Know Today
👀 Keep an eye on
- Explicitly invoked agent skills can reduce unwanted routing and persistent context cost in consequential workflows.
- Recurring AI corrections are becoming a practical input for building small, durable operating instructions.
- Recognised frameworks such as refactoring and premortems may outperform long bespoke prompt definitions.
- Instruction libraries need deletion tests, because unchanged output reveals rules that consume context without changing behaviour.
- AI-assisted coding raises the importance of named human ownership for shipped software outcomes.
🧭 What changed
Skills can be intentionally manual
Automatic invocation is a convenience setting, not an obvious default: manual activation keeps costly or high-impact instructions visible to the operator.
Production rework can become reusable process data
A specific failure can justify one compact instruction, turning ordinary delivery friction into a maintained capability rather than another forgotten chat correction.
Prompt maintenance looks more like software maintenance
Rules can be tested, pruned, and scoped instead of accumulating as a permanent and opaque policy layer.
💭 Opinions worth testing
Opinion: Create a skill for a repeated correction, not simply for a frequent task.
Opinion: Broad end-to-end skills encourage premature artifact production and should be decomposed into process stages.
Opinion: Established technical vocabulary activates more useful model behaviour than invented labels with lengthy definitions.
Opinion: Short, sharply scoped skills are often stronger than comprehensive instruction packs.
Opinion: User-invoked skills are often more controllable and token-efficient than automatically invoked skills.
Opinion: Low-stakes skill iteration can use rework and complaints as feedback, although high-impact workflows require stronger validation.
Opinion: Add instructions only after observing the precise failure they are meant to prevent.
Opinion: State the desired agent behaviour directly instead of centring prompts on prohibited behaviour.
Opinion: AI reduces the value of purely tactical programming but does not reduce responsibility for a merged change.
Opinion: Teams should derive skills from their own operating reality rather than adopt external configurations unchanged.
Reflection: Mine recurring corrections. Encode one constraint per skill. Delete rules that do not change representative output. Assign an accountable owner to every consequential AI-assisted change.
🛠️ Practical workflows
Correction-to-skill loop
Capture a “send it back” moment, identify the missed process step, write the smallest instruction that fixes it, then test it on live work.
Instruction deletion pass
Remove one line from an agent configuration, run representative tasks, and retain the line only when behaviour materially changes.
Authority-aware execution
Let agents prepare and execute within defined scopes; require explicit confirmation before consequential implementation, deployment, or merge actions.
🚀 Build something ambitious
Operating-memory system for an AI-native product team
Outcome: Turn delivery mistakes into compounding execution quality instead of recurring coordination cost.
Workflow: Ingest agent task traces, code-review feedback, support issues, and delivery outcomes; cluster recurring corrections; propose a scoped skill; evaluate it against representative work; route approved skills to the right planning, implementation, testing, and release stages; measure rework and escalation rates.
Connected capabilities: Coding agents, issue tracking, pull requests, test runners, deployment systems, trace storage, lightweight evaluations, and a human escalation channel for authority boundaries.
First meaningful milestone: Ship a service that identifies the five most repeated corrections from a month of engineering work and generates tested, single-purpose candidate skills.
Context-budget governor for multi-agent delivery
Outcome: Give teams faster, more predictable agent execution while preserving intentional control over costly or consequential instructions.
Workflow: Classify tasks by stage and consequence; select only relevant manual or automatic skills; expose the selected instruction set before execution; record token use, routing, rework, and outcome quality; continuously prune rules that do not earn their context cost.
Connected capabilities: Agent runtimes, skill registries, task classifiers, telemetry, evaluation datasets, code repositories, and approval gates for implementation and release authority.
First meaningful milestone: Build a dashboard and routing layer that compares manual versus automatic skill invocation across one engineering workflow and reports quality, cost, and rework deltas.
⚡ Fast builds to try
- Make a correction logger: tag the last ten AI responses you had to revise and produce a ranked list of recurring failure patterns.
- Write a seven-line review skill: use a familiar framework such as a premortem, invoke it before a design decision, and get a compact failure-focused critique.
- Run an instruction subtraction test: remove one line from an existing agent prompt and compare the result on three representative tasks.
- Create an approval phrase library: define explicit confirmations for implementation, merge, and deployment actions and make agent handoffs unambiguous.
⚠️ Caveats
- Reported skill invocation behaviour, context loading, and token effects need confirmation against current platform documentation and usage telemetry.
- Claims about familiar terminology activating model priors are plausible but model-, version-, and task-dependent.
- Informal production feedback is not an adequate evaluation method for security-sensitive, regulated, or high-impact workflows.
🧠 The author's perspective
The author believes that today’s signals align with a future in which software development and company operations become increasingly autonomous through capable runtimes, durable instructions, integrations, data, and appropriate authority. The emphasis on extracting skills from real corrections supports treating operational learning as unique decision data, while scoped skills and staged workflows support harnesses that coordinate planning, implementation, testing, review, and deployment rather than isolated chat tasks. Deliberate invocation, measured context use, explicit confirmation for consequential actions, and named ownership fit an autonomy model where agents act by default within defined authority, escalate when real-world help is needed, and improve from outcomes through reusable skills.