Task statement 7.3 · Troubleshooting and optimization
7.3 · Optimizing Workflows
In Wrenfield Housing, the course example, the operations team writes a monthly operations report for the board: repairs, voids (homes standing empty between tenancies), arrears and complaints, with totals from anonymised summaries. Each month someone rebuilds it in a fresh chat, re-uploads the same files, retypes the instructions and fixes the same formatting, so a report Claude drafts in minutes takes most of two days.
share
The tempting move is to speed up the part you can see, Claude's reply, with a faster model for every step or a lighter check at the end. The reply is rarely the slow part, because most of a recurring job's time goes on setting up, fixing the same faults and retyping. A faster reply saves little, and a lighter check can let a wrong total reach the board, so the job gets quicker and worse at once.
KnowledgeEfficiency, effectiveness and the whole job
A workflow is a job done the same way each time it recurs, as a series of steps. Efficiency is the time and usage one run takes. Usage counts because your usage limits depend on the model, the features and the length of the conversation. Effectiveness is whether the output still does its job and passes the checks it passed before. A change improves the workflow only when efficiency rises and effectiveness holds, so you measure the whole job, set-up and checking included, before and after the change. This rules out judging a change by how fast Claude answers.
A process map times a team's process to find its waits and bottleneck (4.2), and the same timing works on your own recurring job:
- Time one real run: every step from set-up to sign-off, with the hours each one took.
- Mark each step: repeated unchanged every run, done one item at a time, adding nothing, or catching an error that would be costly.
- Change one thing at a time, so each saving traces to one change (5.4).
- Re-run real past examples through the new set-up and compare each with the version that was approved.
- Time the next live run, and keep the change only if the time fell and every check still passes.
Interactive · where the two days go, and which changes keep the report as good
Step through the monthly board report one change at a time, and see which steps must stay.
Before: 14 hours over two days, mostly set-up, the same formatting fixes and retyping.
Reuse: a Project holds the brief and an approved report, so set-up and formatting shrink. The job takes 9.75 hours.
Batch: the four tables share one layout, so they go in one request. The job takes 7.5 hours.
Remove: the report now exports to Word, so nobody retypes it. The job takes 5.5 hours, but is it as good?
Keep the checks: the board acts on these totals, so both people's checks stay. Three past reports, re-run, matched.
KnowledgeFive levers for a recurring job
Each mark from step 2 points to a lever. Reuse moves anything retyped or re-uploaded every run into a Project, a workspace whose knowledge files and instructions every chat in it uses. Projects exist on every plan. Uploaded Project documents are also cached for a while, so using them again soon counts less against usage limits. Batching follows Anthropic's usage advice to group related tasks in one message. The other three levers match the model to each step, cut what adds nothing, and keep a check where an error costs most:
Standing files and instructions in a Project, or a tested prompt saved in a shared document.
Wrong choicePasting last month's instructions into a fresh chat.
Similar items in one request, such as tables that share one layout.
Wrong choiceBatching steps whose output must be checked before the next one starts (1.2).
A faster model for routine steps, once real examples pass on it (3.3).
Wrong choiceOne model for every step, whether fast or capable.
A step whose output nobody uses, or that redoes what a feature already does.
Wrong choiceRemoving a check because it is the slowest step.
A person checks where an error costs most, such as figures for the board (2.4).
Wrong choiceChecking every step twice, or checking nothing.
SkillWorked example: the board report that took two days
The operations officer who builds Wrenfield's report timed one run, the report for the October board. Six steps took 14 hours, and only step 2 involved Claude's replies. The new set-up uses the fastest model family, Haiku, for the tables, and the middle one, Sonnet, for the commentary:
# before: report for the October board, a fresh chat 1. Re-upload last month's report and definitions, retype the brief 2 h 2. Draft the four sections, one message each 3 h 3. Fix the same headings, table layout and terms 3 h 4. Retype the draft into the board's Word template 2 h 5–6. The officer's figures check; the director's sign-off 4 h → 14 hours over two days # after: the Monthly board report Project, November board 1. Upload this month's four summaries 15 min 2. Tables in one request (Haiku); commentary (Sonnet) 45 min 3. A few fixes; the Project holds the brief and a past report 30 min 4. Removed: exported to Word from Claude 0 5–6. Unchanged 4 h → 5.5 hours; July, August and September re-run, 3 of 3 match
The two days went on rebuilding, not drafting. Steps 1, 3 and 4 redid the same work every month, so the officer put the brief, the definitions of terms such as voids and the approved September report into a Project. Claude now builds the report as an artifact, a document shown beside the chat, using the Docs template, which exports to Word. Steps 5 and 6 stayed, because Wrenfield's AI policy requires a person to check anything that goes to the board. Setting up the Project and the re-runs took about four hours once, and the first month's saving covered that.
The re-runs showed the report was still as good. Before going live, the officer ran the July, August and September reports through the new set-up and compared each with the approved version, as 5.4 re-runs the same saved requests after each change. The first try put the commentary on Haiku too, and the August re-run called a rise in arrears a fall. The commentary went back to Sonnet, and the next re-run matched all three approved reports on every total and check.
A Skill is the other way to reuse. A Skill is a folder of instructions that Claude loads only when a task needs it, in any chat, so it suits a repeatable procedure. Skills are on every plan and on by default for Team organisations, but they need code execution, a feature that lets Claude run code.
- Efficiency is the time and usage one run takes; effectiveness is whether the output still does its job. An optimization needs both.
- Time the whole job, set-up and checking included, because a recurring job loses most of its time around Claude's reply, not in it.
- Reuse what repeats, batch what is alike, match the model to each step, cut what adds nothing, and keep a person's check where errors cost most.