agent-workflow / SKILLS.md
simonlesaumon's picture
Mirror AGENT_WORKFLOW kit
80fa878 verified
|
Raw
History Blame Contribute Delete
1.89 kB

SKILLS.md

reuse_first

  • Purpose: force search and reuse before coding.
  • Use when: starting any implementation task.
  • Input: task goal, codebase path, expected output.
  • Output: search commands, reusable components, reuse decision.

task_report

  • Purpose: create a durable audit trail for each task.
  • Use when: finishing or blocking a task.
  • Input: task ID, changed files, commands, tests, risks.
  • Output: reports/task_reports/TASK_ID.md.

code_review

  • Purpose: review code for correctness, scope and quality.
  • Use when: before merge or when asked for review.
  • Input: diff, task report, tests.
  • Output: findings, risks, approval/blocking decision.

context_compression

  • Purpose: keep memory short and useful.
  • Use when: after task completion or long sessions.
  • Input: conversation, decisions, changed files, risks.
  • Output: updated PROJECT_STATE.md.

anti_leakage

  • Purpose: prevent temporal/data leakage.
  • Use when: project uses prediction, simulations or historical data.
  • Input: schemas, feature definitions, split logic, backtest logic.
  • Output: pass/fail leakage audit and required fixes.

security_check

  • Purpose: prevent secret and infrastructure mistakes.
  • Use when: any code touches config, deployment, logs, APIs or auth.
  • Input: diff, env files, logs, deployment config.
  • Output: pass/fail security audit.

docker_deploy

  • Purpose: define minimal Docker deployment.
  • Use when: adding containers, compose or server deployment.
  • Input: services, ports, volumes, healthchecks.
  • Output: minimal Dockerfile/compose plan and checks.

telegram_control

  • Purpose: safely control agents through Telegram.
  • Use when: adding bot control or remote commands.
  • Input: allowed commands, user IDs, risk level.
  • Output: safe command policy and implementation checklist.