Run the delivery line
22 primitives in this stage — 21 skills · 1 agent. Concept-stage catalogue, kept vendor-agnostic.
Auto-follow-up on unanswered requests →
“Manual chasing is the kind of admin that falls through the cracks when I'm heads-down on analysis — this runs in the background so nothing gets forgotten.”
Plan the retrieval across agreed sources →
“I used to just start downloading things — then I'd find I'd grabbed the wrong version from the wrong folder and had to start over.”
Reconcile retrieved files against the source manifest →
“We hit the completeness review and realised one source had never actually been pulled — this catch would have saved us a day.”
Assess inputs against requirements and flag gaps →
“I started cleaning data that turned out to be missing a whole segment the rubric needed — the gap check would have stopped me wasting half a day.”
Make the proceed / blocked recommendation →
“I kept pressing forward with partial data and the client felt ambushed when findings had caveats — a clear blocked call up front would have been cleaner.”
Derive the cleaning and normalisation spec from the raw data →
“I applied cleaning rules in my head as I went and couldn't reproduce the dataset a week later — writing the spec first costs ten minutes and saves hours.”
Map cleaned data onto the methodology's rubric structure →
“The client's data structure never matches the rubric's structure — working out the mapping in advance stops me improvising mid-analysis.”
Confirm the working dataset is diagnostic-ready →
“I started a diagnostic and hit an unlabelled section three modules in — a final readiness check before handoff catches exactly that.”
Engagement brief synthesiser →
“I'd context-switch from another project and start the diagnostic without re-reading the scope — this forces a clean re-entry every time.”
Method version & module selector →
“Running a module that was explicitly out of scope wastes time and confuses the client — locking the config before starting prevents that.”
Diagnostic module execution assistant →
“Each module has its own logic and I can't hold all of it in my head — the assistant walks me through the procedure so nothing gets skipped.”
Anomaly & edge-case triager →
“I either ignored odd data points or over-escalated them — having a structured triage call means I know which edge cases actually matter.”
Structured observation capturer →
“I used to blend observation and interpretation in my notes and couldn't separate them when scoring — capturing them as distinct entries fixes that.”
Rubric dimension scorer →
“Scoring from memory always produced ratings I couldn't defend — walking the rubric with evidence in hand means every score has a reason.”
Issue taxonomy classifier →
“Every finding was described in its own prose and there was no way to see patterns across them — taxonomy tags make the shape of the problem visible.”
Expert verdict arbiter →
“The rubric said one thing and my read of the data said another — having a formal place to record and justify that override stops the verdict from living only in my head.”
Findings record assembler →
“Scores, tags, and verdicts were in three different places and assembly always surfaced gaps or contradictions — one consolidated pass catches them all.”
Draft a targeted input request →
“Vague requests for 'more data' sat in client inboxes for days — a specific, bounded ask gets a faster answer.”
Frame mid-delivery clarifying questions →
“I'd make an interpretation call and later find out I'd guessed wrong — framing the question cleanly upfront saves rework and keeps the client in the loop.”
Capture answers into delivery context →
“The client answered and I carried it in my head — then forgot exactly what they said when the same edge case came up two days later.”
Draft a blocker and revised-timeline note →
“Silently slipping the deadline and hoping nobody noticed never worked — a clean blocker note with a revised date is always better than a missed one with no warning.”
Compose a mid-delivery progress update →
“When I went quiet during a delivery the client started emailing asking if everything was okay — one proactive update prevents ten reactive ones.”