tune-meta-project
Skill0 starsChanged yesterdayMITView on GitHub
Evidence-based calibration of the installed agent environment: analyzes actual work evidence (tasks, telemetry, observations, reports, history), forms verifiable improvement proposals with explicit effect verification, and routes changes through manage / decommission / update mechanisms.
In this repository19 other files it gives its agents.
- AGENTS.mdAGENTS.md
- templates/AGENTS.mdAGENTS.md
- decommission-meta-projectSkill
- ground-meta-projectSkill
- init-meta-projectSkill
- manage-meta-projectSkill
- record-observationSkill
- scan-meta-projectSkill
- sync-architectureSkill
- update-meta-projectSkill
- decommission-meta-projectSkill
- ground-meta-projectSkill
- init-meta-projectSkill
- manage-meta-projectSkill
- record-observationSkill
- scan-meta-projectSkill
- sync-architectureSkill
- tune-meta-projectSkill
- update-meta-projectSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 37
Use existing data: tasks.md, observations, reports, checks, change history, telemetry WITH confirmed identity.
- Distinctive: line 59
Check: duplicate rules across passport / skills / system prompts.
- Distinctive: line 80
If framework version, harness, and task type all changed, do not attribute the result to one setting without grounds.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill 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
Repository Agent Environment Tuner (Tune Meta-Project)
Turns project experience into VERIFIABLE improvements: fewer wasted actions, less context loss, fewer unnecessary handoffs — while preserving verification quality, the project's meta-vision, and agreed authorities.
When to invoke skill
- The user requests: "tune-meta-project", "калибровка обвязки", "прокачай инстанс", "устрани трение", "оптимизируй команду", "recalibrate environment".
- Accumulated friction is SUSPECTED (wasted handoffs, repeated context loss) — as a question to investigate, not a verdict.
Execution steps
Phase 1: Configuration Survey (read-only analysis)
Establish the frame with canonical utilities — calling them directly, never via scan's writing scenario:
python scripts/inspect_meta_project.py --json --harness <h> --installation <i>python scripts/verify_identity.py --checkpython scripts/check_boundaries.py --boundaries .agents/project_knowledge/boundaries.md --audit-tree- installed framework state:
.agents/framework-lock.jsonwhen present; unfinished updates halt tuning of affected components. Record: agreed scenario, active installations and capabilities, roles and bindings, verification policy, calibration scope (whole project or one installation). Calibrating ONE installation never changes shared rules of another; common parts consider both.
Phase 2: Evidence Analysis (not heuristics)
Use existing data: tasks.md, observations, reports, checks, change history, telemetry WITH confirmed identity. Interpretation rules (mandatory):
- repeated edits of one file are a signal TO STUDY — normal TDD is not a verdict;
- thirty days without invocations does not prove a role useless;
- frequent gate blocks may be CORRECT protection;
- frequent ground calls may follow session churn;
- git authorship does not determine the project role;
- missing telemetry means insufficient data — state it, don't guess. Choose the analysis window per question (period / completed-task set / installation); a fixed short commit window is only an optional overview. Separate FACTS (e.g. "five handoffs between roles") from HYPOTHESES (e.g. "the dispatcher allocates poorly" — needs verification).
Phase 3: Verifiable Proposals
For each proposal record EXACTLY: Наблюдаемая проблема / Свидетельства и ограничения данных / Предполагаемая причина (hypothesis) / Предлагаемое изменение и область / Ожидаемый эффект / Способ проверки / Сохраняемые ограничения / Способ отмены / Исполнитель (own the fact–hypothesis split). Select a SMALL set with the highest expected value. "No justified proposals" is a normal result — say so explicitly.
Phase 4: Instruction & Context-Cost Calibration
Check: duplicate rules across passport / skills / system prompts; scenario↔ instruction contradictions; procedures loaded constantly but rarely needed; rules without observable behavioral impact; stale temporary workarounds; excessive mandatory reporting/confirmations. The North Stars limit applies to the TOTAL active set — propose consolidation (tactical details move to skills/knowledge); never delete accepted restrictions to hit a number; a new rule must state its application situation and expected behavior change.
Phase 5: Application via Mechanisms (after user confirmation)
roles → manage-meta-project; binding disable / archive → decommission;
framework → update-meta-project; local instructions → local_repo_data
mechanism; boundaries → existing permission policy (never weaken protection
to bypass a defect — a framework defect is recorded and fixed, not routed
around). Do not re-confirm already-granted authorities; before requesting a
NEW permission present a concrete diff and available checks. A rarely-used
role is judged by purpose (independent verification may stay mandatory).
Phase 6: Effect Verification Record
Persist per applied change: original observation / expected result / re-evaluation condition / actual result / significant condition changes. Statuses: applied-not-verified (applied, not verified) | confirmed | no-effect | regression. If framework version, harness, and task type all changed, do not attribute the result to one setting without grounds. Re-evaluate on comparable tasks or the next tune run; no background observation needed. Rollback restores only the corresponding change, preserving subsequent user work.
Phase 7: Repeatability
Same inputs → no new rules for the sake of running; no repeating rejected
proposals without new grounds; no duplicate observations; no team rebuild;
no files written in analysis mode. Store accepted AND rejected proposals
with rationale and scope (local knowledge slice, via the tuning report and
local_repo_data). Local calibration is an adaptation of the installed
instance — it never bumps the upstream framework version and MUST be
preserved by update-meta-project.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
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.

