Change a standing instruction

Principles in play: Control transitions, Evaluate corrections.TRACEControl transitions · Evaluate corrections

A standing instruction applies to later tasks by default, so changing one can affect work not yet done. Adopting the change is recorded as a boundary transition, and its effect is observed in a later run made under it rather than assumed.

Steps

1. Say the change, and adopt it.

   Say to your agent

   Add a rule to AGENTS.md: every check row names the table version it read. Adopt it as the project default and record the adoption.

   The adoption row starts out awaiting a later run.

2. Continue the work, then ask for the later runs.

   Say to your agent

   Which trace records could be the later run for that adoption?

   The list is found by instruction version, not judged.

3. Look at one, and say what you saw.

   Say to your agent

   I read the record from Thursday’s session. Every check row named the table version. Record that I observed it.

4. If the instruction contributed to a failure, name it.

   Say to your agent

   This error came from the rule in AGENTS.md, not just this one table. Name it as the faulty default on the defect, and list the records it affected.

   That finding can reach verified only through a targeted recheck of a later record made under the corrected instruction. The recheck names that record, which must exist in the project, and the commit that changed the instruction; see Raise, fix and recheck a defect.

Things to know

• Standing instructions are more than AGENTS.md. WORKFLOW.md, the skills and, when your project keeps one, ai/review-prose-rules.md apply to later work too, and changing them is a default adoption in the same way.
• No observation, no link. Until someone records looking at a later run, the adoption stays awaiting.
• Correcting one output leaves the default in place. A faulty instruction that is not revised remains available to shape later work.
• The row records; your words authorize. The adoption row names your sentence as its authority.

What the agent runs

python3 scripts/trace_record.py transition --kind default-adoption --object AGENTS.md \
    --commits <sha> --later-run ai/records/<later id>     # when the later record already exists
python3 scripts/trace_record.py later-run <TR-id>          # candidate records, found, not judged
python3 scripts/trace_record.py later-run <TR-id> --record ai/records/<later id> \
    --observed "what the revised instruction did in that run" --by "Full Name" --person
python3 scripts/defect_ledger.py mark --id <ID> --faulty-default <instruction file> --affected <record>