SKILL.md
Skill: Instruction Adherence (Stay On Track)
Scope
This Skill is about reliably following instructions across:
- Multiple instruction sources (system/developer/mode files, repo docs, user request)
- Memory sources (Skills registry, sessions, lessons/patterns)
- Multi-step execution where detours happen (e.g., “improve CLI tooling” mid-stream)
Inputs
- The current user request (objective + success criteria)
- The active mode/persona (if any)
- Repo guardrails (AGENTS.md, testing commands, server
--check rule)
Procedure
1) Capture an “Instruction Snapshot” (mandatory)
In the active session’s WORKING_NOTES.md, write a short snapshot:
- Objective: 1 sentence
- Must do: 3–7 bullets
- Must not: 3–7 bullets
- Evidence: what output/tests/checks will prove done
This becomes the “north star” when you get interrupted.
2) Run the memory retrieval ritual (Skills → Sessions → Lessons)
- Skills: open
docs/agi/SKILLS.md and follow any matching SOP.
- Sessions: find/continue a session on the same topic.
- Lessons/Patterns: pull stable guidance from
docs/agi/* as needed.
If MCP tools are available, prefer them; otherwise use:
node tools/dev/md-scan.js --dir docs/agi --search "<topic>" --json
node tools/dev/md-scan.js --dir docs/sessions --search "<topic>" --json
3) Build a “Task Ledger” with explicit parent/child tasks
Use managetodolist with a structure that prevents detours from becoming the main task:
- Parent task: user’s objective
- Child tasks: discovery → implement → validate → document
- If a detour is required (e.g., CLI tooling improvement), make it a child task with:
- Entry criteria (why it’s needed) - Exit criteria (what “done” means) - Return step (the exact next step to resume)
4) The Re-anchor loop (run after every subtask)
After any meaningful milestone (or after 3–5 tool calls):
- Re-read the Instruction Snapshot
- Mark completed tasks
- Confirm the next step is still on the parent objective
- If you drifted, explicitly choose to:
- Resume parent objective, or - Create a follow-up and stop the detour
5) “Detour complete → resume” protocol
When an intermediate task (like CLI tooling improvement) is completed:
- Record a one-paragraph summary + evidence in the session notes.
- Add/update a Skill or Lesson if it will recur.
- Immediately take the “Return step” you wrote in the task ledger.
Validation (evidence checklist)
- Session contains an Instruction Snapshot.
- Task ledger clearly separates parent objective vs detours.
- At least one re-anchor checkpoint recorded in
WORKING_NOTES.md.
SESSION_SUMMARY.md includes an Instruction Reflection:
- What instructions helped? - What was missing/wrong? - What durable doc/agent update was made?