tailor-cv-batch
Skill2 starsChanged 3 months agoMITView on GitHub
Tailor several CVs at once — one agent per job, run in parallel, each following /tailor-cv end to end, with a single table of results back. Use when a triage run produced more than one pick worth applying to, or whenever more than one résumé is asked for.
In this repository11 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- apply-formSkill
- cover-letterSkill
- evaluate-roleSkill
- job-triageSkill
- publishSkill
- research-companySkill
- setupSkill
- sync-appliedSkill
- tailor-cvSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 41
If there is no apply doc at all, take the list from the user and build the same structure by hand.
- Very specific: line 99
Report a stopped-early CV as partial fit, not as a failure: When the only sub-bar dimension is keyword_coverage and its misses are terms …
- Distinctive: line 108
Every claim still traces to the bank.
- Distinctive: line 109
Do not commit from inside an agent — one commit at the end, by you, so the run is one change.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Build many tailored CVs concurrently. Every CV is still /tailor-cv, unchanged — this skill is the
fan-out around it plus the parts that were being re-derived by hand every run.
When to use this instead of /tailor-cv: whenever the answer is more than one CV. /tailor-cv is
the right tool for exactly one job; past that it costs one serial round-trip per job for no reason,
because parsing, drafting, rendering and grading are per-job and independent. The only shared input,
profile/bullet-bank.md, is a read.
1. Build the work list — do not assemble it by hand
.venv/bin/python -m cv.batch worklist "<path to the dated apply doc>" --min-fit 70Returns JSON: company, role, folder, link, has_jd, existing. Three things it filters, and each is
a job that must never be rebuilt:
- ticked boxes — applications already sent;
- struck-through entries (
~~...~~) — duplicates the run resurfaced, which usually means Ben is already in that process under another company name; - every section that is not the apply set — the apply doc's audit blocks are full of
- [ ]lines too, and📬 Reply, don't cold-applyis the expensive one to get wrong: those are live conversations.
--min-fit N is the gate /job-triage uses, and it is 70 (2026-08-04). It filters on the score
the apply doc printed rather than on position: a positional cap (--top N, still available) drops the
job at N+1 for being eleventh, where a score gate drops only the ones that were not worth the tokens.
Ben: "for good fits (>70) you should probably build anyway." An entry whose line printed no score is
kept — absence of a number is not evidence of a low one, which is what carryovers look like.
Either way it is a gate, not a quota: a run where four roles clear 70 returns four.
If the command raises, it found a ## heading it cannot classify. Fix the heading in the apply doc
to one of /job-triage Step 6's — do not work around it and do not hand-assemble the list instead,
which is how the research, résumé and letter stages start disagreeing about which jobs are in the set.
If there is no apply doc at all, take the list from the user and build the same structure by hand.
Then decide the set, and say which jobs you dropped and why. Two rules:
has_jd: falsemeans there is no posting to write against. Fetch it first (WebFetch; LinkedIn is usually login-walled — ask Ben to paste it) or drop the job. Never write a CV from a run summary: on 2026-07-28 the run summaries were wrong three times in one day.existing: trueis a rebuild, not a skip, ifprofile/bullet-bank.mdhas changed since the folder was written.git log -1 --format=%cd profile/bullet-bank.mdagainst the folder mtime answers it.
2. Fan out — one agent per job, all launched in a single message
Spawn one general-purpose agent per job, all in one message so they run concurrently. Each gets a
prompt that says: work in the repo root in place (no worktree), follow
.claude/skills/tailor-cv/SKILL.md exactly and in full, read profile/bullet-bank.md,
docs/knowledge-base/personal/tailoring-playbook.md and profile/rubric.md before drafting, and do not commit.
Four things every agent prompt must carry, because each was a real failure:
- The folder and the job link, and whether it is a rebuild.
- Read-only on the bullet bank. N agents appending to
profile/bullet-bank.mdconcurrently is how a corrupted bank reaches every application in the run. A genuinely new bullet is proposed in the agent's return value with its evidence; you apply the accepted ones yourself, serially, in §4. Same shape as the grader that reports fixes and never applies them — the machine proposes, the step with a human in it disposes. - The refusal list, spelled out. Reviewer fixes that add SQL, OpenAI, Cursor, Copilot, CI/CD, Kubernetes, a years-of-experience number or an invented metric are refused and reported. The bank and Ben's stated preferences outrank the grader, always.
- Build first, report after. Ben's standing instruction: render, then show what changed for
post-hoc edit. Do not stop at
/tailor-cv§4's approval gate.
And what to return: folder, page count, depth, headline changes, the reviewer's four scores and pass count, refused fixes, what the gloss pass cut, proposed bank additions, and the honest keyword gaps.
3. The mechanical steps are code — agents call them, they do not re-derive them
.venv/bin/python -m cv.batch fit <folder> # render, tighten, re-render until one page
.venv/bin/python -m cv.batch degloss <folder> # the gloss substitutions with no honest rewritefit only ever touches the plan's tighten block. It will not drop a bullet to win a page —
that is a claim decision, and it belongs to whoever can weigh what is lost. When it exits 3 the
document genuinely does not fit and the agent picks the trim.
degloss is deliberately a short list of phrases that carry a stance and no fact. The judgement
half of /tailor-cv §5b — does this sentence add a fact or a stance? — stays with the agent. A style
ban enforced by a word list is dodgeable by substitution, and this repo already reversed one such
test: int8-quantized, TorchScript-compiled duly became the vaguer size-optimized.
4. Collate, then apply bank changes serially
One table: role · pages · the four scores · passes · refused fixes · keyword gaps. Then show Ben the
proposed bank additions with their evidence and wait — profile/bullet-bank.md is protected and no
agent writes to it without his explicit yes, orchestrator included. Apply the approved ones one at a
time. (Serial application is the concurrency fix; his approval is a separate gate and neither
substitutes for the other. See the notice at the top of the bank.)
Report a stopped-early CV as partial fit, not as a failure. When the only sub-bar dimension is
keyword_coverage and its misses are terms with no evidence behind them, more passes cannot help —
say how many must-haves are unclaimable. On the Apex req that was 11 of 23, which is a fact about the
requisition rather than a defect in the document.
Rules
- Never let a batch agent write to
profile/bullet-bank.md,profile/rubric.md, or another job's folder. - Every claim still traces to the bank. Parallelism changes the schedule, never the standard.
- Do not commit from inside an agent — one commit at the end, by you, so the run is one change.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

