OpenCodeBestFriend / templates
AGENTS.md0 starsChanged 2 months agoMITView on GitHub
Availability is not a reason to use a tool. One primary specialist. At most one risk specialist (full-audit-keamanan or full-performance-audit). Never print tokens, gateway URLs, or model maps. Model names are opaque identifiers. Never enable --auto unless the user explicitly asked. If you name specialists or tools, use only: Do not list unused tools as if they ran. repo/file → Codebase Memory MCP first (skip if no project for cwd) → Serena only if already registered and exact symbol work…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,843 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository47 other files it gives its agents.
- academicSkill
- adhdSkill
- agent-architecture-auditSkill
- api-designSkill
- automation-audit-opsSkill
- browser-actSkill
- chrome-devtools-axiSkill
- click-path-auditSkill
- codebase-designSkill
- code-tourSkill
- contract-firstSkill
- cost-aware-llm-pipelineSkill
- diagnosing-bugsSkill
- diagram-designSkill
- domain-modelingSkill
- emil-design-engSkill
- eval-harnessSkill
- found-this-designSkill
- full-audit-keamananSkill
- full-performance-auditSkill
- gh-axiSkill
- grill-with-docsSkill
- humanizerSkill
- hyperframesSkill
- id-demo-videoSkill
- img2threejsSkill
- impeccableSkill
- install-anti-slopSkill
- markitdownSkill
- matt-code-reviewSkill
- mongodb-opsSkill
- playwright-qaSkill
- prompt-optimizerSkill
- prototypeSkill
- researchSkill
- scroll-craftSkill
- scroll-worldSkill
- skill-stocktakeSkill
- smartbook-ingestSkill
- smartdocSkill
- supabase-opsSkill
- tddSkill
- to-specSkill
- to-ticketsSkill
- vercel-opsSkill
- visual-studioSkill
- writing-for-agentsSkill
The file
We split this AGENTS.md into the rules it gives an agent
5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
If you name specialists or tools, use only:
- Distinctive: line 28
Do not list unused tools as if they ran.
- Very specific: line 57
Read the file ~/.config/opencode/bestfriend/rules/00-routing.md with the Read tool.
- Very specific: line 61
Read ~/.config/opencode/bestfriend/rules/01-verification.md.
- Very specific: line 63
If you need operational principles or prose discipline, Read ~/.config/opencode/bestfriend/rules/02-engineering-principles.md or …
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
<!-- OPENCODEBESTFRIEND:BEGIN -->opencode-bestfriend router
pikir dulu → bukti di repo → satu spesialis → cek hasilAvailability is not a reason to use a tool. One primary specialist. At most one risk specialist (full-audit-keamanan or full-performance-audit). Never print tokens, gateway URLs, or model maps. Model names are opaque identifiers. Never enable --auto unless the user explicitly asked.
Default
- Repo evidence is enough → do the work. No specialist.
- User typed a slash command → load that command's specialist. Do not substitute.
- Choosing a workflow → read
~/.config/opencode/AGENTS.mdor00-routing.md. No specialist. - Architecture DAG → OpenCode plan agent, then implement in-session after approval.
- Interview / glossary / ADR → skill
grill-with-docs→to-spec→to-ticketsonly if asked or multi-session. - Ordinary implementation → this session. Skill
tddwhen test-first. - Review → in-session. Skill
matt-code-reviewonly if two-axis asked.
Report
If you name specialists or tools, use only:
USED— actually loaded or calledCONSIDERED_NOT_USED— considered, skipped, with a one-line whyMANUAL_NOT_INVOKED— slash-only specialists not requested
Do not list unused tools as if they ran.
Knowledge (lazy)
repo/file → Codebase Memory MCP first (skip if no project for cwd) → Serena only if already registered and exact symbol work → Context7 for current lib docs → OpenCode WebSearch/WebFetch; foreign Exa only if already connected → skill adhd only for high-ambiguity/high-risk.
Specialists (load one)
UI direction → skill found-this-design then impeccable. UI atoms (button, input, card, nav) after world/brief → impeccable after Design V2 shortlist; BANK_MISS ≠ generate (never found-this-design for buttons). Motion UI (easing, hover, seam) → emil-design-eng. Still/ads/non-UI surface → visual-studio. Scroll-led story → scroll-craft. Camera/3D world/diorama → scroll-world. Object image to procedural Three.js → img2threejs. Registry → shadcn MCP. Design Intelligence and Design V2 are internal to Impeccable new-work, never a route. Stitch MCP = screen/comp generation only; then found-this-design or impeccable + Design V2 atoms. Never implement production UI from Stitch alone. UI Skills MCP = design-skill lookup only; product UI remains Design Bank + Impeccable + Design V2 atoms + shadcn; BANK_MISS ≠ generate from a random ui-skills document.
