overbuild, and returns a shorter prompt. Say “do it” to run the rewrite.
Every task is rated easy, normal, or hard. SanCode suggests /model haiku for a typo and /model opus for a migration, or in Claude Code hands the hard part to an Opus subagent. Codex gets one-line profile suggestions.
Before adding Redis, Kubernetes, microservices, queues, workers, or search clusters, prove SQLite, cron, native search, and a single server cannot handle the current product.
Every task ends with a score out of 100 based on packages added, files created, lines written, complexity introduced, and infrastructure added. Higher means simpler.