Provider-aware model + reasoning-level routing for AI execution workflows.
TEMO Efficiency helps an AI decide which tool, which model, and which reasoning/effort level to use before execution. It avoids using the strongest model by default, prevents repeated verified work, and escalates only when evidence shows that the current checkpoint needs more capability.
Use the smallest capable model. Preserve the acceptance criteria. Escalate only when evidence says you need to.
Load TEMO Efficiency
→ refresh the current canonical rules when possible
→ identify/confirm the AI provider/tool
→ verify the user's real selectable models + reasoning levels
→ ask for screenshots/list only when the catalog is not visible
→ build/reuse FAST / BALANCED / DEEP / MAX
→ score the current checkpoint
→ choose the exact model + level
→ SHOW THE CHOICE BEFORE EXECUTION
→ execute one bounded micro-checkpoint
→ verify the result
→ protect PASS / VERIFIED work
→ escalate one step only when evidence justifies it
→ optionally prepare evidence-based feedback
TEMO is an execution behavior contract, not just a prompt that says “use fewer tokens.”
EXECUTION CHOICE
Tool / Environment: Codex Local
Model: <exact verified model>
Profile: FAST
Level / Effort: <exact verified level>
Boost / Speed: OFF or Not exposed
Consumption: Low or Not exposed
Deploy: NO
Reason: The current checkpoint is deterministic and narrow.
Then copy and execute the command below.
When TEMO has enough verified information, it chooses for the user. The user should not have to guess which model or level to select.
Escalation is only:
FAST → BALANCED → DEEP → MAX
TEMO can be used with:
- ChatGPT / OpenAI
- Codex
- Claude / Claude Code
- Gemini
- Cursor
- Copilot
- Cloud Code
- local models
- other AI tools
TEMO does not assume that two users on the same provider have the same model catalog. Free/paid/professional tiers, experiments, apps, regions, and product surfaces may expose different options.
Use the real current model + level controls directly.
TEMO asks for the smallest evidence needed:
Send screenshots of:
1) the expanded model picker,
2) the expanded reasoning/level/thinking picker if separate,
3) any boost/speed/mode selector if present.
If screenshots are inconvenient, paste the exact visible labels instead.
If the AI/tool itself is unclear, TEMO may ask for one screenshot of the app/site header or settings page.
TEMO never builds a full ladder from memory, provider-family assumptions, or plan/tier assumptions.
See templates/CATALOG_CONFIDENCE_GATE.md.
TEMO uses a canonical refresh model.
At the start of a new session, when repository access exists:
- read
TEMO_LATEST.mdonce; - compare the loaded version with the current manifest;
- if the loaded copy is older, load the current canonical
SKILL.mdorTEMO_PORTABLE.mdbefore routing the first task; - do not keep checking repeatedly during the same session.
If the update check fails, TEMO continues with the loaded copy instead of blocking the user's task.
TEMO does not silently rewrite this GitHub repository from arbitrary user sessions.
Improvements enter through reviewed maintainer changes, GitHub Issues, or Pull Requests. This avoids a bad run automatically changing the canonical skill for everyone.
Feedback loop: learn from real users without hidden telemetry
TEMO does not silently upload conversations, screenshots, prompts, code, logs, account information, provider catalogs, or telemetry.
When a meaningful signal happens, TEMO may prepare an optional compact feedback report, for example:
- routing was too strong or too weak;
- provider/model discovery failed;
- a new provider exposes unfamiliar controls;
CATALOG_HALLUCINATION_FAILoccurred;- GitHub/portable fallback failed;
- a regression guard prevented repeated work;
- documentation was unclear;
- a cross-device/provider test produced a useful PASS/FAIL.
The user decides whether to submit it.
Use the GitHub TEMO Efficiency Feedback issue form or see docs/FEEDBACK_LOOP.md.
Feedback is useful when it contains minimal reproducible evidence, not private data.
TEMO has three official access modes:
A. Full repository access
→ use the canonical files
B. Repository search/navigation fails
→ stop after one failed lookup
→ load the raw portable contract:
https://raw.githubusercontent.com/luaysameer/temo-efficiency/main/TEMO_PORTABLE.md
C. No web access
→ upload/paste TEMO_PORTABLE.md once
→ use the core workflow from that one file
TEMO_PORTABLE.md is self-contained.
Paste:
Use TEMO Efficiency.
First try the canonical repository:
https://github.com/luaysameer/temo-efficiency
If repository navigation/search fails, do NOT keep retrying it. Immediately load:
https://raw.githubusercontent.com/luaysameer/temo-efficiency/main/TEMO_PORTABLE.md
Then follow TEMO Efficiency for my next task.
If the AI has no web access, upload only TEMO_PORTABLE.md and say:
Use the attached TEMO_PORTABLE.md as the TEMO Efficiency behavior contract for this chat.
TEMO scores the current checkpoint, not the importance of the whole project.
Five dimensions are scored 0–2:
- Complexity
- Risk
- Scope
- Verification burden
- Uncertainty
| Score | Profile | Typical level | Intended use |
|---|---|---|---|
| 0–2 | FAST | lowest reliable | deterministic edits, extraction, formatting, narrow checks |
| 3–5 | BALANCED | medium/default | focused implementation, normal debugging, targeted integration |
| 6–8 | DEEP | medium/high | difficult regressions, architecture, coupled systems |
| 9–10 | MAX | highest justified | exceptional complexity or high security/data/infrastructure risk |
A full exact ladder is created only from a verified catalog.
If only the current model is known:
CATALOG STATUS: PARTIAL
ROUTING MODE: CURRENT_MODEL_ONLY
Large work is split into narrow checkpoints.
Each checkpoint contains:
- Tool/environment
- exact model + profile
- exact reasoning/effort level when exposed
- boost/consumption state when exposed
- one primary objective
- protected / do-not-repeat work
- targeted diagnostics/tests
- success condition
- stop condition
- Deploy YES/NO
After meaningful verified success:
diagnose → fix → targeted test → regression guard → real acceptance → LOCKED_PASS
Future work should preserve that PASS unless relevant code, dependencies, environment, provider catalog, requirements, or evidence changes.
If a task requires local USB, Android ADB, local files, GPU, desktop UI, browser state, or attached hardware, TEMO should choose an execution environment that can physically reach it.
Use docs/CROSS_DEVICE_TEST.md.
It covers:
- full repository access;
- raw portable fallback;
- no-web single-file mode;
- unknown provider/catalog;
- different account model catalogs;
- catalog hallucination prevention;
- escalation behavior;
- PASS preservation.
SKILL.md— normative skill behavior, v1.5.0TEMO_LATEST.md— current-version manifestTEMO_BOOTSTRAP.md— startup/update/fallback flowTEMO_PORTABLE.md— self-contained portable contracttemplates/PROVIDER_DISCOVERY.md— provider/model/level discoverytemplates/CATALOG_CONFIDENCE_GATE.md— blocks model-catalog guessingtemplates/EXECUTION_CHOICE.md— model + level shown before executiontemplates/CHECKPOINT.md— micro-checkpoint contractdocs/FEEDBACK_LOOP.md— opt-in feedback workflowdocs/CROSS_DEVICE_TEST.md— portability/regression testsCONTRIBUTING.md— contribution guidanceREADME.ar.md— الشرح العربي
TEMO does not bypass quotas, billing, subscriptions, rate limits, plan restrictions, or safety controls.
It does not guarantee a fixed saving percentage.
It does not use hidden telemetry or silently self-modify the canonical repository.
The purpose is to reduce avoidable AI work while preserving the required quality bar.
MIT — use it, test it, adapt it, and improve it.