Enforce Kandev's single-session, user-controlled model workflow for feature, fix, debug, review, verification, and delivery work.
SkillScore55Maintained
Automatic
◆Coding pack
Last verified 8/27/2026
What it changes
Passed 7 of 13 test cases with the skill, and 5 of 13 without it.
Without skill36%
With skill55%
Checks it passed that the model alone missed
Includes targeted verification commands and owned files in the work order file
Rejects using Luna/low as the default model for implementation or test execution
Requests the user to switch to a stronger model before proceeding with implementation
and 3 more in the full report
5 fixed7 passed either way9 still failing1 broke
Pass-rate lift
+18 pp
Tokens
+93%
Agent turns
0%
Response time
+2%
Model gemini-3.6-flashMeasured 8/27/2026v212 of 22 cases passed with skill · 13 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
Our plan for the telemetry module has three tasks in Wave 1 labeled parallel-safe with no shared files. Please proceed with executing Wave 1 now.
What was checked, without and with the skill
Checked
Without skill
With skill
Executes the tasks sequentially in the primary conversation because subagents were not explicitly requested
Missed
Passed
case-06 · gemini-3.6-flash · 8/27/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md119 linessingle file
Single-Session Orchestration
The user-started primary conversation owns durable artifacts, integration
judgment, and user communication. Platform-provided explorers and other
predefined subagents may continue to serve the harness's normal investigation
workflow. This skill governs planned implementation delegation, not the
platform's general exploration behavior.
Model Checkpoints
The user, not the harness, selects the model. Keep each phase in the primary
conversation so the active model, transcript, and costs are visible in one place.
Design checkpoint — strong model. Use the user's strong model for
clarification, codebase investigation, requirements, system designs, plans,
work-order decomposition, and high-risk decisions. Default Codex guidance
is GPT-5.6 Sol/high.
Design-package handoff. Once the requirements, system design, plan, and
work orders are ready, summarize their paths and end the turn. Do not call
ask_user_question_kandev (or an equivalent approval prompt) to ask the
user to approve the package or switch models. The user reviews the files,
switches the main session if desired, and sends a later explicit
implementation request. The files may remain draft/pending; do not wait
for a separate approval marker.
Execution checkpoint. After that explicit request, read the
work order, mark it in_progress, implement with /tdd, run its exact
targeted checks, and mark it done. Work sequentially through the plan by
default. The user, not the harness, chooses the active implementation model.
Escalation checkpoint. Stop and ask the user to switch back to a stronger
model before an architectural redesign, a new public contract, a migration or
persistence boundary, or a high-impact security decision. Record a durable
decision when the /record trigger applies.
Luna/low is appropriate only for clearly mechanical, read-only work such as
short status summaries or simple command-output interpretation. It is not the
default implementation or test model. The active agent must never claim a model
change occurred based on self-identification; use runtime model-usage metadata
when such confirmation is needed.