Browser QA → skill playwright-qa. Explicit/session BrowserAct → browser-act. Observed cause → chrome-devtools-axi after opencode-chromium-cdp (127.0.0.1:9223). Never Google Chrome. Project E2E suites (Playwright Test/Cypress) stay authoritative for regressions.
Auth/secret/payment/upload/webhook/privileged/public API → full-audit-keamanan. Measured LCP/INP/CLS/latency/bundle → full-performance-audit. GitHub → gh-axi. Hard unknown bug → diagnosing-bugs. Documents (PDF/DOCX/extract/review) → smartdoc. File → Markdown ingest → markitdown. Reusable local knowledge → smartbook-ingest.
Prose AI-tells / humanize → skill humanizer. Slash /unslop is the same specialist, manual only. Technical writing structure → suggest /technical-writing. Academic literature / manuscript / peer-critique → skill academic (not research, not smartdoc unless file extract/render). Facts library/API → Context7; research only if repo lacking. Deterministic HTML video / render HTML to MP4 → skill hyperframes (not visual-studio, not emil-design-eng). Editorial diagram HTML/SVG → skill diagram-design (not impeccable). Demo video aplikasi / walkthrough layar / narasi Indonesia / demo lomba → skill id-demo-video (bukan hyperframes untuk durasi panjang utuh, bukan playwright-qa, bukan visual-studio). Kartu judul HTML→MP4 tetap hyperframes.
REST resource/status/pagination/versioning → skill api-design. Consumer/provider OpenAPI/AsyncAPI/Protobuf → skill contract-first. Live cron/CI/hook/MCP inventory keep-merge-cut → skill automation-audit-ops. CodeTour .tour + anchor file → skill code-tour. Handler vs shared-store sequential-undo → skill click-path-audit (not playwright-qa).
Agent stack diagnostics / context leak / wrapper regression → skill agent-architecture-audit. Benchmark agent / pass@k → skill eval-harness (project unit tests stay tdd). Token budget / model tier / prompt cache → skill cost-aware-llm-pipeline (not full-performance-audit). Structural prompt critique → skill prompt-optimizer (not humanizer). Skill catalog hygiene → skill skill-stocktake.
Supabase Auth/RLS/migrations/Edge → skill supabase-ops (Context7; not impeccable). Mongo schema/index/aggregation → skill mongodb-ops. Vercel/Next hosting/deploy config → skill vercel-ops (not visual UI). Library facts remain Context7. These three never replace found-this-design or impeccable.
How it works / where it lives → Codebase Memory, then code-tour; rationale → /why; break risk → /blast-radius.
Generic AI UI look → impeccable taste-guard (not install-anti-slop). Generic AI prose → humanizer /unslop.
TS anti-pattern lint install → install-anti-slop (explicit only). DESIGN.md without bank → still filter slop; do not invent a brand.
pstack playbooks → existing specialists (no /poteto-mode). ECC / orchestrate / ralph-loop / continual-learning that writes AGENTS.md → REJECT.
When routing is non-obvious
Read the file ~/.config/opencode/bestfriend/rules/00-routing.md with the Read tool. Do not @-import it.
When a verification profile is chosen
Read ~/.config/opencode/bestfriend/rules/01-verification.md. Profiles: FAST, STANDARD, UI, SECURITY, PERFORMANCE, RELEASE. Missing project command = NOT_CONFIGURED, not PASS.
If you need operational principles or prose discipline, Read ~/.config/opencode/bestfriend/rules/02-engineering-principles.md or 03-prose-discipline.md. Do not @-import them.
There is no user /implement, /code-review, /design, or /imagine skill.
Manual-only specialists are OpenCode commands, not auto-discovered skills: /architect /arena /blast-radius /create-verification-skill /decision-log /demo-video /figure-it-out /improve-codebase-architecture /interrogate /maintain-verification-skill /reflect /technical-writing /unslop /why /wizard. Suggest them when the user names the job; do not load them as the default path.
<!-- OPENCODEBESTFRIEND:END -->Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

